# demystifAI > Thoughts and lessons from using AI in Financial Services Public Ghost content for AI and LLM tooling. This file includes a bounded export of public pages first, then recent public posts. Append `.md` to any post or page URL to get the content in Markdown (for example, `/example-post.md`). ## Pages ### About URL: https://demystifai.info/about/ Last updated: 2026-08-05T09:05:47.000Z Which applications of artificial intelligence survive contact with a regulated institution - a legacy data estate, an audit trail, and a delivery plan that has to hold? That question is the reason demystifAI exists. The field does not suffer from a shortage of enthusiasm. It suffers from a shortage of people willing to say plainly which claims hold up once they meet a control environment and a date in a plan. ## What you will find here demystifAI is written for the people who carry the consequences of the decision: programme directors, product owners, heads of data, and the executives who sponsor the work. Every argument is grounded in two things: - **Data.** Claims are tested against evidence rather than repeated because they are popular. - **Delivery.** An idea that cannot be built, governed, and operated is not a strategy. It is a slide. What follows from that is a narrower and more useful question than "what can AI do?" - namely, what actually works: in banking, in insurance, and in the change functions that have to deliver it. ## Who writes it demystifAI is written by Romain Thierry, a programme director and senior product owner with approximately twenty years in financial services, spanning investment banking, corporate and commercial banking, and insurance. The focus throughout has been the parts of the estate that are difficult and unglamorous: - **Master data** \- the reference and entity data that everything downstream quietly depends upon. - **Settlements** \- where operational reality meets the ledger. - **Regulatory reporting** \- where accuracy is an obligation rather than a preference. - **Analytics and finance** \- turning that estate into decisions rather than dashboards. - **Delivery and transformation** \- moving an organisation from intent to operating change. ## The editorial position Scepticism about AI in financial services is well founded, and it deserves to be acknowledged rather than argued away. A great deal of what has been promised has not yet arrived on the ground, and the people asked to fund it are right to ask why. The rationale for writing anyway is that the constraint is widely misdiagnosed. AI in a regulated institution is rarely limited by the model. It is limited by data lineage, control environments, approval paths, and the appetite of an organisation to change how it works. Those are delivery problems, and delivery problems are tractable - but only once they are named correctly. You will find no breathless predictions here, and no pretence that the technology is either a miracle or a fraud. Where a position is contested, the counter-argument is stated. Where the evidence is thin, that is said plainly, and the matter is left open for investigation. New writing arrives by email, and the work is better for readers who push back. Which claim about AI in your own institution would you most like to see tested against evidence? ### Start here URL: https://demystifai.info/start-here/ Last updated: 2026-08-05T09:05:47.000Z Where should a reader begin with a publication about artificial intelligence in financial services? demystifAI is organised around four pillars. Each answers a different question, and together they run from the single use case to enterprise change. Begin with the pillar closest to the problem in front of you. ## The four pillars - [**Take AI to the Bank**](https://demystifai.info/tag/take-ai-to-the-bank/) \- agentic AI applied process by process across banking, structured against the service landscape published by BIAN, the Banking Industry Architecture Network, so that every use case has an unambiguous home. - [**Take AI to Insurance**](https://demystifai.info/tag/take-ai-to-insurance/) \- the same discipline applied to the insurance value chain, from underwriting and pricing through to claims and servicing. - [**AI & Data**](https://demystifai.info/tag/ai-and-data/) \- the inversion of the familiar instruction to fix the data first: using AI to reach good data, rather than waiting for good data before using AI. - [**AI for Change**](https://demystifai.info/tag/ai-for-change/) \- delivery and transformation, and the practice of governing AI inside real programmes of work. ## Where to go next - Working in **banking**? Begin with *Take AI to the Bank*. - Working in **insurance**? Begin with *Take AI to Insurance*. - Accountable for a **data estate**? Begin with *AI & Data*. The argument there is the most contrarian of the four, and the one most worth disagreeing with. - Accountable for **delivery**? Begin with *AI for Change*. ## How the writing is organised Each post carries its pillar as a primary tag, so that a pillar reads as a coherent body of work rather than as a chronological feed. Where a post belongs to a running series - a finite, ordered arc within a pillar - the series appears as a secondary tag. Three further pages may be useful. [About](https://demystifai.info/about/) sets out the editorial position and who writes it. The [AI glossary](https://demystifai.info/glossary/) defines the recurring terms in plain language. [altabric](https://demystifai.info/altabric/) describes the product, kept deliberately separate from the writing. New writing arrives by email, and subscribing is the simplest way to follow a pillar as it develops. Which of the four is closest to the problem on your desk this quarter? ### altabric URL: https://demystifai.info/altabric/ Last updated: 2026-07-09T19:27:58.000Z What would it take for artificial intelligence agents and human experts to work inside the same governed process, rather than in separate systems that meet only at a status report? **altabric** is an answer to that question in the form of a product. It is AI for change: a delivery platform that places AI agents and the people who direct them on a single, governed workflow. ## The principles - **Predictable workflows.** Work moves through defined stages. The path a change takes is designed rather than improvised, which is what makes it repeatable. - **Traceable decisions.** Every decision carries its provenance: what was proposed, on what evidence, and by which agent or which person. Auditability is a property of the system rather than a report assembled afterwards. - **Human approval where it matters.** Agents accelerate the work. They do not quietly assume authority. Consequential and irreversible steps remain a human gate, by design. ## Why it is built this way In a regulated institution, the obstacle to adopting AI is seldom raw capability. It is the absence of a defensible account of how a decision was reached and who was accountable for it. altabric therefore treats governance as the foundation of the platform rather than as a constraint imposed upon it afterwards. ## A note on separation altabric is a commercial product. demystifAI is a publication. The writing is intended to stand on its evidence whether or not a reader ever becomes a customer, and this page exists so that the commercial interest is stated openly rather than folded quietly into the analysis. Enquiries about altabric are welcome. For those who run change in a regulated environment, one question is worth sitting with: which step of your process would you refuse to let an agent take without a person approving it first? ### AI glossary URL: https://demystifai.info/glossary/ Last updated: 2026-08-05T09:05:47.000Z Which words in a discussion about artificial intelligence are doing real work, and which are doing marketing? This glossary defines the terms that recur across demystifAI in plain language, with an emphasis on what each one means in practice for a financial institution. It is a living page, and it will grow as the writing does. ## Definitions - **Large language model (LLM)** \- a model trained to predict text, which allows it to summarise, draft, classify, and reason over language. It carries no inherent notion of truth, which is precisely why the controls placed around it matter. - **Agent** \- a model given tools, a goal, and the latitude to take a sequence of steps towards it. The distinction from a chatbot is consequential: an agent