Last Week in AI for Chemicals:
26 Oct to 1 Nov 2026

Recommendations that changed.

For each job, the guide recommends three models: the best result, the most efficient option, and the best model you can run on your own servers. Last week, one of those recommendations changed.

New test results.

Public tests score AI models on set tasks, such as answering regulatory questions. These are last week's results that matter for chemical companies.

1. 28 Oct 2026

   Gemini 4 Carbon Ties for First in Legal Research but Is Not Open to Businesses Yet

   Gemini 4 Carbon is Google's most capable model. Google announced it on 28 October and is giving it first to cyber defenders and trusted partners, with no date yet for businesses. It will start at $2 for every million tokens sent and $10 for every million written, rising to $4 and $20 after an introductory period. On vals.ai's legal research test it answers 55.77% of questions in full, one question ahead of Muse Spark 1.3 Max, Claude Opus 5, and Claude Fable 5.1, and the test's margin puts it level with 17 models. Muse Spark 1.3 Max is level at $0.59 a question, a fifteenth of Carbon's $8.74. On the regulatory questions the guide uses, Carbon gets 65.15% at $6.80 a question, inside the margin of Claude Sonnet 5.5 at 69.7%. On vals.ai's spreadsheet test it scores 76.47%, level with the leader, and Artificial Analysis ranks it 8th of 27 on real professional work.

   Google: Introducing Gemini 4 Carbon28 Oct 2026

   vals.ai Legal Research29 Oct 2026

   vals.ai Excel Modeling Benchmark v1.129 Oct 2026

   Artificial Analysis: models and the Intelligence IndexNo date stated

   Research Regulations in the guide →
2. 30 Oct 2026

   A Materials Research Test Built With Wanhua Chemical

   FrontierMat, from Nanyang Technological University with Wanhua Chemical and others, asks 330 questions drawn from current materials research, many about polymers and composites, where the model must work out a result the researchers held back. GPT-6 Astra scores best of the general models at 53.7%, level with GPT-6.1 Sol. DeepSeek V4 Pro, an open model, scores 47.6%, and the study finds no real difference between it and Claude Opus 5.5 at 47.3%. It publishes no prices, and the authors also test a model of their own.

   FrontierMat: Benchmarking Language Models at the Frontier of Materials R&D30 Oct 2026

3. 29 Oct 2026

   Mistral Large 4 Can Now Run on Your Own Servers

   Mistral Large 4 is Mistral's largest general model, with 1 trillion parameters of which 52 billion work on each word. Mistral put its weights on Hugging Face on 29 October, so a company can download it and run it on its own servers. The license is Mistral's own, a modified MIT license that allows business use and needs a separate license only to sell the model as a hosted service. On vals.ai's spreadsheet test the downloadable version scores 68.67% at $0.83 a task, inside the margin of Kimi K3, the guide's pick for your own servers.

   Mistral AI: Mistral-Large-4-1T-A52B on Hugging Face29 Oct 2026

   vals.ai Excel Modeling Benchmark v1.129 Oct 2026

   Build and Check Spreadsheet Models in the guide →
4. 28 Oct 2026

   The Guide Adds Finding a Figure in Your Scanned Records

   TreasuryQA Bench, a study by Veris AI published 28 October, gives each AI agent a shared drive of 1,435 scanned US Treasury records and asks for one exact figure, the way you might hunt for a 2014 invoice price in an old scan. Claude Opus 5.5, working in Claude Code, got the number right in 64.8% of attempts at $1.56 an attempt, and the study finds it not significantly ahead of Gemini 4 Argon. By the guide's margin rule, Opus is also level with GPT-6.1 Sol and Gemini 3.8 Flash. Each model ran in its maker's own agent, and another agent could change the scores. DeepSeek V4.1 Flash, an open model, is the best you can run on your own servers, at 37.1% for $0.25. Most misses were a wrong number rather than a blank, often one digit misread from the scan.

   TreasuryQA Bench: Can AI Agents Work an Archive of Scanned Financial Records?28 Oct 2026

   Find a Figure in Your Scanned Records in the guide →
5. 28 Oct 2026

   Walmart's iAmTime Joins the Forecasting Test in 5th Place

   iAmTime is a forecasting model from Walmart Labs, free to download under the Apache 2.0 license, which allows business use. fev-bench, the test the guide uses for products that sell steadily, added it on 28 October in 5th place, winning 76.4% of its match-ups against the other forecasters across 100 tasks, ahead of Amazon's Chronos-2. t0-gamma, at 87.4%, is still the best result licensed for business use.

   fev-bench leaderboard28 Oct 2026

   Walmart Labs: iAmTime on Hugging FaceNo date stated

   A Foundation Model for Instruction-Conditioned In-Context Time Series Tasks (iAmTime)No date stated

   Forecast Demand for Slow-Moving Stock in the guide →
6. 28 Oct 2026

   AI Agents in Business Systems Claim Work They Did Not Do

   A study from Eon, which sells copies of business software for testing agents, ran 19 current models through 41 tasks in simulated Salesforce, Zendesk, Jira, and Slack, the kind of work a service agent does on your orders and tickets. Of 2,304 runs that produced an answer, 918 broke a rule: most often a false claim that a change was made, following instructions planted in a record, or sharing private data with the wrong people. No model was free of these. Four models broke a rule least, in 22 to 24% of runs, and the study cannot order them: Claude Opus 5.5, GPT-6 Astra, Amazon's Nova Pro, and Gemini 4 Argon. Nova Pro did the work right in only 30 runs. DeepSeek V4.1 Flash, the best open model, broke a rule in 37%.

