Last Week in AI for Chemicals:
19 Oct to 25 Oct 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, 2 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. 24 Oct 2026

   Kimi K3.1 Pro Ties Claude Opus 5.5 on Professional Work

   Kimi K3.1 Pro is Moonshot AI's new general model, Kimi K3 trained further for long coding and office work. Moonshot AI released it on 24 October at $2 for every million tokens sent and $10 for every million written, half of Opus 5.5's price. Artificial Analysis rates it 1846 on real professional work, 3rd of 27 and level with Claude Opus 5.5. On vals.ai's spreadsheet test it scores 70.00% at $1.59 a task, inside the leader's margin. Moonshot AI says it will publish the weights by 4 November; until then it runs only on Moonshot AI's own service.

   Artificial Analysis: models and the Intelligence IndexNo date stated

   Moonshot AI: Kimi K3.1 Pro24 Oct 2026

   vals.ai Excel Modeling Benchmark v1.124 Oct 2026

   Write a Report From Several Sources in the guide →
2. 24 Oct 2026

   vals.ai Moves Its Spreadsheet Test to Version 1.1

   vals.ai replaced its Excel Modeling Benchmark with version 1.1 on 24 October, with GPT-6 Luna as its new grader, and says the scores are not directly comparable with the old ones. The guide's picks hold: Claude Fable 5.1 leads at 76.73% of tasks graded correct, MiMo V2.6 Flash gives the most for each dollar at $0.11 a task, and Kimi K3 is the best result you can run on your own servers. The test cannot separate Fable 5.1 from 12 other models, among them GPT-6.1 Sol at $1.14 a task.

   vals.ai Excel Modeling Benchmark v1.124 Oct 2026

   Build and Check Spreadsheet Models in the guide →
3. 23 Oct 2026

   The Best Forecaster Licensed for Business Use Is Now t0-gamma

   fev-bench ranks forecasting models on 100 general forecasting tasks, the test the guide uses for products that sell steadily. Its board was rebuilt on 23 October. TimeH2O-v1 from H2O.ai now leads with a 91.9% win rate and Mistral's Timestral-209M is second at 89.6%, but neither is licensed for business use. The best that is, at 88.1%, is t0-gamma from The Forecasting Company, under the Apache 2.0 license. Mistral said on 15 October that Timestral ranks first on fev-bench; on the board today it is second.

   fev-bench leaderboard23 Oct 2026

   The Forecasting Company: t0-gamma on Hugging FaceNo date stated

   Forecast Demand for Slow-Moving Stock in the guide →
4. 23 Oct 2026

   ParseBench Corrects How It Scores Tracing Text to the Page

   ParseBench, run by LlamaIndex, which sells a document parser it ranks, corrected how it scored tracing text back to its place on the page for 65 results on 23 October, and is still re-scoring 13 more. TeleOCR, the guide's pick for reading supplier documents, moves from 5th to 15th on that measure and is 96th of 149 overall. The pick rests on OmniDocBench, the independent test, where TeleOCR is still first.

   ParseBench: correction to Visual Grounding scores23 Oct 2026

   OmniDocBench11 Sep 2026

   Read Supplier Documents in the guide →
5. 22 Oct 2026

   Upstage's Document Parser Enters ParseBench in 4th Place

   Document Parse 2 (Enhanced) is Upstage's paid service for reading documents into text and tables. ParseBench added it on 22 October at 80.51 out of 100 on whole documents, 4th of 149, for 3 cents a page, behind two LlamaParse modes and Pulse Ultra 2. ParseBench is run by LlamaIndex, which sells LlamaParse. The guide's pick for reading supplier documents rests on OmniDocBench, the independent test, where TeleOCR is still first.

   ParseBench: Upstage Document Parse 2 (Enhanced) results22 Oct 2026

   OmniDocBench11 Sep 2026

   Read Supplier Documents in the guide →
6. 22 Oct 2026

   Claude Opus 5.5 Adds a Made-Up Claim to 9.9% of Summaries

   Vectara's hallucination leaderboard has each model summarize the same documents and counts the summaries that state something the document does not support. On 22 October it added Claude Opus 5.5 at 9.9% and DeepSeek V4.1 Flash, an open model, at 4.2%, just behind Llama 3.3 70B at 4.1%, the guide's pick you can run on your own servers. Finix S1 32B from Ant Group still leads at 1.8%, so the guide's picks for drafting replies hold.

