Every week

Last Week in AI for Chemicals:
16 to 22 Nov 2026

OpenAI's GPT-6.2 Astra and Google's Gemini 4 Neon arrived a day apart, and Neon now shares first place on regulatory research.

71.21%of vals.ai's regulatory research questions answered fully right, by Gemini 4 Neon, level with two others at the top
Vailent model guide · 5 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, 3 of those recommendations changed.

Research RegulationsBest result

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

BeforeGemini 4 Hydrogen69.7%, $4.84 a question
NowGemini 4 Neon71.21%, $5.76 a question

Google's Gemini 4 Neon answers 71.21% of vals.ai's regulatory research questions fully right, at $5.76 a question. Gemini 4 Hydrogen and Claude Sonnet 5.5 both score 69.70%, at $4.84 and $11.67. Each score carries a margin of error of about 6 points, so the three share the top, and the test can't separate them from 29 other models. The guide now names Neon, the highest score; Hydrogen gives a result the test can't tell apart for $0.92 less a question.

vals.ai Legal Research19 Nov 2026

Build and Check Spreadsheet ModelsMost efficient option

The most score for each dollar spent.

BeforeMuse Ember72.17%, $0.091 a task
NowTencent Hy4 Flash62.97%, $0.034 a task

Tencent released Hy4 Flash on 17 November, with its weights on Hugging Face under Tencent's own licence. It gets 62.97% of vals.ai's spreadsheet modeling tasks right at $0.034 a task, the most score for each dollar. That is outside the margin of error of Claude Fable 5.2, the leader at 77.38%, so it gets fewer right. Muse Ember, at $0.091 a task, is still inside that margin.

vals.ai spreadsheet modeling benchmark v1.119 Nov 2026

Search Market NewsBest result

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

BeforeKeenable Search83 out of 100, $40.21 a thousand tasks
NowParallel Search83.4 out of 100, $30.05 a thousand tasks

Artificial Analysis pairs each search tool with the same answer model and grades the answers. It measured Parallel Search's new settings on 19 November: with its advanced setting it scores 83.4 out of 100 at $30.05 a thousand tasks, search and answer together, against 83.0 for Keenable Search at about $40. The board publishes no margin of error, so the two tie, and the guide names Parallel because it costs about a quarter less.

Artificial Analysis Search Index19 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. GPT-6.2 Astra and Gemini 4 Neon Join the Top Three on the Index

    OpenAI released GPT-6.2 Astra, its new frontier model for the hardest projects, and GPT-6.2 Sol, its efficient model for everyday work, on 17 November. Astra costs $8 for every million tokens sent and $40 for every million written, 20% less than GPT-6 Astra, and Sol $2 and $10. Google released Gemini 4 Neon on 18 November, at $3 and $15, and it reads up to a million tokens at once. On the Artificial Analysis Intelligence Index, Astra is 2nd at 62.1 and Neon 3rd at 61.4, behind Claude Fable 5.2 at 64.6.

  2. Inkling 2 Mini Is an Open Model Small Enough for a Laptop

    Thinking Machines released Inkling 2 Mini on 20 November with open weights. It is small enough to run on a laptop, so a company could try it on its own machines. Artificial Analysis already lists it, well below the frontier models on its Intelligence Index.

  3. OpenAI Tests Its Own Models on Medicinal Chemistry

    OpenAI published MedChemBench on 20 November, its own test of 99 medicinal chemistry problems drawn from published research, and its GPT-6.2 Astra earns 84.5% of the available credit. OpenAI made both the test and the model, so the guide reads the result as OpenAI's claim, and the problems come from drug discovery, not a distributor's daily work.

  4. GPT-6.2 Astra Builds Spreadsheet Models Level With the Leader

    On vals.ai's spreadsheet modeling test, updated 19 November, GPT-6.2 Astra gets 76.03% of the tasks right at $2.80 a task, and Gemini 4 Neon 76.46% at $3.82. Both are inside the margin of error of Claude Fable 5.2, the leader at 77.38% for $6.67, so the test can't separate the three. Astra costs under half as much as Fable 5.2 a task, and Neon a little over half. GPT-6.2 Sol is inside the margin too, at $0.84.

    Build and Check Spreadsheet Models in the guide
  5. Astra for Law 6.2 Joins the Regulatory Test

    OpenAI released Astra for Law 6.2, a version of GPT-6.2 Astra for legal work, on 17 November, and any ChatGPT Enterprise customer can ask for access. On vals.ai's test it answers 60.61% of the regulatory research questions fully right, joint 15th of 101 and inside Neon's margin, at $2.47 a question. Across all eight areas of US law in the test it scores 53.85%, joint 7th.

    Research Regulations in the guide
  6. Qwen4 Max Rates 6th on Real Professional Work

    Alibaba released Qwen4 Max on 16 November, a general model it does not offer as open weights. On Artificial Analysis's test of real professional tasks it rates 6th, behind Claude Opus 5.5, which still leads. GPT-6.2 Astra and Gemini 4 Neon have no rating there yet.

    Write a Report From Several Sources 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.

  • Ignite Ran 17 to 20 November

    Last week's edition flagged Ignite, Microsoft's yearly conference for business customers. Its announcements on the Foundry blog were tools for building and running agents, such as agents that stay inside a company's own network, and a service for running open models in production.

    Came due

    Foundry blog19 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.

    Due

    OpenAI: deprecationsNo date stated

  • Gemini 4 Carbon and Argon Prices Double on 1 December

    Google's introductory prices for Gemini 4 Carbon and Gemini 4 Argon, $2 for every million tokens sent and $10 for every million written, run through 30 November. From 1 December both cost $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 to Gemini 3.8 Flash, and every request for Gemini 2.5 Flash to Gemini 3.5 Flash Lite. Check which Gemini model your tools call before then.

    Due

    Google: Gemini API deprecationsNo 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 GPT-6.1 Sol, GPT-6.1 Terra, and GPT-6 Luna 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 19 Nov 2026 · Vailent checked 21 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 21 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 21 Nov 2026
  4. OpenAI: Introducing GPT-6.2 Astra and Sol The maker's own page.Published 17 Nov 2026 · Vailent checked 21 Nov 2026
  5. Google: Introducing Gemini 4 Neon The maker's own page.Published 18 Nov 2026 · Vailent checked 21 Nov 2026
  6. OpenAI: Astra for Law The maker's own page.Published 17 Nov 2026 · Vailent checked 21 Nov 2026
  7. Tencent Hy4 Flash on Hugging Face The maker's own model page, which carries its licence.Published 17 Nov 2026 · Vailent checked 21 Nov 2026
  8. Thinking Machines: Introducing Inkling 2 Mini The maker's own page.Published 20 Nov 2026 · Vailent checked 21 Nov 2026
  9. OpenAI: Introducing MedChemBench OpenAI, which makes the models it tests.Published 20 Nov 2026 · Vailent checked 21 Nov 2026
  10. Google: Gemini API pricing The maker's own price list.States no date · Vailent checked 21 Nov 2026
  11. Google: Gemini API deprecations The maker's own page.States no date · Vailent checked 21 Nov 2026
  12. OpenAI: deprecations The maker's own page.States no date · Vailent checked 21 Nov 2026
  13. Foundry blog The maker's own blog.Published 19 Nov 2026 · Vailent checked 21 Nov 2026