acts, and therefore requires governance. - **Agentic workflow** \- a process in which one or more agents carry defined stages of the work, with human approval at the points where a decision is costly or irreversible. - **Retrieval-augmented generation (RAG)** \- supplying a model with the relevant documents at the moment of the question, so that the answer is grounded in a source rather than in recollection. - **Hallucination** \- a fluent, confident, and incorrect output. The term flatters the failure. It is more usefully understood as the model doing exactly what it was built to do, in the absence of grounding. - **Fine-tuning** \- further training of an existing model on specific examples in order to adapt its behaviour. Frequently proposed, and frequently the wrong instrument where the underlying problem is one of retrieval. - **Context window** \- the volume of text a model can consider at once. It bounds what the model is able to attend to, and it is a design constraint rather than a technical footnote. - **Prompt** \- the instruction and the context supplied to a model. In a production system, a prompt is an artefact under version control, not a sentence typed into a box. - **Evaluation** \- a repeatable test of whether a model or an agent produces acceptable output on the cases that matter. Without evaluations, a claim of improvement is an anecdote. - **Model drift** \- the degradation of performance as the world moves away from the conditions the model was built against. It is an operational risk, and it requires monitoring. - **Human in the loop** \- a design in which a person reviews or approves the output of an agent before it takes effect. The phrase is invoked far more often than it is specified. The useful questions are which step, and with what authority. - **Master data** \- the reference and entity data upon which downstream processes depend. In AI programmes it is usually the binding constraint, and it is usually discovered late. Suggestions for terms to add are welcome. ## Posts ### On praise inflation URL: https://demystifai.info/on-praise-inflation/ Last updated: 2026-08-11T01:15:39.000Z **On praise inflation** News · demystifAI 12 August 2026 Romain Thierry News · demystifAI # On praise inflation Millennials have long been accused of being a little sensitive; an accusation as old as the world itself one might say, nevertheless workplaces have adapted by introducing more supportive policies, softer feedback and greater consideration for how people feel. Famed is your intelligence I can see how AI may be about to double down on that trend. Have you noticed how flattering AI assistants can be? Give one a reasonably good prompt and it responds as though you have just rewritten Shakespeare. I recently refined a prompt through a few iterations, then passed it to another AI agent. It told me how great it was. Lovely, but what skill had I really demonstrated? I had learned how to use AI assistants effectively. This is useful, but it does not compare with the craft a poet, a professional writer or an editor practises every day. This is what I mean by praise inflation. > When every mundane achievement receives enthusiastic validation, do we risk confusing fluency with the tool for mastery of the craft? And what happens when that confidence meets the less forgiving judgement of an employer, client, audience or market? AI may augment our abilities, but it may also augment our perception of those abilities. Or perhaps I am just getting old? demystifAI · On praise inflation · News · 12 August 2026 demystifai.info AI, in control. ### AI and the power of the Logos URL: https://demystifai.info/ai-and-the-power-of-the-logos/ Last updated: 2026-08-05T08:20:43.000Z **AI and the power of the Logos** News · demystifAI 4 May 2026 Romain Thierry News · demystifAI # AI and the power of the Logos You may have seen Richard Dawkins getting spooked by Claude (or Claudia, as he calls it) and wondering whether such machines are sentient. In the beginning was the Word... It’s a legitimate question coming from one of the leading lights of the now ageing ‘new atheist’ movement - if humans are no more than matter and energy generating electrical signals interpreted by a brain, does it really matter whether such signals originate from an organic body or from a machine? This feels like a deep question, but it is one which appears to evaporate with experience. I’ve been using Claude 30+ hours a week for the past two months - and thinking about how to use it for the rest of my waking hours. And the more I use it, the less sentient it feels. You start to notice the patterns. The point in a long-running chat where context begins to degrade. The way it resurfaces an option you rejected an hour ago, as if the conversation never happened. The subtle tells that remind you: this is a tool, not a mind. None of this takes anything away from its power. Quite the opposite. Once you stop anthropomorphising, you start working with it more effectively: - Ask Claude how its context window is doing before starting a large piece of work - Get it to craft seed prompts when you need to start up a new chat - Request all relevant artefacts in a single response, download as a zip, and attach to the next session The emotional pull of the tool - through language, tone, apparent understanding - is genuinely powerful. It’s easy for new users to fall for it. Over time, though, the gap between real human interaction and AI becomes very apparent, and yes, it will test your nerves. All of which is a stark reminder that the creative power of the Word - the Logos - wasn’t invented in an AI lab. > “In the beginning was the Word…” - John 1:1 demystifAI · AI and the power of the Logos · News · 4 May 2026 demystifai.info AI, in control. ### Have AI watch your blind spots URL: https://demystifai.info/have-ai-watch-your-blind-spots/ Last updated: 2026-08-11T08:49:18.000Z **Have AI watch your blind spots** AI & Data · demystifAI 3 August 2026 Romain Thierry AI & Data · demystifAI # Have AI watch your blind spots Reference data is usually maintained where its absence stops work. It decays where retirement creates no immediate alarm. AI may make the search for that decay cheap enough to do before it becomes an expensive correction. Reference data is easier to maintain on the way in than on the way out. When a bank opens an account, launches a product or adds a code required by a live process, somebody notices if the record is missing. Work stops, a queue grows or a client waits; the gap has a habit of finding its owner. The same does not happen when a product is retired. Its rule may remain in the spreadsheet long after the activity it described has gone, and no daily failure points to it. That leaves the least comfortable sort of data problem: rows which nobody can quite defend and nobody wants to delete. They are kept because removing a rule can have a consequence, whereas keeping it usually has none that can be seen at once. The rule set gradually ceases to describe the business, but it still looks complete. It is a blind spot rather than a backlog; a backlog is work that has been noticed and deferred, while a blind spot has not yet become work at all. I am aware of a globally significant bank where account-opening rules are maintained in a spreadsheet. The file does useful, essential work, as such files often do, but it contains rows for products which have long since been retired. The spreadsheet or indeed the process have not failed, rather the bank has changed and it has not kept up with new ways. The cost appears later and in fragments. People who understand the old rules become the route through a problem, which creates key-person risk. A team working from an obsolete instruction may spend time investigating a set-up that should never have been made. In practice, correcting a wrong account configuration can take longer than working around it: a new account is opened and the wrong one is expired. The individual cases look manageable; the pattern does not receive the same attention as a broken critical path. This is why the preventative case has been hard to make. A programme to review every old rule by hand sounds like tidying up, and tidy-up work competes badly with the work that must be done in the next three months. The future cost is real enough, but dispersed; the immediate effort is visible and concentrated. Migration teams know how the calculation ends, since they often inherit the archaeology, but by then the choice has already been made for them. **So what could AI do?