Last Week in AI for Chemicals:
2 to 8 Nov 2026
Kimi K3.1 Pro, level with the best model on real professional work, is now free to run on your own servers.
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, 3 of those recommendations changed.
Write a Report From Several SourcesFree to run in-house
The best model you can download and run on your own servers.
Kimi K3.1 Pro is Moonshot AI's general model, the kind you would ask to draft a report from your sales data and market news. Moonshot put its weights on Hugging Face on 2 November, and Artificial Analysis now lists it as an open model. On real professional work it ranks 2nd of 27, level with Claude Opus 5.5 within the test's margin, and 112 points ahead of DeepSeek V4.1 Flash 1017, the next open model. Its license is free, and its conditions, which cover companies that sell the model as a hosted service or show it to more than 100 million users, do not apply to a company's internal use. It is very large, about 2.8 trillion parameters, so running it yourself takes a large cluster of servers. Artificial Analysis lists its hosted price at $2 for every million tokens sent and $10 for every million written.
Artificial Analysis: GDPval-AANo date stated
Read Scans and PhotosBest result
The top score on the most independent public test for this job.
OvisOCR3 is a small model from ATH-MaaS, 0.85 billion parameters, built to read documents. MDPBench, the guide's test of reading photographed and scanned pages, added it on 6 November, and on photographed pages it scores 87.4 out of 100, 2.3 points above Gemini 3 Pro Preview. The test publishes no margin of error. It is run by the lab that also makes MonkeyOCR, a model it ranks, and the newest Claude, GPT, and Gemini models have not been tested on it.
MDPBench leaderboard6 Nov 2026
Read Scans and PhotosFree to run in-house
The best model you can download and run on your own servers.
OvisOCR3's weights are free to download under the Apache 2.0 license, which allows business use, so it is also the best result you can run on your own computers, 10.2 points above dots.mocr.
MDPBench leaderboard6 Nov 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.
Inkling 2 Small Is Level With the Leader on Regulatory Research
Inkling 2 Small is a general model from Thinking Machines Lab, which says it is especially strong on finance work over spreadsheets, filings, and documents. Thinking Machines released it on 5 November through its own platform, Tinker, at $0.30 for every million tokens sent and $1 for every million written. On vals.ai's regulatory research questions it answers 60.61% fully right at $0.24 a question, inside the margin of the leader, Claude Sonnet 5.5, at 69.7%. On Artificial Analysis's test of real professional work it ranks 16th of 27. Its weights are closed for now, and Thinking Machines says it intends to release them next week.
Research Regulations in the guideThinking Machines: Introducing Inkling 2 Small5 Nov 2026
vals.ai Legal Research5 Nov 2026
Artificial Analysis: GDPval-AANo date stated
Artificial Analysis: models and the Intelligence IndexNo date stated
Tencent's Hy4 Lands Just Outside the Regulatory Research Leaders
Hy4 is Tencent's new general model, released on 3 November, and Artificial Analysis lists it at $0.83 for every million tokens sent and $2.50 for every million written. On vals.ai's regulatory research questions it answers 51.52% fully right at $2.21 a question, just outside the leader's margin, while Hy4 Preview, its earlier version, is inside it at 54.55%. Its weights are closed.
Research Regulations in the guidevals.ai Legal Research5 Nov 2026
Artificial Analysis: models and the Intelligence IndexNo date stated
AI Customer Service Agents Slip in Korean and Farsi
τ-bench-LILT, a study by LILT, which ranks no model of its own, gives AI agents customer service tasks in retail, airlines, telecoms, and banking, each worked against the company's own policy, in English, German, Korean, Turkish, and Farsi. Asked to get the same retail task right in four tries out of four, Claude Opus 5.5 held up best, at 91.2% in English and 78.1% in Korean. NVIDIA's Nemotron 3 Ultra fell from 78.1% in English to 17.5% in Farsi, and Claude Opus 5 from 81.6% to 26.3% in Korean. If your customers write to you in other languages, test an agent in each one. The study publishes no prices.
Answer Service Requests From Company Policy in the guideOffice Files Can Mislead AI Assistants
An AI assistant reads the text a converter pulls out of a Word or PowerPoint file, and a study published on 4 November shows that text can differ from what a person sees on the page. Moving the text boxes on a slide, for example, can change the order the AI reads them while the slide looks the same. Across five such tricks, Gemini was misled in every attempt, Claude and Microsoft 365 Copilot in 80%, and ChatGPT in 40%. Claude Code was misled in 24%, so tools from the same maker differ by how the file reaches the model. Check the values an AI pulls from your supplier files against the page itself, as you would a scan.
Read Supplier Documents in the guide
Released, not yet tested.
No independent test has run these, so the guide shows no score.
