Every week

Last Week in AI for Chemicals:
9 to 15 Nov 2026

Claude Fable 5.2 now builds spreadsheet models best and leads the general index, and two cheaper models joined the top group on regulatory research.

77.38%of spreadsheet modeling tasks graded correct, by Claude Fable 5.2, the best result
Vailent model guide · 6 min read

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, 5 of those recommendations changed.

Build and Check Spreadsheet ModelsBest result

The top score on the most independent public test for this job.

BeforeClaude Fable 5.176.73%, $15.93 a task
NowClaude Fable 5.277.38%, $6.67 a task

vals.ai's test of building financial models in spreadsheets added Claude Fable 5.2 on 11 November. It gets 77.38% of the tasks right at $6.67 a task, against 76.73% at $15.93 for Fable 5.1. The test can't separate the two, or 15 other models, inside its margin of error, so the reason to switch is the bill: Fable 5.2 does the same work for well under half the price. The test's tasks are financial models such as cash-flow valuations, not price lists.

vals.ai spreadsheet modeling benchmark v1.111 Nov 2026

Build and Check Spreadsheet ModelsMost efficient option

The most score for each dollar spent.

BeforeMiMo V2.6 Flash68.88%, $0.11 a task
NowMuse Ember72.17%, $0.091 a task

Meta's Muse Ember gets 72.17% of the same spreadsheet tasks right at $0.091 a task, inside Fable 5.2's margin of error, so it gives the most score for each dollar. Fable 5.2 costs about 73 times as much per task.

vals.ai spreadsheet modeling benchmark v1.111 Nov 2026

Research RegulationsBest result

The top score on the most independent public test for this job.

BeforeClaude Sonnet 5.569.7%, $11.67 a question
NowGemini 4 Hydrogen69.7%, $4.84 a question

Google's Gemini 4 Hydrogen answers 69.70% of vals.ai's regulatory research questions fully right, the same score as Claude Sonnet 5.5, at $4.84 a question against Sonnet's $11.67. The two share first place, and the guide now names Hydrogen because it costs less than half as much. Each score carries a margin of error of about 6 points, so the test can't separate the leaders from 30 other models.

vals.ai Legal Research11 Nov 2026

Research RegulationsMost efficient option

The most score for each dollar spent.

BeforeMiMo V2.6 Flash46.97%, $0.067 a question
NowMuse Ember54.55%, $0.056 a question

Muse Ember answers 54.55% of the regulatory questions fully right at $0.056 a question. That is inside the leaders' margin of error, at about 87 times less than Gemini 4 Hydrogen and about 210 times less than Sonnet 5.5, so it gives the most score for each dollar.

vals.ai Legal Research11 Nov 2026

Search Market NewsBest result

The top score on the most independent public test for this job.

BeforeExa · dynamic highlights80.3 out of 100, $63.78 a thousand tasks
NowKeenable Search83 out of 100, $40.21 a thousand tasks

Artificial Analysis pairs each search tool with the same answer model and grades the answers. Keenable Search, with its pro setting, now scores 83.0 out of 100 at $40.21 a thousand tasks, search and answer together, and 82.1 with its realtime setting. Exa, the previous leader, scores 80.3 for $63.78. The board publishes no margin of error.

Artificial Analysis Search Index9 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.

  1. Claude Fable 5.2 Leads the Index and Matches Opus on Professional Work

    Anthropic released Claude Fable 5.2 on 9 November, at $8 for every million tokens sent and $40 for every million written, a lower price than Fable 5.1's. Artificial Analysis puts it first on its Intelligence Index at 64.6, ahead of Gemini 4 Carbon at 60.0 and Claude Opus 5.5 at 59.1. On real professional work it rates 1831, 5th of 28 and level with Opus 5.5 inside the test's margin. Claude Mythos 5.2, released the same day, goes only to organizations Anthropic vets.

    Write a Report From Several Sources in the guide
  2. Gemini 4 Hydrogen Ties for First on Regulatory Research

    Google released Gemini 4 Hydrogen to everyone on 12 November, a fast model at $1.50 for every million tokens sent and $7.50 for every million written. On vals.ai's regulatory research questions it shares first place with Claude Sonnet 5.5 at 69.70%, for $4.84 a question, and it is now the guide's pick for the job.

    Research Regulations in the guide
  3. MiniCPM-OCR 1.0 Joins the Document Reading Test

    OmniDocBench added OpenBMB's MiniCPM-OCR 1.0 on 12 November. It scores 96.44 out of 100 on whole documents, 4th behind OXR-1.0, TeleOCR, and OvisOCR2, and 94.40 on tables. The guide's picks for reading documents don't change.

    OmniDocBench12 Nov 2026

    Read Supplier Documents in the guide
  4. ParseBench 2 Adds a Test of Reading Forms

    LlamaIndex, which sells the LlamaParse document reader, published ParseBench 2 on 12 November: 4,317 test cases across six measures, including a new one on forms with 600 form pages. Its maker says the scores can't be compared with the first version, so the guide keeps the first version's results for now.

    Pull Tables and Price Lists in the guide
  5. A New Test of Work on Drawings, Plans, and Visual Files

    Artificial Analysis and Google DeepMind, which makes Gemini, launched the Multimodal Knowledge Agent test on 10 November: 200 private tasks across 10 industries, worked from dense visual project files such as drawings and marked-up images. Claude Fable 5.2 leads at 57%, ahead of Gemini 4 Carbon at 54% and Claude Opus 5.5 at 53%, but it costs $31.97 a task against Carbon's $6.29. Because one of its builders sells a model it ranks, the guide reads it as a shortlist, not a verdict.