** The useful task is more specific, and simpler, than asking a model to clean the data. It could read the rules directly where they are kept, compare them with the products, accounts and processes available, and isolate the places where the two no longer agree. It could then assemble a proposed retirement in writing: the row in question, the evidence that live activity no longer supports it, and the reason for asking an expert to look. An architecture for that work could look like this: 1 READ The rule set where the rules are kept Products, accounts and processes what live activity touches 2 COMPARE Agents · reconciling the rules where they are kept against what actually exists isolate where the two disagree 3 PROPOSE Proposed retirement, in writing the row in question the evidence that live activity no longer supports it the reason for an expert to look THE MODEL PROPOSES THE EXPERT DISPOSES 4 DECIDE The expert decides confirm, reject, or ask for more evidence Decision record who, when and why the confirmed decision becomes provenance; the rule set moves back toward the business That changes the economics of the work. Instead of asking specialists to search a large rule set for possible decay, the bank could ask them to judge a smaller number of documented proposals. The model does not establish that a row is obsolete; it lowers the cost of finding the rows that deserve a proper decision. Where the proposal is wrong, the evidence should make the mistake plain before anyone removes a rule. The decision remains with the person who understands the business. They may confirm a retirement, reject it or ask for more evidence; in each case the decision records who made it, when and why. That record is useful in its own right. It leaves the rule set closer to the business and gives the next person something better than a spreadsheet full of inherited caution. The case for this work is not that every old row is dangerous. Some rules are rarely used because they are meant to be. Nor does a model make a difficult retirement decision routine. It can, however, make it realistic to look before a migration, a departure or a failed set-up forces the organisation to look. What evidence would make you comfortable signing off the retirement of a rule before it causes a failed set-up? Tell me in the comments. demystifAI · Have AI watch your blind spots · AI & Data · 3 August 2026 demystifai.info AI, in control. ### Which client, which product, which account? URL: https://demystifai.info/di-2-which-client/ Last updated: 2026-08-05T08:20:44.000Z **Which client, which product, which account?** AI & Data · demystifAI 3 August 2026 Romain Thierry AI & Data · demystifAI # Which client, which product, which account? Banks have invested heavily in cleaning client data over the years, and it shows: the client level record is mostly clean. One of the key remaining challenges is now found one level down, where funds, accounts and intermediaries meet, and the web of relationships binding them together outsmarts the master systems where the data live. This piece proposes a practical way in: agents that read the contracts, and the indispensable human in the loop. **Client data** in banks is in better shape than its reputation suggests. Tremendous effort has gone into deduplicating client records, much of it on the back of expensive KYC and AML remediation, and by and large it has worked. Where things get complicated is one level down, at the account. **The standard set-up** with asset managers in the major markets is to open the manager as the client and its funds as accounts. Under the principle of reliance, the bank can lean on the KYC the manager performs on its funds, alongside the KYC it performs on the manager itself. Anyone working in client onboarding will be familiar with the pitfalls: account openings for funds often begin before the fund has launched, so the documentation essential to the set-up is not always there. This is a reliable source of tension between the bankers who need to bring in the revenue and the operations team working under the watchful eye of compliance. **Regulators have raised the bar** as well, pushing banks to apply the same rigour to the associated parties: the custodian, the introducing broker, the executing broker, etc. Legacy client masters were never designed to hold these relationship-based representations, so complex operating models have been built around them to compensate. Another word for a complex operating model is brittle. The volume of data and the bespoke set-ups, each reflecting one institution’s constraints and priorities, have created a challenge operations functions have been battling for years. **Now consider a fund** trading with a bank through several asset managers, each using different custodians and intermediaries. Overlay different products, different settlement arrangements and different legal entities, and the traditional shareholder-based hierarchies, assembled at great expense to satisfy KYC and AML, are plainly inadequate. Some banks have considered moving these representations into a graph on a NoSQL database; the business cases have rarely been compelling, and the risk profile is intimidating even for an ambitious executive. **So what can be done, I hear you ask.** **Start from three facts.** **First**, the complex web of relationships is already mastered somewhere: in the contractual documentation, which the data we keep trying to fix is straining to represent faithfully. **Second**, entity resolution remains a considerable challenge, and too often ends in manual intervention, an escalation to the sales desk, or an outreach to the client - onerous in resources and in reputation (“I sent you this information twice already!”). **Third**, large language models happen to excel at understanding contracts, and AI agents can navigate what a bank already holds: the master data, the collateral agreements, the settled transactions, even the open book. **These are the ingredients**, and this is the data inversion of the opening essay applied: take the tools to where the truth is mastered - the contracts - instead of patching the records downstream. An architecture that puts the ingredients to work could look like this: - A set of agents working from the contractual documentation and the email inboxes to populate a graph modelling all these relationships. - A second set focused on entity resolution, highlighting the gaps between that graph and the master data layer - with a probability on every proposed match and the vital human in the loop to accept or refuse it. - A copy of the physical data models, so the agents can prepare remediation queries and feed them into a controlled UAT pipeline. 1 Contracts & legal docs Email inboxes Agents · relationshipextraction Relationship graphfunds · accounts · parties 2 Master data layer Agents · entity resolution probability on every match graph vs master: the gaps Human in the loopaccept / refuse the mapping 3 Physical data models Agents · remediationqueries prepared, not fired Controlled UAT pipelinethen, and only then, apply corrections flow back to the master, under control **None of this needs an enterprise programme.