Gemini Now Accepts Older Office Files
No independent testGoogle's Gemini API now accepts PowerPoint 97 to 2003 files and OpenDocument presentations and spreadsheets, and reads Word 97 to 2003 files as page images. Older supplier files, such as an archived safety data sheet or a price list saved years ago, can now go straight to the model. No public test has measured how well it reads them.
Published
Google: Gemini API release notes5 Nov 2026
Reducto Releases r-1 mini, a Cheaper Document Reader
No independent testReducto, which sells document reading to businesses, released r-1 mini on 5 November at $1 per 1,000 pages. Its accuracy claims come from Reducto's own test, and no independent test has scored r-1 mini yet.
Published
Reducto: Introducing r-1 mini5 Nov 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.
Kimi K3.1 Pro's Weights Shipped
Last week's edition flagged Moonshot AI's promise to release Kimi K3.1 Pro's weights by 4 November. They went up on Hugging Face on 2 November, and they make Kimi K3.1 Pro the guide's pick for writing reports on your own servers.
Microsoft Ignite Runs 17 to 20 November
Microsoft holds Ignite, its yearly conference for business customers, from 17 to 20 November, and its AI platform team tells developers to plan their viewing. If your company runs on Microsoft 365 or Azure, watch for changes to the AI tools you already pay for.
Microsoft Foundry blog: What's new in Microsoft Foundry, October 20265 Nov 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. An agent or a check built in either stops working that day unless it moves.
OpenAI: deprecationsNo date stated
Gemini 2.5 Pro and Flash Move to Newer Models on 1 December
Google deprecated Gemini 2.5 Pro and Gemini 2.5 Flash on 2 November. From 1 December it sends every request for either one to a newer Gemini model and bills it at that model's price. Check which Gemini model your tools call before then.
Google: Gemini API deprecations2 Nov 2026
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.
OpenAI: deprecationsNo date stated
Inkling 2 Small's Weights, Promised for Next Week
Thinking Machines says: “We intend to make the full weights available under an Apache 2.0 license next week after the completion of external safety testing and red-teaming.” Apache 2.0 allows business use, so once they ship a company could run Inkling 2 Small on its own servers.
Gemini 4 Carbon Opens to Businesses
Google is still giving Gemini 4 Carbon only to cyber defenders and trusted partners, and says it will reach developers, businesses, and consumers “as soon as possible.” Until then a company cannot build on it.
Google: Introducing Gemini 4 Carbon28 Oct 2026
Sources.
- Artificial Analysis: GDPval-AA Artificial Analysis, an independent testing company. It rates models on real professional tasks and marks which have open weights.States no date · Vailent checked 6 Nov 2026
- Moonshot AI: Kimi-K3.1-Pro on Hugging Face The maker's own model page, which gives the license.Published 2 Nov 2026 · Vailent checked 6 Nov 2026
- MDPBench leaderboard The OCRBench authors' lab, which also makes MonkeyOCR, a model it ranks.Published 6 Nov 2026 · Vailent checked 6 Nov 2026
- ATH-MaaS: OvisOCR3 on Hugging Face The maker's own model page, which gives the Apache 2.0 license.States no date · Vailent checked 6 Nov 2026
- Thinking Machines: Introducing Inkling 2 Small The maker's own page.Published 5 Nov 2026 · Vailent checked 6 Nov 2026
- vals.ai Legal Research vals.ai, an independent testing company.Published 5 Nov 2026 · Vailent checked 6 Nov 2026
- Artificial Analysis: models and the Intelligence Index Artificial Analysis, an independent testing company. Each model's record gives its price and its release date.States no date · Vailent checked 6 Nov 2026
- τ-bench-LILT: Evaluating Conversational Agents on Realistic Multilingual Tool-Use Tasks A published study by LILT, which ranks no model of its own.Published 4 Nov 2026 · Vailent checked 6 Nov 2026
- When File Ingestion Misleads Language Models: Measuring and Mitigating Divergences in Office Documents A published study whose author ranks no model of their own.Published 4 Nov 2026 · Vailent checked 6 Nov 2026
- Google: Gemini API release notes The maker's own changelog.Published 5 Nov 2026 · Vailent checked 6 Nov 2026
- Google: Gemini API deprecations The maker's own page.Published 2 Nov 2026 · Vailent checked 6 Nov 2026
- Reducto: Introducing r-1 mini The maker's own blog.Published 5 Nov 2026 · Vailent checked 6 Nov 2026
- Microsoft Foundry blog: What's new in Microsoft Foundry, October 2026 The maker's own blog.Published 5 Nov 2026 · Vailent checked 6 Nov 2026
- OpenAI: deprecations The maker's own page.States no date · Vailent checked 6 Nov 2026
- Google: Introducing Gemini 4 Carbon The maker's own page.Published 28 Oct 2026 · Vailent checked 6 Nov 2026