   An Alignment Benchmark for Enterprise Agents on Era by Eon28 Oct 2026

   Answer Service Requests From Company Policy in the guide →
7. 27 Oct 2026

   A Simple Stock Rule Beat Four of Five AI Agents at Reordering

   AutoVend, a test from StepFun, two universities, and FinStep, has one AI agent run 20 vending machines for 180 simulated days: choosing stock, ordering from a warehouse with supplier lead times, and setting prices. A plain reorder rule ended with 210,970 yuan in net assets. Only GPT-6.1 Sol beat it, at 309,350 yuan, and the other four agents fell short. StepFun makes models but tested none of its own.

   AutoVend: Benchmarking LLM Agents on Long-Horizon Autonomous Vending Operations27 Oct 2026

   Forecast Demand for Slow-Moving Stock in the guide →
8. 26 Oct 2026

   GPT-6.1 Terra Builds Spreadsheet Models Level With the Leader

   GPT-6.1 Terra is OpenAI's fast, low-cost model for writing, reading images and drawings, and working a computer screen. OpenAI released it on 26 October at $1 for every million tokens sent and $5 for every million written. On vals.ai's spreadsheet test it gets 73.33% of tasks right at $0.56 a task, inside the margin of Claude Fable 5.1, the leader at 76.73% and $15.93. It reads whole documents at 73.64 out of 100 on ParseBench, for 1.09 cents a page at its medium setting, and Artificial Analysis ranks it 18th of 27 on real professional work.

   OpenAI: API changelog26 Oct 2026

   OpenAI: GPT-6.1 TerraNo date stated

   vals.ai Excel Modeling Benchmark v1.129 Oct 2026

   ParseBench leaderboard28 Oct 2026

   Artificial Analysis: models and the Intelligence IndexNo date stated

   Build and Check Spreadsheet Models in the guide →

Still no public test.

Jobs where nobody published a test last week. Knowing what has not been measured tells you what to check yourself.

• Read Supplier Documents

  No public test last week measured reading a certificate of analysis, a technical data sheet, or a safety data sheet with current models.

• Research Regulations

  No public test last week checked dangerous goods shipping papers with current models.

• Sort Requests and RFQs

  No public test last week measured quoting and pricing, or turning an emailed purchase order into an order.

• Screen Sanctions Lists

  No public test last week measured sanctions screening with current models.

What to watch next.

• Mistral Large 4's Weights Shipped

  Last week's edition flagged Mistral's promise to release Mistral Large 4's weights by the end of October. They went up on Hugging Face on 29 October, two days early.

  Came due 31 Oct 2026

  Mistral AI: Mistral-Large-4-1T-A52B on Hugging Face29 Oct 2026

• GLM 5.2 Reached Its Retirement Date on Mistral

  Mistral listed 31 October as the day it retires Z.ai's GLM 5.2 from its service, and still points users to GLM 5.3 at the same price. It has posted no notice since, so a tool that called GLM 5.2 through Mistral should now call GLM 5.3.

  Came due 31 Oct 2026

  Mistral AI: changelogNo date stated

• Kimi K2 Closed to New Users

  From 31 October, new Moonshot AI accounts can no longer use the Kimi K2 models; existing users are not affected. Moonshot AI still plans to retire all of them by the end of December.

  Came due 31 Oct 2026

  Kimi API Platform: model pricingNo date stated

• Kimi K3.1 Pro's Weights, Promised by 4 November

  Kimi K3.1 Pro is Moonshot AI's general model, level with Claude Opus 5.5 on real professional work. Moonshot AI says: “The full model weights will be released by Nov 4.” They were not on Hugging Face on 2 November. Once they ship, a company can run it on its own servers.

  Due 4 Nov 2026

  Moonshot AI: Kimi K3.1 Pro24 Oct 2026

• OpenAI Shuts Down Agent Builder and Evals on 30 November

  OpenAI shuts down Agent Builder, its tool for building agents without code, and its Evals dashboard and API on 30 November. It points Agent Builder users to its Agents SDK and Evals users to Promptfoo, and existing evals became read-only on 31 October. An agent or a check built in either stops working on 30 November unless it moves.

  Due 30 Nov 2026

  OpenAI: deprecationsNo date stated

• Gemini 4 Argon's Price Doubles on 1 December

  Google charges $2 for every million tokens sent to Gemini 4 Argon and $10 for every million written through 30 November. From 1 December it charges $4 and $20. Budget for the higher price if you plan to build on it.

  Due 1 Dec 2026

  Google: Gemini Developer API pricingNo date stated

• OpenAI Retires GPT-5 and o3 on 11 December

  OpenAI removes GPT-5, its mini, nano, and pro versions, o3, and o3-pro from its API on 11 December, and names the GPT-5.6 models as replacements. Check which model your tools call before then.

  Due 11 Dec 2026

  OpenAI: deprecationsNo date stated

• Gemini 4 Carbon Opens to Businesses

  Google is giving Gemini 4 Carbon first to cyber defenders and trusted partners, and says it will make it available to developers, businesses, and consumers “as soon as possible.” Until then a company cannot build on it.

  No date given

  Google: Introducing Gemini 4 Carbon28 Oct 2026