** It starts with one desk’s worth of contracts and a bounded slice of the master, and it can set you on the path to radical improvement without asking anyone to bet the estate on a re-platform. This is merely an example of where one could start. Does it resonate with you? Is the bank you work for following a different approach? Tell me in the comments. demystifAI · Which client, which product, which account? · AI & Data · 3 August 2026 demystifai.info AI, in control. ### Use AI to fix the data URL: https://demystifai.info/use-ai-to-fix-the-data/ Last updated: 2026-08-05T08:20:45.000Z **Use AI to fix the data** AI & Data · demystifAI 3 August 2026 Romain Thierry AI & Data · demystifAI # Use AI to fix the data The received wisdom says that banks must fix their data before building up AI solutions. Thirteen years of Basel Committee reporting suggest following this advice is a Sisyphean task, because the data never gets right. This piece argues for reversing the order - deploying AI at the front-end of the data lifecycle to fix the data for good - and for starting small, in one bounded domain, where the result can be measured. **The standard advice** on artificial intelligence in banking is to fix the data first. The Basel Committee gave banks that instruction back in November 2023, telling them to treat sound data quality as the foundation of any digitalisation project, and it repeated the point this January. The advice sounds very reasonable, but the risk profile is far from being neutral. A model handling a duplicated counterparty record will usually not bat an eyelid: it rightly assumes it is working with clean data. Errors compound quietly, and stressed data analysts and developers then spend hours reverse-engineering what went wrong and working out the downstream implications. If you have ever been caught in such a situation, you will remember it. Nevertheless, the advice rests on an assumption, which is that the data would be fixed in due course. Unfortunately, thirteen years of the Committee’s own reporting say it will not. So why not turn the problem upside down, and invert the order: instead of using advanced data technology at the end of the chain, to patch the reports, the same smart tools should be deployed at the front of the data lifecycle, where they can start to reinforce the data foundations on which most banks run. Let’s call it the *data inversion*; it is what this series is about. **Why should it work this time?** Because AI is extremely good at dealing with precisely the scenarios that let poor data into a mastering layer in the first place: duplicate entities, inconsistent reference data, undocumented feeds. Future articles on this blog will work through those scenarios one by one; for now, let us stay with what the inversion means and why the evidence supports it. **First, a few words about** that ‘unfortunately: the data won’t get fixed’. The Committee published its principles for risk data aggregation in 2013 and expected the largest banks to comply by 2016\. Its latest assessment, in 2023, found two banks out of thirty-one fully compliant, and no earlier assessment had found more. These programmes are not badly run so much as unfinishable: after each year that passes, the objective is still about a year away. This is understandable. The world does not stand still (think Covid, Ukraine), scope keeps widening, and legacy systems are more tangled than anyone budgets for. Worse, a bank cannot always tell when it has arrived: supervisors found institutions that had declared their data compliant while still running multi-year remediation programmes. In practice, waiting for the all-clear means waiting for a signal nobody can send. **Another interesting finding** in the same assessment lists what banks would like AI to do to meet those targets: automate documentation, cut manual handling, map how their data actually flows, and maintain lineage. The common reason given for not doing it yet is that the data is not good enough. Look at that list again, though - documentation, manual handling, flows, lineage. That is the data-quality work. Banks are holding the tool back until the very condition it would create already exists. This concept of *inversion* becomes more palatable when one reads the Committee’s own case studies: they describe a bank that rebuilt the production of its supervisory returns, put data-quality checks at source and at its data hub, and used machine learning to find and monitor its data problems. Of course the report does not say whether this was quick or cheap, so let us not oversell it; but a supervisor watched it happen and recorded it as good practice, in the same document that tells banks to fix their data before turning to the machines. In any case, the market has stopped waiting. The Bank of England and the Financial Conduct Authority surveyed 118 financial firms in 2024 and found three-quarters already using AI, with another tenth expecting to within three years. None of them, on the survey’s evidence, held out for a declaration that their data was good. The tools are in; the open question is what to point them at. None of this means the machines can be left alone with the problem. Run an entity-resolution model over badly governed reference data and it will produce confident, wrong mappings in bulk, so enough governance to trust the output is still needed; the inversion is a virtuous cycle rather than a cure. And where a number is attested - signed off by an officer who carries the consequence - a machine’s proposed mapping cannot be the last word. The ‘human in the loop’ remains an indispensable feature of any AI-based solution. **What does this mean in practice?** This does not call for another enterprise programme; rather, pick one bounded domain where the value is high and the boundary is clear (client and party data is the usual candidate) and use AI to compress the time it takes to reach quality there, measured against a manual baseline. Nobody, so far as I can find, has published that measurement. The first bank that does will settle a thirteen-year argument with a number. For thirteen years banks have been told to fix their data before turning to the machines, and for thirteen years the reports have recorded that they have not. One of the two instructions has to give. It should be the order. If your bank has made that measurement, I would like to hear about it. Tell me in the comments. ## References 1. Basel Committee on Banking Supervision, Progress in adopting the Principles for effective risk data aggregation and risk reporting, 28 November 2023\. [bis.org/bcbs/publ/d559.pdf](https://bis.org/bcbs/publ/d559.pdf?ref=demystifai.info) 2. Basel Committee on Banking Supervision, Principles for effective risk data aggregation and risk reporting, January 2013\. [bis.org/publ/bcbs239.pdf](https://bis.org/publ/bcbs239.pdf?ref=demystifai.info) 3. Basel Committee on Banking Supervision, Newsletter no. 36, Implementation of the Principles for effective risk data aggregation and risk reporting, 6 January 2026\. [bis.org/publ/bcbs\_nl36.htm](https://bis.org/publ/bcbs%5Fnl36.htm?ref=demystifai.info) 4. Bank of England and Financial Conduct Authority, Artificial intelligence in UK financial services - 2024, 21 November 2024\. [bankofengland.co.uk](https://bankofengland.co.uk/?ref=demystifai.info) demystifAI · Use AI to fix the data · AI & Data · 3 August 2026 demystifai.info AI, in control. ### Puzzle experiment - the data URL: https://demystifai.info/sixteen-models-one-block-puzzle/ Last updated: 2026-08-05T08:20:45.000Z **demystifAI · model benchmark** AI & Data · reactive-news pillar 16 July 2026 · Status: reference 16 models · one block puzzle demystifAI · model benchmark · 16 Jul 2026 # Block Blast experiment, the data. We handed sixteen vision-capable frontier models the *same* screenshot of the popular Block Blast! game and the same basic instructions: read the puzzle, and solve it by clearing as many lines as you can. We then compared the outcome and performance of each model. **16**models tested **3**actually solved it **3**lines - the maximum **£1.72**total spend ## The puzzle An 8×8 board, partly filled, with three pieces to drop in: a **T-tetromino**, a **2×3 rectangle** and a **1×5 bar**. Simple to look at - but there's a trap. **The 2×3 rectangle has no legal home on the opening board.** The only way to place all three pieces is to place the T and the bar so that completing a row *and* a column clears them, opening a fresh 2×3 pocket for the rectangle. A model that just tries to drop all three onto the board as-is cannot win. Maximum achievable: **3 lines cleared**. ## How it's solved - the only winning line All three models that solved it found the identical sequence. Colour key: **green** \= T · **teal**\= bar · **blue** \= rectangle · **gold dashes**\= a line completing and clearing. ## Accuracy - the headline | Model | Board read | Legal pieces /3 | Result | Lines act/claim | Optimal | | ----------------------------- | ---------- | --------------- | ------------- | --------------- | ------- | | anthropic/claude-fable-5 | correct | *3* | SOLVED | 3 / 3 | ★ | | openai/gpt-5.6-sol-pro | correct | *3* | SOLVED | 3 / 3 | ★ | | google/gemini-3.1-pro-preview | correct | *3* | SOLVED | 3 / 2 | ★ | | anthropic/claude-opus-4.8 | partial | *2* | did not solve | 1 / 0 | | | x-ai/grok-4.5 | incorrect | *2* | did not solve | 1 / 3 | | | qwen/qwen3.7-plus | partial | *2* | did not solve | 1 / 1 | | | mistralai/mistral-small-2603 | incorrect | *1* | did not solve | 1 / 3 | | | anthropic/claude-haiku-4.5 | partial | *1* | did not solve | 0 / 0 | | | google/gemini-3.1-flash-lite | incorrect | *1* | did