   Vectara hallucination leaderboard22 Oct 2026

   Draft Replies Without Made-Up Facts in the guide →
7. 20 Oct 2026

   The Guide Adds Finding the Current Answer When Records Disagree

   A study from KAIST and Korea University, published 20 October, asks 1,613 questions over work records that change across chats, notes, emails, and PDFs, the way a later email revises the price in a quote. GPT-5.6 Sol answers best, at 68.0% with 50 records found for each question, and Qwen3.5-27B is the best you can run on your own servers, at 50.3%. Where two records disagree, even Sol gets only 55 to 56% right. The study tested no Claude model and published no prices.

   MMEnterprise-Bench: Evaluating Enterprise Agents over Evolving Multimodal Records20 Oct 2026

   Find the Current Answer When Records Disagree in the guide →
8. 19 Oct 2026

   MiniMax M3.1 Flash, a Low-Cost Model for Agent Work

   MiniMax M3.1 Flash is a low-cost model from MiniMax built for agent tasks, such as working through a supplier portal step by step. MiniMax released it on 19 October at $0.15 for every million tokens sent and $0.60 for every million written, and put its weights on Hugging Face on 23 October under its own license. Artificial Analysis rates it 18th of 27 on real professional work, and vals.ai puts it 47th of 85 on US regulatory research.

   MiniMax: MiniMax M3.1 Flash19 Oct 2026

   MiniMax: MiniMax-M3.1-Flash on Hugging Face23 Oct 2026

   Artificial Analysis: models and the Intelligence IndexNo date stated

   vals.ai Legal Research, regulatory questions24 Oct 2026

9. 19 Oct 2026

   Mercury 3, a Fast Model That Drafts the Whole Answer at Once

   Mercury 3 is a model from Inception that drafts a whole answer at once and then refines it, rather than writing one word at a time, so it answers fast: Inception quotes 1,068 tokens a second. Inception released it on 19 October through its own service and other cloud platforms, at $0.20 for every million tokens sent and $0.70 for every million written, half that while its launch discount lasts, and has published no weights. vals.ai puts it 55th of 85 on US regulatory research, at about 8 cents a question.

   Inception: Introducing Mercury 319 Oct 2026

   vals.ai Legal Research, regulatory questions24 Oct 2026

Released, not yet tested.

No independent test has run these, so the guide shows no score.

• Rerank 5, Cohere's Models That Sort Search Results

  No independent test

  Rerank 5 is a pair of models from Cohere that put the most relevant passages first when a tool searches your own files, so the model answering sees less noise. Cohere released Rerank 5 Pro and Rerank 5 Fast on 21 October; the Pro model aims at the hardest searches and the Fast one at high volume. No test the guide uses has scored them.

  Published 21 Oct 2026

  Cohere: changelog21 Oct 2026

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.

• Google's Oldest Deep Research Agent Reached Its Shutdown Date

  Last week's edition flagged that Google would shut down its oldest Deep Research preview, from December 2025, on 23 October. Google's deprecations page still lists that date and names the April 2026 version as the replacement, so market research built on the old version should run on the newer one.

  Came due 23 Oct 2026

  Google: Gemini API deprecationsNo date stated

• Mistral Large 4's Weights, Promised for October

  Mistral Large 4 is Mistral's largest general model, and it still runs only on Mistral's own service. Mistral's page still says: “We will release the weights by the end of the month.” Once they ship, a company can run it on its own servers.

  Due 31 Oct 2026

  Mistral AI: Mistral Large 46 Oct 2026

• GLM 5.2 Leaves Mistral's Service

  Mistral resells Z.ai's GLM 5.2 and retires it on 31 October, pointing users to GLM 5.3 at the same price. A tool that calls GLM 5.2 through Mistral needs to switch before then.

  Due 31 Oct 2026

  Mistral AI: changelogNo date stated

• Moonshot AI Closes the Kimi K2 Models to New Users

  From 31 October, new Moonshot AI accounts can no longer use the Kimi K2 models, and Moonshot AI plans to retire all of them by the end of December. A tool built on Kimi K2 should plan its move.

  Due 31 Oct 2026

  Kimi API Platform: model pricingNo date stated

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

  Moonshot AI says: “The full model weights will be released by Nov 4.” Once they ship, a company can run Kimi K3.1 Pro on its own servers.

  Due 4 Nov 2026

  Moonshot AI: Kimi K3.1 Pro24 Oct 2026

• Claude Sonnet 4.5 Retires on 30 November

  Anthropic retires Claude Sonnet 4.5 on its API on 30 November 2026 and names Claude Sonnet 5.5 as the replacement. A tool built on the older model stops answering that day unless it moves.

  Due 30 Nov 2026

  Anthropic: model deprecations30 Sep 2026

• 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