Last Week in AI for Chemicals:
12 Oct to 18 Oct 2026

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. 16 Oct 2026

   Muse Spark Open Ties Claude Haiku 5.5 on Regulatory Research

   Muse Spark Open is a general model from Meta that drafts, summarizes, and works through research, and a company can run it on its own servers once Meta approves access to the weights. Meta published it on 15 October. On vals.ai's US regulatory research test it answers 50% of questions fully right, 28th of 79 and level with Claude Haiku 5.5, at $1.42 a question. Claude Sonnet 5.5 leads that test at 69.7%, and Muse Spark Open sits outside its margin. Artificial Analysis lists it at $1.25 for every million tokens sent and $4.25 for every million written.

   vals.ai Legal Research, regulatory questions16 Oct 2026

   Artificial Analysis: models and the Intelligence IndexNo date stated

   Meta: Muse Spark Open on Hugging FaceNo date stated

   Research Regulations in the guide →
2. 16 Oct 2026

   Three Models You Can Download Join the Professional Work Ranking

   Artificial Analysis ranks models on real professional tasks, such as a report a customer will read. Three models with downloadable weights joined: Meta's Muse Spark Open at 22nd of 27, Reflection AI's Beam 1.0 at 25th, and NVIDIA's Nemotron 3.5 Super VL at 26th. Beam 1.0 is a general text model, and Reflection published its weights on 16 October under the Apache 2.0 license, as it promised last week. Nemotron 3.5 Super VL reads images, scanned documents, and video as well as text. MiMo V2.6 Pro, 7th of 27, stays the best model for this job that you can run on your own servers.

   Artificial Analysis: models and the Intelligence IndexNo date stated

   Reflection AI: Beam 1.0 on Hugging Face16 Oct 2026

   NVIDIA: Nemotron 3.5 Super VL on Hugging Face15 Oct 2026

   Write a Report From Several Sources in the guide →
3. 16 Oct 2026

   Gemini 4 Argon Joins the Test for Made-Up Facts

   Vectara's leaderboard counts how often a model adds a claim its source does not support when it summarizes a document, the same risk as an invented lead time in a drafted customer reply. Vectara added Google's Gemini 4 Argon on 16 October: it added an unsupported claim to 12.1% of summaries, 81st of 109 models. Finix S1 32B, from Ant Group, stays the guide's pick at 1.8%. Google made Argon generally available on 13 October.

   Vectara hallucination leaderboard16 Oct 2026

   Google: Gemini API release notesNo date stated

   Draft Replies Without Made-Up Facts in the guide →
4. 16 Oct 2026

   The Guide Adds Spreadsheet Models and Service Requests

   The guide now covers building financial spreadsheets in Excel, on vals.ai's Excel Modeling Benchmark, updated 16 October. Claude Fable 5.1 leads with 76.67% of tasks graded correct, at $15.93 a task, and the test cannot separate it from 12 other models, down to GPT-6.1 Sol at $1.14. MiMo V2.6 Flash, an open model, gives the most for each dollar at $0.11 a task. The tasks are financial models such as buyouts and valuations, so check any margin or pricing sheet yourself. The guide also adds answering service requests from your own policy documents, on Sierra's tau-bench, where Qwen 3.8 Max resolved 55.2% of banking requests on the first try; its newest results date from July.

   vals.ai Excel Modeling Benchmark16 Oct 2026

   tau-bench leaderboard, BankingNo date stated

   Build and Check Spreadsheet Models in the guide →

Released, not yet tested.

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

• Nemotron 3.5 Bolt, a Model Built to Run on One Machine

  No independent test

  Nemotron 3.5 Bolt is a model from NVIDIA that reads text and images and is built to run everyday agent tasks on a single local machine. NVIDIA published it on 15 October with downloadable weights under its OpenMDW license, including a version compressed to fit in 24GB of memory. No test the guide uses has scored it.

  Published 15 Oct 2026

  NVIDIA: Nemotron 3.5 Bolt on Hugging Face15 Oct 2026

• DeepSeek V4.1 Flash Update and DeepSeek's Search API

  No independent test

  DeepSeek-V4.1-Flash-1017 is what DeepSeek calls a minor update to V4.1 Flash, its open model. DeepSeek says it finishes tasks faster and writes shorter answers. The same day DeepSeek opened its own web search and page fetch service to developers, so a model can look up current pages such as a supplier's price notice. The guide's tests scored the earlier version; none has scored the update.

  Published 17 Oct 2026

  DeepSeek: API change log17 Oct 2026

• Timestral, Free Forecasting Models From Mistral

  No independent test

  Timestral is a family of four small models from Mistral that forecast a series of numbers, such as weekly sales of a grade, without training on your data first. Mistral released them on 15 October under a modified MIT license, from 5 million to 209 million parameters, small enough to run without a graphics card. Mistral says the largest ranks first on three public forecasting tests. The leaderboard the guide uses has not shown it yet.

  Published 15 Oct 2026

  Mistral AI: Introducing Timestral15 Oct 2026

• LFM2.5-24B-A2B, a Model for Your Own Devices From Liquid AI

  No independent test

  LFM2.5-24B-A2B is a text model from Liquid AI built to run agents on a laptop or other device, so the data stays on it. Liquid AI released it on 13 October under its own license rather than a standard open source one. Only Liquid AI's own scores exist so far.

  Published 13 Oct 2026

  Liquid AI: LFM2.5-24B-A2B13 Oct 2026

• Granite Switch 4.1, IBM's Model That Answers, Sorts, and Transcribes

  No independent test

  Granite Switch 4.1 is a model from IBM that writes replies, picks from set answers such as which team a request goes to, and turns speech into text, all from one service. IBM released it on 14 October in 8B and 30B sizes under the Apache 2.0 license, free to run on your own computers. No public test the guide uses has scored it.

  Published 14 Oct 2026

  IBM: Granite Switch 4.1 on Hugging Face14 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. The nearest study, on the sea transport code, was last revised on 28 September.

• 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 Shuts Down Its Oldest Deep Research Agent

  Google's Deep Research agent searches the web and writes a sourced report. Google shuts down its oldest preview version, from December 2025, on 23 October and points users to the versions from April 2026. Market research that runs on the old version stops that day.

  Due 23 Oct 2026

  Google: Gemini API release notesNo date stated

• Mistral Large 4's Weights, Promised for October

  Mistral Large 4 is Mistral's largest general model, and today it runs only on Mistral's own service. Mistral 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 on its own service and retires it there 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

• 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

• Timestral on the Forecasting Leaderboard

  The team behind fev-bench, the test the guide uses for products that sell steadily, posted Timestral's results to its code on 16 October. Its public leaderboard was last rebuilt on 2 October and does not show them yet.

  No date given

  fev-bench code and results16 Oct 2026