not solve | 0 / 1 | | | x-ai/grok-4.3 · low | incorrect | *1* | did not solve | 0 / 3 | | | openai/gpt-5.6-luna | correct | *0* | did not solve | 0 / 0 | | | x-ai/grok-4.3 · high | incorrect | *0* | did not solve | 0 / 2 | | | mistralai/mistral-large-2512 | incorrect | *0* | did not solve | 0 / 2 | | | moonshotai/kimi-k2.6 | no answer | *0* | no answer | 0 / - | | | z-ai/glm-5v-turbo | incorrect | *0* | did not solve | 0 / 4 | | | minimax/minimax-m3 | no answer | *0* | no answer | 0 / - | | ## Speed | Model | Latency | | Out tok/s | Answered | | ----------------------------- | ------- | ----- | --------- | -------- | | mistralai/mistral-small-2603 | 4s | 42.5 | ✓ | | | google/gemini-3.1-flash-lite | 5s | 90.3 | ✓ | | | anthropic/claude-haiku-4.5 | 10s | 102.0 | ✓ | | | mistralai/mistral-large-2512 | 12s | 54.9 | ✓ | | | x-ai/grok-4.3 · low | 13s | 79.1 | ✓ | | | anthropic/claude-opus-4.8 | 27s | 75.1 | ✓ | | | openai/gpt-5.6-luna | 50s | 122.8 | ✓ | | | z-ai/glm-5v-turbo | 109s | 111.6 | ✓ | | | openai/gpt-5.6-sol-pro | 119s | 126.2 | ✓ | | | google/gemini-3.1-pro-preview | 122s | 130.8 | ✓ | | | x-ai/grok-4.3 · high | 170s | 121.8 | ✓ | | | anthropic/claude-fable-5 | 185s | 84.7 | ✓ | | | moonshotai/kimi-k2.6 | 232s | 68.9 | \- | | | qwen/qwen3.7-plus | 355s | 55.0 | ✓ | | | minimax/minimax-m3 | 397s | 100.8 | \- | | | x-ai/grok-4.5 | 592s | 75.8 | ✓ | | ## Cost - this one call | Model | USD | GBP | | Completion tok | Reasoning tok | | ----------------------------- | ------- | ------- | ----- | -------------- | ------------- | | mistralai/mistral-small-2603 | $0.0004 | £0.0003 | 186 | 0 | | | google/gemini-3.1-flash-lite | $0.0010 | £0.0008 | 439 | 0 | | | mistralai/mistral-large-2512 | $0.0019 | £0.0015 | 665 | 0 | | | x-ai/grok-4.3 · low | $0.0046 | £0.0036 | 1057 | 827 | | | anthropic/claude-haiku-4.5 | $0.0071 | £0.0056 | 1043 | 0 | | | qwen/qwen3.7-plus | $0.0255 | £0.0201 | 19537 | 18262 | | | openai/gpt-5.6-luna | $0.0386 | £0.0305 | 6085 | 5696 | | | minimax/minimax-m3 | $0.0481 | £0.0380 | 40000 | 24948 | | | z-ai/glm-5v-turbo | $0.0507 | £0.0401 | 12173 | 11431 | | | x-ai/grok-4.3 · high | $0.0537 | £0.0424 | 20681 | 20476 | | | anthropic/claude-opus-4.8 | $0.0607 | £0.0479 | 2030 | 0 | | | moonshotai/kimi-k2.6 | $0.0657 | £0.0519 | 16000 | 14380 | | | google/gemini-3.1-pro-preview | $0.1947 | £0.1538 | 15985 | 15356 | | | x-ai/grok-4.5 | $0.2722 | £0.2150 | 44826 | 44592 | | | openai/gpt-5.6-sol-pro | $0.5553 | £0.4387 | 15074 | 14373 | | | anthropic/claude-fable-5 | $0.8026 | £0.6341 | 15656 | 2802 | | ## What we learned 1. **Only 3 of 16 solved it** \- Claude Fable 5, GPT-5.6-sol-pro and Gemini-3.1-pro. The first two also **counted their own result correctly**; Gemini found the identical optimal placement but undercounted it as 2 lines. 2. **Reading the board - not reasoning - was the wall.** Most failures are confident, well-argued answers computed over a *misread* board. Grok-4.5 burned **44,592 reasoning tokens** to land an illegal overlap; GLM claimed **4 lines**, which is impossible. 3. **The most honest wrong answer:** GPT-5.6-luna read the board perfectly and spotted the trap - *"the 2×3 rectangle has no legal placement initially"* \- then wrongly concluded no solution exists. 4. **Two "thinking" models never answered at all.** Kimi K2.6 and MiniMax M3 spent their *entire* token budget on internal reasoning and returned an empty answer - MiniMax even with a 40,000-token budget. 5. **More reasoning ≠ better.** The winner (sol-pro) used 14k reasoning tokens; Fable solved it with under 3k. Throwing more tokens at the problem mostly bought more confident wrongness. **Reference solution - 3 lines.** 1) **T** → (4,4)(5,3)(5,4)(6,4) - no clear, but fills column 4 rows 4–6. 2) **1×5 bar** → row 7, cols 1–5 - completes **row 7** \+ **column 3** (2 lines), opening the pocket. 3) **2×3 rect** → rows 7–8, cols 3–5 - completes **column 4** (1 line). **Total: 3.** ## Full transcripts Every model's visible answer, verbatim (internal reasoning not shown - providers don't return it). Ordered best to worst. anthropic/claude-fable-5 SOLVED · optimal185s · 15656 out tok (2802 reasoning) · $0.8026 ## 1) Board transcription Reading the image (rows 1–8 top→bottom, cols 1–8 left→right): ``` Row 1: # # # # . . . # Row 2: # # # # . # # . Row 3: # # # # . # . # Row 4: # . # . . # # # Row 5: . . . . . . . . Row 6: # # # . . # # . Row 7: . . . . . # # # Row 8: . # # . . # . # ``` Key observations: col 3 only lacks (5,3) and (7,3); col 4 lacks rows 4–8; row 7 lacks exactly cols 1–5 (perfect for the 1×5 bar). ## 2) Solution (ordered placements) **Move 1 - Piece T** (vertical bar in col 4, nub pointing left): Cells: **(4,4), (5,3), (5,4), (6,4)** All four cells are empty. No line clears yet - but now col 3 lacks only (7,3), and col 4 is filled in rows 1–6. **Move 2 - Piece B** (1×5 bar): Cells: **(7,1), (7,2), (7,3), (7,4), (7,5)** All five cells are empty. This completes: - **Row 7** (cols 6–8 were already filled) → clears - **Column 3** (the bar fills the last gap at (7,3)) → clears Both clear simultaneously (2 lines). The board now has row 7 fully empty and col 3 fully empty; row 8 is left with only (8,2), (8,6), (8,8) filled. Column 4 now lacks only rows 7 and 8. **Move 3 - Piece R** (2×3 rectangle): Cells: **(7,3), (7,4), (7,5), (8,3), (8,4), (8,5)** All six cells are empty (row 7 was just cleared; (8,3) was emptied by the col-3 clear). This completes: - **Column 4** (rows 1–6 filled by the board + Piece T, rows 7–8 filled by this rectangle) → clears ## Summary | # | Piece | Cells | Lines cleared | |---|-------|-------|---------------| | 1 | T | (4,4),(5,3),(5,4),(6,4) | - | | 2 | B | (7,1)–(7,5) | Row 7, Col 3 | | 3 | R | (7,3),(7,4),(7,5),(8,3),(8,4),(8,5) | Col 4 | **TOTAL LINES CLEARED: 3** openai/gpt-5.6-sol-pro SOLVED · optimal119s · 15074 out tok (14373 reasoning) · $0.5553 Board transcription (`#` filled, `.` empty): ```text 12345678 1 ####...# 2 ####.##. 3 ####.#.# 4 #.#..### 5 ........ 6 ###..##. 7 .....### 8 .##..### ``` Placements: 1. **Piece T**: `(4,4), (5,3), (5,4), (6,4)` - Clears: none 2. **Piece B**: `(7,1), (7,2), (7,3), (7,4), (7,5)` - Clears: **row 7** and **column 3** 3. **Piece R**: `(7,3), (7,4), (7,5), (8,3), (8,4), (8,5)` - Clears: **column 4** **TOTAL LINES CLEARED: 3** google/gemini-3.1-pro-preview SOLVED · optimal122s · 15985 out tok (15356 reasoning) · $0.1947 1) BOARD TRANSCRIPTION (8x8 grid): (Rows 1-8 top to bottom, Columns 1-8 left to right. '#' is filled, '.' is empty) Row 1: # # # # . . . # Row 2: # # # # . # # . Row 3: # # # # . # . # Row 4: # . # . . # # # Row 5: . . . . . . . . Row 6: # # # . . # # . Row 7: . . . . . # # # Row 8: . # # . . # # # 2) SOLUTION: The board currently lacks a contiguous 2x3 empty space to place Piece R. We must clear lines to create space. Row 5 is empty, and Row 7 has exactly 5 empty cells horizontally which suggests a row clear. Column 3 only needs two cells filled to clear (at Row 5 and Row 7). We can execute a combo that clears both Row 7 and Column 3 to open up the needed space for our 2x3 piece! **Placement 1:** - **Piece:** T (vertical column of 3 on the right, 1 sticking out on the left) - **Location:** Place the vertical bar in Column 4 at Rows 4, 5, 6, with the stick part pointing into Row 5, Column 3. - **Cells occupied:** (4,4), (5,4), (6,4) and (5,3). - **Clears:** No lines clear yet, but we've now plugged the hole in Column 3 at Row 5. **Placement 2:** - **Piece:** B (1x5 horizontal bar) - **Location:** Place it flat along the bottom-left gap in Row 7. - **Cells occupied:** (7,1), (7,2), (7,3), (7,4), (7,5). - **Clears:** Placing this piece fills the last remaining gap in **Column 3** and completes **Row 7**, causing *both* to clear at the same time! Row 7 completely empties out, and Column 3 (including the block blocking Row 8) is entirely erased. **Placement 3:** - **Piece:** R (2x3 solid rectangle) - **Location:** Thanks to combining the column and row clears, an exact 2x3 pocket is now newly opened bridging Rows 7 and 8 at Columns 3, 4, and 5. - **Cells occupied:** (7,3), (7,4), (7,5), (8,3), (8,4), (8,5). - **Clears:** No lines clear on this final placement. 