  6. Muse Ember Answers Regulatory Questions for Cents

    Muse Ember is Meta's efficient model for high-volume, repeated work, such as pulling fields out of thousands of files, at $0.10 for every million tokens sent and $0.50 for every million written. Meta's page gives no release date, and Artificial Analysis lists it from 9 November. It answers 54.55% of vals.ai's regulatory questions fully right at $0.056 a question, inside the leaders' margin, and gets 72.17% of the spreadsheet modeling tasks right at $0.091 a task, inside Fable 5.2's margin.

    Research Regulations in the guide
  7. Keenable Search Tops the Search Tool Test

    On Artificial Analysis's test of search tools, measured on 9 November, Keenable Search scores 83.0 out of 100 with its pro setting and 82.1 with its realtime setting, at about $40 a thousand tasks. Exa, the previous leader, scores 80.3 at $63.78. TinyFish still gives the most score for the money, 71.2 at $8.62.

    Search Market News in the guide

Released, not yet tested.

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

  • Google Releases Gemini Decision 1 for Yes-or-No Questions

    No independent test

    Gemini Decision 1 answers yes-or-no, category, and checklist questions about text, images, audio, video, and PDFs, and writes no text back, so Google bills only what you send, at $0.10 for every million tokens. It suits sorting work such as tagging incoming quote requests or supplier documents. No public test has measured it yet.

    Published

    Google: Gemini API release notes12 Nov 2026

  • DeepSeek Opens Web Search and File Reading in Its API

    No independent test

    DeepSeek now lets developers switch on web search, file input for PDFs and other documents, and reading a linked page inside its existing API, and it offers search, OCR, and speech as separate APIs. A company could ask its model about a supplier's PDF without building its own file reader. No public test has measured these tools.

    Published

    DeepSeek: API change log13 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.

  • Gemini 4 Carbon Opened to Businesses

    Last week's edition said Google gave Gemini 4 Carbon only to cyber defenders and trusted partners. On 10 November Google made it available to every paying developer and business, at an introductory $2 for every million tokens sent and $10 for every million written.

    Came due

    Google: Gemini API release notes12 Nov 2026

  • Inkling 2 Small's Weights Shipped

    Last week's edition flagged Thinking Machines' promise to release Inkling 2 Small's weights. They went up on Hugging Face on 12 November, so a company can now run the model on its own servers. On regulatory research it sits inside the leaders' margin at $0.24 a question.

    Came due

    Thinking Machines: Introducing Inkling 2 Small12 Nov 2026

  • Ignite Runs 17 to 20 November

    Microsoft holds Ignite, its yearly conference for business customers, from 17 to 20 November. If your company uses its cloud or office software, watch for changes to the AI tools inside them.

    Due

    Foundry blogNo date stated

  • 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.

    Due

    OpenAI: deprecationsNo date stated

  • Gemini 4 Carbon's Price Doubles on 1 December

    Google's introductory price for Gemini 4 Carbon, $2 for every million tokens sent and $10 for every million written, runs through 30 November. From 1 December it is $4 and $20.

    Due

    Google: Gemini API pricingNo date stated

  • Gemini 2.5 Pro and Flash Move to Newer Models on 1 December

    From 1 December Google sends every request for Gemini 2.5 Pro or Gemini 2.5 Flash to a newer Gemini model and bills it at that model's price. Check which Gemini model your tools call before then.

    Due

    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.

    Due

    OpenAI: deprecationsNo date stated

Sources.

  1. vals.ai spreadsheet modeling benchmark v1.1 vals.ai, an independent testing company. Its tasks are financial models such as cash-flow valuations.Published 11 Nov 2026 · Vailent checked 13 Nov 2026
  2. Artificial Analysis: models and the Intelligence Index Artificial Analysis, an independent testing company. Each model's record gives its index score, its price, and its release date.States no date · Vailent checked 13 Nov 2026
  3. Artificial Analysis: GDPval-AA Artificial Analysis, an independent testing company. It rates models on real professional tasks.States no date · Vailent checked 13 Nov 2026
  4. Artificial Analysis: Multimodal Knowledge Agent benchmark Artificial Analysis, built with Google DeepMind, which makes Gemini, a model it ranks.Published 10 Nov 2026 · Vailent checked 13 Nov 2026
  5. Anthropic: Claude Fable 5.2 and Mythos 5.2 The maker's own page.Published 9 Nov 2026 · Vailent checked 13 Nov 2026
  6. Google: Gemini API release notes The maker's own changelog.Published 12 Nov 2026 · Vailent checked 13 Nov 2026
  7. Google: Gemini API pricing The maker's own price list.States no date · Vailent checked 13 Nov 2026
  8. Meta: Muse Ember The maker's own page, which gives the model's price.States no date · Vailent checked 13 Nov 2026
  9. DeepSeek: API change log The maker's own change log.Published 13 Nov 2026 · Vailent checked 13 Nov 2026
  10. OmniDocBench OpenDataLab (Shanghai AI Lab), which also makes MinerU, a model it ranks.Published 12 Nov 2026 · Vailent checked 13 Nov 2026
  11. ParseBench on GitHub LlamaIndex, which sells LlamaParse, a document reader it ranks.Published 12 Nov 2026 · Vailent checked 13 Nov 2026
  12. Thinking Machines: Introducing Inkling 2 Small The maker's own page.Published 12 Nov 2026 · Vailent checked 13 Nov 2026
  13. Foundry blog The maker's own blog.States no date · Vailent checked 13 Nov 2026
  14. OpenAI: deprecations The maker's own page.States no date · Vailent checked 13 Nov 2026
  15. Google: Gemini API deprecations The maker's own page.Published 2 Nov 2026 · Vailent checked 13 Nov 2026