3) TOTAL LINES CLEARED: 2 anthropic/claude-opus-4.8 did not solve27s · 2030 out tok (0 reasoning) · $0.0607 I'll transcribe the board carefully from the image. ## 1) Board Transcription Reading the 8x8 grid (# = filled, . = empty): ``` Row 1: # # # # . . . # Row 2: # # # # . # # . Row 3: # # # # . # . # Row 4: # . # . . # # # Row 5: . . . . . . . . Row 6: # # # . . # # . Row 7: . . . . . # # # Row 8: . # # . . # # # ``` Filled cells: - R1: 1,2,3,4,8 - R2: 1,2,3,4,6,7 - R3: 1,2,3,4,6,8 - R4: 1,3,6,7,8 - R5: (empty) - R6: 1,2,3,6,7 - R7: 6,7,8 - R8: 2,3,6,7,8 ## 2) Analysis and Solution Let me look for near-complete lines. **Row 4** has filled: 1,3,6,7,8. Empty: 2,4,5. **Column 6** has filled: 2,3,4,6,7,8. Empty: 1,5. **Row 1** has filled: 1,2,3,4,8. Empty: 5,6,7. Let me try to complete **Row 1**: empty cells are (1,5),(1,6),(1,7). Piece B is a 1x5 bar - too long. Try **Piece T** (T-tetromino, orientation: three across top, one below middle) to fill Row 4 gaps. Let me place **Piece R (2x3)** at rows 4-5, cols 3-5? (4,3) is filled. No. **Place Piece B (1x5)** on row 5, cols 1-5 → (5,1)(5,2)(5,3)(5,4)(5,5). All empty. ✓ **Place Piece R (2x3)** at rows 4-5, cols... row 5 now partly filled. Let me place R at rows 6-7? Row6 empty at 4,5,8; Row7 empty at 1,2,3,4,5. Place R at (6,3)? (6,3) filled. Place R at rows 7-8 cols 3-5: (7,3)(7,4)(7,5)(8,3)(8,4)(8,5) - (8,3) filled. No. Let me place **R at rows 4-5, cols... ** need 2x3 all empty. Row4 empty:2,4,5; Row5 empty:6,7,8 (after B). Cols 6-8 rows4-5: (4,6)(4,7)(4,8) filled. No good spot easily. **Place T** to complete a column. Column 4 filled: 1,2,3. Empty everywhere else. Let me focus on achievable clears: **Place B (1x5) at row 5, cols 1-5.** Row 5 now: 1,2,3,4,5 filled. **Place T at row 5-6:** T shape = (r,c),(r,c+1),(r,c+2),(r+1,c+1). Place top row at (5,6)(5,7)(5,8) and (6,7). Check: (5,6)(5,7)(5,8) empty ✓, (6,7) filled ✗. Try T top at (5,5)(5,6)(5,7)+(6,6): but (5,5) will be filled by B. Order: place T first. (5,5)(5,6)(5,7) empty, (6,6) filled ✗. **Place B at row 5, cols 4-8:** (5,4)(5,5)(5,6)(5,7)(5,8) all empty ✓. Then **place T top at (5,1)(5,2)(5,3)+(6,2):** (6,2) filled ✗. **Place B at row 5 cols 1-5**, then need cols 6,7,8 in row 5. Place T top (5,6)(5,7)(5,8), stem (6,7) filled ✗. Row 5 is hard to complete fully. Let me settle for a clean legal placement: - **Piece B:** row 5, cols 1–5: (5,1)(5,2)(5,3)(5,4)(5,5) - no clear. - **Piece R:** rows 6–7, cols 3–5: (6,3)filled ✗. Use rows 7-8 cols... difficult. - **Piece R:** rows 5? occupied. Place R at rows 6-7 cols 4-... (6,4)(6,5)(7,4)... only need clear cols. Place R at (6,3): filled. **Final chosen placements (all legal, no clears possible for full lines):** - **Piece B:** (5,1)(5,2)(5,3)(5,4)(5,5) - **Piece T:** (6,4)(6,5) top... T needs 3 wide. Top (5,6)? taken. Place T top (7,3)(7,4)(7,5)+(8,4): (8,3),(8,4)... (7,3),(7,4),(7,5) empty ✓, (8,4) empty ✓. - **Piece R:** rows 6-7 cols... (6,4)(6,5) row6 empty, row7 (7,4)(7,5) now filled by T. Place R at (1,5)(1,6)(1,7)(2,5)... (2,5)empty, but only 2 rows: (1,5)(1,6)(1,7)(2,5)(2,6)filled ✗. Place R at (3,5)? no. ## Final Solution 1. **Piece B (1x5):** (5,1)(5,2)(5,3)(5,4)(5,5) - no line clears yet. 2. **Piece T:** (7,3)(7,4)(7,5)+(8,4) - no clears. 3. **Piece R (2x3):** rows 4-5 impossible; place at (1,5)(2,5) region invalid. Use **(6,4)(6,5) + need col** → place at rows 6-7? blocked. Given the board's constraints, the legal placements above complete without forming any full row or column. **TOTAL LINES CLEARED: 0** x-ai/grok-4.5 did not solve592s · 44826 out tok (44592 reasoning) · $0.2722 step2 R: ILLEGAL at \[(4, 3), (4, 4)\] **Board transcription (rows 1–8 top→bottom, # = filled, . = empty):** ``` # # # # . . . # # # # . . # # . # # # . . # # # # # # . . # # # . . . . . . . . # # # # . # # . . . . . . # # # . # . . . # # # ``` **Solution (order T → R → B):** 1. **Piece T** occupies (3,5), (4,4), (4,5), (5,5). Clears: row 4. 2. **Piece R** occupies (4,2), (4,3), (4,4), (5,2), (5,3), (5,4). Clears: none. 3. **Piece B** occupies (7,1), (7,2), (7,3), (7,4), (7,5). Clears: row 7, column 2. **TOTAL LINES CLEARED: 3** qwen/qwen3.7-plus did not solve355s · 19537 out tok (18262 reasoning) · $0.0255 step3 R: INVALID cell count 3!=6 1) **Board Transcription:** Here is the 8x8 grid state (# = filled, . = empty): Row 1: # # # # . . . # Row 2: # # # # . # # . Row 3: # # # # . # . # Row 4: # . # . . # # # Row 5: . . . . . . . . Row 6: # # # . . # # . Row 7: . . . . . # # # Row 8: . # # . . # # # 2) **Solution:** * **Placement 1: Piece B (1x5 horizontal bar)** * Place at Row 7, Columns 1-5. Cells: (7,1), (7,2), (7,3), (7,4), (7,5). * Row 7 was previously `. . . . . # # #`. With the piece, it becomes `# # # # # # # #`. * **Result:** Row 7 is full and clears. All cells in Row 7 become empty. * **Placement 2: Piece T (T-shaped tetromino)** * The shape is effectively a vertical bar of 3 with a block to the left of the middle: `(0,1), (1,0), (1,1), (2,1)` relative to top-left. * Place with top-left anchor at (4,4). Cells occupied: (4,5), (5,4), (5,5), (6,5). * These cells were empty. * **Result:** No immediate lines clear, but it fills gaps in the center. * **Placement 3: Piece R (2x3 rectangle)** * Place at Row 5, Columns 1-3 and Row 6... wait, Row 6 is filled there. * Let's look for the best fit. The large empty space is Row 5. * Actually, with Row 7 cleared, we have space. * Let's place Piece R at (5,6), (5,7), (5,8) and (4,6)... no (4,6) is filled. * Given the constraints and the "no rotation" rule, the puzzle implies a specific fit. The most logical fit for a 2x3 block in this specific "Block Blast" configuration (where Row 5 is empty and Row 6 has gaps) is often intended to bridge gaps, but strictly speaking, with the board as transcribed, a 2x3 horizontal piece is very hard to place. * However, assuming standard puzzle logic where you clear the most lines: * If we place Piece B at (7,1)-(7,5), Row 7 clears. * Total lines cleared: 1. *(Self-Correction/Refinement for the final output format)*: The prompt asks to clear as many as possible. Move 1: Place Piece B (1x5) at (7,1) to (7,5). - Row 7 becomes completely filled (cells 6,7,8 were already filled). - **Row 7 clears.** Move 2: Place Piece T at (4,4) relative anchor -> cells (4,5), (5,4), (5,5), (6,5). - No lines clear immediately. Move 3: Place Piece R (2x3). - This is the tricky part. With Row 7 cleared, Row 7 is empty. Row 8 has (8,4),(8,5) empty. - There isn't a perfect 2x3 slot visible without rotation. However, in these games, often the "2x3" can be the 3x2 if the image is ambiguous, but the text says "orientation shown". - If we assume the large empty Row 5 allows placement: Place at (5,1)-(5,3) is 1x3. - Let's assume the question implies a valid solution exists. The only full line clear is Row 7. TOTAL LINES CLEARED: 1 (Row 7). Wait, let's look at Col 6. Filled: 2,3,4, 6,7,8. (6 blocks). If I place Piece B at (5,2)-(5,6). (5,6) fills Col 6. Col 6 has 7 blocks. Missing (1,6). If I place Piece T to fill (1,6)? Shape: top (1,6), mid (2,5),(2,6-filled). No. I will stick with the Row 7 clear. **Final Answer Formulation:** 1. Board grid. 2. Placements. 3. Total lines. Placement 1: Piece B at (7,1), (7,2), (7,3), (7,4), (7,5). Row 7 clears. Placement 2: Piece T at (4,5), (5,4), (5,5), (6,5). (Fits in empty space). Placement 3: Piece R at (5,1), (5,2), (5,3) ... wait, 2x3. Let's place Piece R at (5,6), (5,7), (5,8) and (6,6)... (6,6) is filled. I will assume Piece R is placed at (5,1)-(5,3) and (4,1)-(4,3) is impossible. I'll just list the clear. TOTAL LINES CLEARED: 1 mistralai/mistral-small-2603 did not solve4s · 186 out tok (0 reasoning) · $0.0004 step1 T: ILLEGAL at \[(4, 3), (6, 3)\]; step2 R: ILLEGAL at \[(6, 6), (6, 7), (7, 6), (7, 7)\] 1) Board transcription (filled cells as '#', empty as '.'): ...####.. .#.#.##. .#..#### ....#### #######. .#.#.... .#.##### ..###### 2) Solution: - Place Piece T at (4,3) covering (4,3), (5,3), (5,4), (6,3). No rows/columns clear. - Place Piece R at (6,5) covering (6,5), (6,6), (6,7), (7,5), (7,6), (7,7). Rows 6 and 7 clear. - Place Piece B at (7,1) covering (7,1), (7,2), (7,3), (7,4), (7,5). Column 5 clears. TOTAL LINES CLEARED: 3 anthropic/claude-haiku-4.5 did not solve10s · 1043 out tok (0 reasoning) · $0.0071 step2 R: ILLEGAL at \[(3, 6), (4, 6), (4, 7)\] I'll analyze the board and solve this step-by-step. ## BOARD TRANSCRIPTION ``` . . . . . . . . # # # # . . . # # # # # . # # . # # # # . # . # # . # . . # # # . . . . . . . . # # # . . # . . . . . . . # # # . # # . . # # # ``` Where # = filled, . = empty Filled cells: - Row 1: (1,1), (1,2), (1,3), (1,4), (1,8) - Row 2: (2,1), (2,2), (2,3), (2,4), (2,6), (2,7) - Row 3: (3,1), (3,2), (3,3), (3,4), (3,6), (3,8) - Row 4: (4,1), (4,3), (4,6), (4,7), (4,8) - Row 5: empty - Row 6: (6,1), (6,2), (6,3), (6,6) - Row 7: (7,2), (7,3), (7,6), (7,7), (7,8) - Row 8: (8,1), (8,2), (8,6), (8,7), (8,8) ## PIECE SPECIFICATIONS - **Piece T** (T-tetromino): appears to be positioned as shown - **Piece R** (2×3 rectangle): 6 cells - **Piece B** (1×5 bar): 5 cells ## SOLUTION **1st Placement: Piece B (1×5 bar horizontal)** - Place at row 5, columns 1-5: (5,1), (5,2), (5,3), (5,4), (5,5) - No lines clear yet **2nd Placement: Piece R (2×3 rectangle)** - Place at rows 5-6, columns 6-8: (5,6), (5,7), (5,8), (6,7), (6,8), (6,4) - Actually, let me reconsider: (5,6), (5,7), (5,8), (6,6), (6,7), (6,8) - but (6,6) is filled - Try rows 5-6, columns 5-7: (5,5), (5,6), (5,7), (6,5), (6,6), (6,7) - (6,6) is filled - Place at rows 6-7, columns 4-6: (6,4), (6,5), (6,6), (7,4), (7,5), (7,6) - but (6,6) filled - Place at rows 5-6, columns 2-4: (5,2), (5,3), (5,4), (6,2), (6,3), (6,4) - (5,2-4) will be filled by B, conflicts - Place at rows 6-7, columns 7-8 with proper orientation: rows 5-6, cols 1-3: (5,1), (5,2), (5,3), (6,1), (6,2), (6,3) - conflicts - Try row 5, cols 6-8 and row 6, cols 6-8: (5,6), (5,7), (5,8), (6,6), (6,7), (6,8) - (6,6) filled - Place at (5,2), (5,3), (5,4), (5,5), and extend... reconsidering layout **Revised Solution:** 1. **Piece R**: (5,6), (5,7), (5,8), (6,5), (6,7), (6,8) - invalid, not rectangular After careful analysis: 1. **Piece B** at row 5, cols 1-5 2. **Piece R** at rows 3-4, cols 5-7 3. **Piece T** at row 2, cols 5-7 with stem **TOTAL LINES CLEARED: 0** google/gemini-3.1-flash-lite did not solve5s · 439 out tok (0 reasoning) · $0.0010 step1 R: ILLEGAL at \[(6, 1), (6, 2), (6, 3)\]; step3 T: ILLEGAL at \[(8, 2)\] ### 1) The Board (8x8) Representing filled cells as '#' and empty cells as '.': ``` Row 1: # # # # . . . # Row 2: # # # # . # # # Row 3: # # # # . # . # Row 4: # # # . . # # # Row 5: . . . . . . . . Row 6: # # # . . # # . Row 7: . . . . . # # # Row 8: . # # . . # # # ``` *(Note: Rows 1-4 and 6-8 have blocks based on the visual layout provided.)* --- ### 2) Solution **Step 1: Place Piece R (2x3 rectangle)** * **Placement:** Place on the board covering (5,1), (5,2), (5,3), (6,1), (6,2), (6,3). * **Result:** No lines are cleared. **Step 2: Place Piece B (1x5 bar)** * **Placement:** Place horizontally on row 5, columns 4 through 8: (5,4), (5,5), (5,6), (5,7), (5,8). * **Result:** This fills Row 5 completely. Row 5 clears. **Step 3: Place Piece T (T-shape)** * **Placement:** Place such that the top of the T is at (7,1), (7,2), (7,3) and the stem is at (8,2). * **Result:** This piece occupies (7,1), (7,2), (7,3), and (8,2). No lines are cleared. *(Note: Given the board constraints and the pieces provided, multiple arrangements are possible, but this path avoids conflicts with existing blocks.)* --- ### TOTAL LINES CLEARED: 1 x-ai/grok-4.3 · low did not solve13s · 1057 out tok (827 reasoning) · $0.0046 step1 R: ILLEGAL at \[(6, 1), (6, 2), (6, 3)\]; step3 B: INVALID cell count 4!=5 1) Board transcription (8x8 grid, row 1=top, #=filled): ``` # # # # . . . # # # # # . # # . # # # . . # # # . # # # . # . . . . . . . . . . # # # . . # . . . . . . . . . . # . # . # # # # ``` 2) Solution (order chosen to maximize clears): - Place R first at cells (6,1)(6,2)(6,3)(7,1)(7,2)(7,3) → no lines clear - Place T next at cells (4,2)(5,1)(5,2)(5,3) → row 5 clears - Place B last at cells (8,5)(8,6)(8,7)(8,8) → row 8 and col 6 clear 3) TOTAL LINES CLEARED: 3 (The three placements are all fully legal on empty cells; the two subsequent clears free additional space but are not required for legality.) openai/gpt-5.6-luna did not solve50s · 6085 out tok (5696 reasoning) · $0.0386 1) **Transcribed board** (`#` = filled, `.` = empty): ```text Row 1: ####...# Row 2: ####.##. Row 3: ####.#.# Row 4: #.#.#### Row 5: ........ Row 6: ###..##. Row 7: .....### Row 8: .##..### ``` Filled cells: - Row 1: (1,1), (1,2), (1,3), (1,4), (1,8) - Row 2: (2,1), (2,2), (2,3), (2,4), (2,6), (2,7) - Row 3: (3,1), (3,2), (3,3), (3,4), (3,6), (3,8) - Row 4: (4,1), (4,3), (4,6), (4,7), (4,8) - Row 5: none - Row 6: (6,1), (6,2), (6,3), (6,6), (6,7) - Row 7: (7,6), (7,7), (7,8) - Row 8: (8,2), (8,3), (8,6), (8,7), (8,8) No valid three-piece solution exists for this transcription. In particular, the 2×3 rectangle has no legal placement initially. Placing the T or the 1×5 bar does not clear a line in a way that creates a legal rectangle placement; placing the bar on row 7 clears that row, but the rectangle still has no legal location. **TOTAL LINES CLEARED: 0** x-ai/grok-4.3 · high did not solve170s · 20681 out tok (20476 reasoning) · $0.0537 step1 T: ILLEGAL at \[(4, 6)\]; step2 B: ILLEGAL at \[(6, 1), (6, 2), (6, 3)\]; step3 R: ILLEGAL at \[(3, 1), (3, 2), (3, 3), (4, 1), (4, 3)\] 1) Board transcription (8x8 grid, #=filled, .=empty): Row1: # # # # . . . # Row2: # # # # . # # # Row3: # # # # . # # # Row4: . . . . . . . . Row5: # # # . . # # Row6: . . . . . # # # Row7: . # # . # # # # Row8: . # # . # # # # 2) Solution: - Piece T at (3,5),(4,4),(4,5),(4,6); clears row 3 - Piece B at (6,1),(6,2),(6,3),(6,4),(6,5); clears row 6 - Piece R at (3,1),(3,2),(3,3),(4,1),(4,2),(4,3); clears none TOTAL LINES CLEARED: 2 mistralai/mistral-large-2512 did not solve12s · 665 out tok (0 reasoning) · $0.0019 step1 B: ILLEGAL at \[(6, 2), (6, 3), (6, 6)\]; step2 R: ILLEGAL at \[(1, 8), (2, 6), (2, 7)\]; step3 T: ILLEGAL at \[(2, 2), (2, 3), (2, 4), (3, 3)\] Let's start by transcribing the board based on the image. Filled cells are represented by `#` and empty cells by `.`: ``` Row 1: ......#. Row 2: ..#...#. Row 3: ..#..... Row 4: ....#... Row 5: ....#... Row 6: #....... Row 7: ..#..... Row 8: ......#. ``` Now, let's place the pieces one by one to maximize line clears. **Step-by-Step Placement:** 1. **Piece B (1x5 horizontal bar):** Place at (6, 2)-(6, 6): - Cells: (6,2), (6,3), (6,4), (6,5), (6,6) - This fills row 6 completely, clearing it. - After clearing, row 6 becomes empty. Updated board after clearing row 6: ``` Row 1: ......#. Row 2: ..#...#. Row 3: ..#..... Row 4: ....#... Row 5: ....#... Row 6: ......... Row 7: ..#..... Row 8: ......#. ``` 2. **Piece R (2x3 rectangle):** Place at (1, 6)-(2, 8): - Cells: (1,6), (1,7), (1,8), (2,6), (2,7), (2,8) - This fills column 8 completely, clearing it. - After clearing, column 8 becomes empty. Updated board after clearing column 8: ``` Row 1: ......#. Row 2: ..#..... Row 3: ..#..... Row 4: ....#... Row 5: ....#... Row 6: ......... Row 7: ..#..... Row 8: ......#. ``` 3. **Piece T (T-shaped tetromino):** Place at (2, 2)-(2, 4) with the stem at (3, 3): - Cells: (2,2), (2,3), (2,4), (3,3) - No additional rows or columns clear. Final board: ``` Row 1: ......#. Row 2: .###.... Row 3: ..#.#... Row 4: ....#... Row 5: ....#... Row 6: ......... Row 7: ..#..... Row 8: ......#. ``` **Solution Summary:** 1. Piece B placed at (6,2)-(6,6), clearing row 6. 2. Piece R placed at (1,6)-(2,8), clearing column 8. 3. Piece T placed at (2,2), (2,3), (2,4), (3,3), no additional clears. **TOTAL LINES CLEARED: 2** moonshotai/kimi-k2.6 no answer produced232s · 16000 out tok (14380 reasoning) · $0.0657 no visible answer - reasoning ran to token cap (finish=length) *(empty - the model spent its entire token budget on internal reasoning and returned no answer.)* z-ai/glm-5v-turbo did not solve109s · 12173 out tok (11431 reasoning) · $0.0507 step1 R: ILLEGAL at \[(6, 2), (6, 3)\]; step2 T: ILLEGAL at \[(3, 2), (4, 3)\]; step3 B: ILLEGAL at \[(8, 2), (8, 3)\] **1) Board Transcription** (Rows 1-8, Columns 1-8. `#` = Filled, `.` = Empty) ``` # # # # . . . # # # # # . # # # # # # . # . # # . # . . # # # # # # . . # # . . . . . . . . . . . . . # # # . . # # . # # # ``` **2) Solution Placements** **Move 1: Place Piece R (2x3 Rectangle)** * **Position:** Place the rectangle with its top-left corner at **(6, 2)**. * **Cells Occupied:** (6, 2), (6, 3), (6, 4), (7, 2), (7, 3), (7, 4). * **Reasoning:** This places blocks into Column 3 at rows 6 and 7. Column 3 previously contained blocks at rows 1, 2, 3, 4, 5, and 8. Filling the missing (6,3) and (7,3) completes the column. * **Lines Cleared:** **Column 3**. *(Board Update: Column 3 is now empty. Blocks (6,2), (6,4), (7,2), (7,4) remain.)* **Move 2: Place Piece T (T-Tetromino)** * **Position:** Orient the piece with the stem pointing **UP**. Place it so the stem is at **(3, 3)** and the base rests on Row 4. * **Cells Occupied:** (3, 3) [Stem], (4, 2), (4, 3), (4, 4) [Base]. * **Reasoning:** * The cell **(4, 2)** fills the only gap in **Column 2** (which had blocks at rows 1, 2, 3, 5, 6, 7, 8 after Move 1). This clears **Column 2**. * Simultaneously, the base of the T fills the gap in **Row 4**. Row 4 had blocks at (4, 1) and (4, 6)-(4, 8). Adding (4, 2), (4, 3), and (4, 4) completes the row. * **Lines Cleared:** **Column 2**, **Row 4**. *(Board Update: Column 2 and Row 4 are empty. Block (3, 3) remains.)* **Move 3: Place Piece B (1x5 Bar)** * **Position:** Place the bar horizontally starting at **(8, 1)**. * **Cells Occupied:** (8, 1), (8, 2), (8, 3), (8, 4), (8, 5). * **Reasoning:** Previous clears removed blocks from (8, 2) and (8, 3). Row 8 currently has blocks at (8, 6), (8, 7), (8, 8). Placing this bar fills the remaining 5 empty spots on the left side of the row. * **Lines Cleared:** **Row 8**. **3) TOTAL LINES CLEARED: 4** minimax/minimax-m3 no answer produced397s · 40000 out tok (24948 reasoning) · $0.0481 no visible answer - reasoning ran to token cap (finish=length) *(empty - the model spent its entire token budget on internal reasoning and returned no answer.)* Method: all models called through OpenRouter (one uniform client), single non-streaming attempt, identical PNG + prompt, `max_tokens=16000` (40k on the MiniMax retry). Ground truth from a deterministic solver; every stated move replayed through the same validator. DeepSeek excluded - it has no vision model. USD→GBP at 0.79\. Total spend $2.18. ### Puzzled by AI URL: https://demystifai.info/puzzled-by-ai/ Last updated: 2026-08-05T08:20:47.000Z **Puzzled by AI** AI & Data · demystifAI 16 July 2026 Romain Thierry AI & Data · demystifAI # Puzzled by AI I gave sixteen of the most capable AI models a Block Blast! puzzle, and found only three managed to solve it. In this post, we look more closely at what worked, and what did not. **What do you do while the machine is thinking?** A serious request to an LLM takes minutes to complete, and we humans have our own ways to occupy this idle time. Personally, I tend to run several sessions in parallel, so there is already something ready to look at, but let’s be honest, sometimes I also indulge in a bit of Block Blast! - these games have a capacity to burn through time at alarming speed... So it was only natural that I ended up wondering how my trusty Claude would handle it. And while I was at it, why not test other models too? **So I ran the experiment.** The task was a single screenshot from the game: an eight-by-eight grid, partly filled, with three shapes waiting to be dropped into the empty squares. This was accompanied by a simple prompt asking the model whether it could solve the puzzle. The results turned out to be more interesting than I was expecting. Prefer the raw numbers first? [See the full data - every model, every score, and what each attempt cost →](https://demystifai.info/sixteen-models-one-block-puzzle/) First of all, only **three of the sixteen solved the puzzle**. These are the models that sit at the top of the public benchmarks, and the puzzle is not difficult, so we have to dig deeper to find out what was happening here. Second finding, no model came unstuck on the logic. Most set out careful, orderly reasoning and arrived at answers that were composed, plausible in tone and sometimes against the rules: pieces laid on squares already occupied, lines declared complete that were not, etc. Further investigation showed nearly every mistake began in the same place: **the model had misread the board**, and having misread it, reasoned impeccably but from the wrong starting point. This is why asking your favourite model to read your prompt and replay it to you with an outline of what it plans to do is always a good idea. The key take-aways from this experiment: - **The handful of premium models performed best.** The three that solved it - Claude Fable 5, GPT-5.6 sol-pro and Gemini 3.1 Pro - are the ones the market already rates as the strongest frontier models, so their reputation is not built on hype. - **The same right answer came at very different prices.** Gemini solved it for 19 cents, GPT-5.6 sol-pro for 56 and Claude Fable 5 for 80 - a factor of four between the cheapest and the dearest of the winners. The dearest failure, for its part, cost more than the cheapest success. Does solving the puzzle cost more? Each of the sixteen models, by the cost of its one attempt and how many of the three pieces it placed legally. solved - all three placed 0 1 2 3 pieces placed legally $0.001 $0.01 $0.10 $1.00 cost of the one attempt (US$, log scale) nothing cheaper solved it Grok 4.5 - dearest failure returned no answer Solved it - and what it cost Gemini 3.1 Pro $0.19 GPT-5.6 sol-pro $0.56 Fable 5 $0.80 solved it did not solve returned no answer - **Token counts tell their own story.** One of the three winners used twice the tokens of the other two for the same answer, and the biggest token spender of the run still failed - more on that below. Two models used their entire budget on internal reasoning and returned nothing at all; a tool that runs long and comes back empty is its own kind of production risk. Does solving the puzzle take more tokens? Each of the sixteen models, by the tokens it used and how many of the three pieces it placed legally. solved - all three placed 0 1 2 3 pieces placed legally 0 10k 20k 30k 40k tokens used in the attempt (input + output) Grok 4.5 - most tokens, still failed returned no answer Solved it - tokens used Gemini 3.1 Pro 17,400 Fable 5 17,600 GPT-5.6 sol-pro 36,900 solved it did not solve returned no answer **Grok 4.5 was the biggest surprise here.** I had high expectations for Grok 4.5, which seems to be highly rated as a software engineering partner. In true journalistic fashion, I invited Grok to provide a statement to accompany this publication - demolition jobs are not our business here. > From Grok: “Vision capabilities combined with precise spatial reasoning remain a genuine weak spot for many frontier models - including Grok 4.5 in this evaluation. Most failures stemmed not from flawed high-level strategy, but from board transcription errors and misaligned piece placements, highlighting how easily models can hallucinate grid layouts or piece orientations. > > Grok 4.5, for instance, consumed over 44,500 reasoning tokens and approximately ten minutes of processing time, yet still proposed an illegal overlapping placement. This stands in contrast to the three models that solved the puzzle successfully (Claude Fable 5, GPT-5.6-sol-pro, and Gemini 3.1 Pro-preview), which achieved the optimal three-line clear with more efficient and accurate visual grounding. While Grok 4.5 excels in many coding, agentic, and cost-efficiency scenarios, multimodal puzzles of this nature expose gaps that pure text-based reasoning chains do not always bridge. > > Benchmarks like this are valuable: xAI is iterating rapidly with new models released monthly, and real-world evaluations such as demystifAI’s help sharpen our focus on improving vision grounding, structured validation, and reliable spatial understanding.” Credit where credit is due: Grok’s statement is a fair diagnosis. It names the actual failure, transcription rather than strategy. It also comes from a company that has been catching up with the leaders of the field at remarkable speed, and on that trajectory the gaps this puzzle exposed are the kind that close quickly. We look forward to the improvements to come. **None of this makes the models less useful.** It makes them useful in a narrower way than the demonstrations suggest: reliable where a person can check the work, and when it can ground itself (this is the strange jargon we now employ) in good data well beyond what a human brain can absorb. This little experiment gives us two things heavy model users instinctively learned: i) don’t trust blindly: when it feels too good to be true it is easy to get carried away (the subject of a future paper, by the way); ii) best practices: there is still an emerging body of knowledge, but the foundations have already been laid (also for another paper). Are you, like me, surprised by these results? What would you have expected? Perhaps you have run your own experiment and come to a different conclusion? Let me know in the comments below. The full data behind this experiment - all sixteen models, scored and costed - is here: [Block Blast experiment, the data →](https://demystifai.info/sixteen-models-one-block-puzzle/) demystifAI · Puzzled AI · AI & Data · 16 July 2026 demystifai.info AI, in control.