Model highlight

Claude Sonnet 5.5 for Chemical Companies: At the Top on Regulations, in a Crowd

On an independent test of regulatory research, Claude Sonnet 5.5 is tied at the top with Gemini 4 Hydrogen and Gemini 4 Neon. The test cannot separate the three from 29 other models, one at about a two-hundredth of Sonnet's price. For a chemical distributor or producer, the question becomes which model is good enough for the price, not which one is first.

69.7%of US regulatory research questions answered fully right, tied at the top of 101 models
Anthropic · Released 28 Sep 2026

Latest update

vals.ai added Google's Gemini 4 Neon at 71.21% on regulatory research, the highest score, 1.5 points above Claude Sonnet 5.5 and inside the test's margin. Sonnet now shares the top of 101 models with Neon and Gemini 4 Hydrogen, in a leading group of 32. OpenAI's GPT-6.2 Astra and Astra for Law 6.2 joined that group, and four models at 53.03%, GPT-6.1 Sol among them, now sit just outside it. On the whole legal test Neon leads at 56.25%.

1 of 13

Claude Sonnet 5.5 at a glance

Made by
Anthropic5
Released
28 Sep 20265
Price
$2 per million tokens sent, $10 per million written6
Reads
Text and images, up to 1 million tokens3
Settings
Five, from low to max3
Index
54.5 on the Artificial Analysis Intelligence Index, 10.0 points behind Claude Fable 5.2, the leader2
Regulatory
Tied at the top of 101 with Gemini 4 Hydrogen and Gemini 4 Neon, in a leading group of 321

What Claude Sonnet 5.5 is.

Anthropic released Claude Sonnet 5.5 on 28 September 2026. It reads text and images, holds up to 1 million tokens, and costs $2 for every million tokens you send and $10 for every million it writes, the same as Sonnet 5.

Anthropic says it “runs 30% faster and costs up to 30% less for most work” than Sonnet 5. An independent test found otherwise on cost: at its max setting it used more tokens than any model the testers have measured, and a task cost more than on Sonnet 5.

Tied at the top on regulatory research, in a leading group of 32.

vals.ai's test asks 66 US regulatory research questions, and an answer counts only if it meets every point lawyers listed. Gemini 4 Neon scores highest, and Sonnet 5.5 and Gemini 4 Hydrogen share the top with it, 1.5 points behind. Every model below Neon here scores within the test's margin of error of it. Sort by price to see what the same result costs.

31 results the test cannot separate from the leader, on 66 questions. The cheapest of them, Muse Ember, costs 103 times less a question.

  1. Gemini 4 Neon71.21%$5.76
  2. Claude Sonnet 5.569.7%$11.67
  3. Gemini 4 Hydrogen69.7%$4.84
  4. Gemini 4 Argon68.18%$4.81
  5. Claude Fable 5.168.18%$20.31
  6. Claude Opus 565.15%$6.58
  7. Claude Opus 5.565.15%$20.86
  8. Gemini 4 Carbon65.15%$6.80
  9. Kimi K3Open63.64%$3.44
  10. Muse Wonder63.64%$3.92
  11. Muse Wonder Full63.64%$3.19
  12. Fireworks Ember 163.64%$3.17
  13. Grok 4.662.12%$1.52
  14. GLM-5.3Open62.12%$2.13
  15. Muse Spark 1.3 Max60.61%$0.54
  16. Inkling 2 Small60.61%$0.24
  17. GPT-5.6 Sol60.61%$19.69
  18. Astra for Law 6.260.61%$2.47
  19. MiMo V2.6 ProOpen59.09%$0.18
  20. Qwen3.8 Max59.09%$2.20
  21. DeepSeek V4.1 Flash 1017Open59.09%$0.65
  22. Claude Fable 5.259.09%$3.65
  23. GPT-6.2 Astra59.09%$11.82
  24. Claude Fable 557.58%$9.38
  25. GLM-5.3 FlashOpen57.58%$0.20
  26. GPT-5.6 Terra57.58%$6.87
  27. Kimi K3.1 ProOpen56.06%$6.34
  28. Qwen4 Max56.06%$2.07
  29. Grok 4.754.55%$4.18
  30. Claude Sonnet 554.55%$2.51
  31. Muse Ember54.55%$0.056
  32. Tencent Hy4 Preview54.55%$0.83

The margin ends here. The next result, DeepSeek V4.1 Flash, GPT-6 Astra, GPT-6.1 Sol, and GPT-5.5 at 53.03%, is the first the test can tell apart from the leader. Scores are drawn from 50%; prices on a scale where each step is ten times dearer.

vals.ai Legal Research, regulatory questions19 Nov 2026

Where Sonnet 5.5 lands on the other jobs.

Two more of the guide's jobs have a test that ran it.

Preference rating on real professional tasks (GDPval-AA), at the max setting

  1. Claude Opus 5.51866Tie
  2. Kimi K3.1 Pro1861Tie
  3. Muse Wonder Full1857Tie
  4. Claude Sonnet 5.51838Tie
  5. Claude Fable 5.21831Tie
  6. Qwen4 Max1809
  7. Muse Wonder1770
  8. Claude Fable 5.11765

Sonnet 5.5, Opus 5.5, Moonshot AI's Kimi K3.1 Pro, Meta's Muse Wonder Full, and Claude Fable 5.2 are level, inside the test's margin of error. Alibaba's Qwen4 Max is next, 29 points under Sonnet. The test can't separate it from Sonnet or Fable 5.2. It can separate it from Opus.

Artificial Analysis: models and the Intelligence IndexNo date stated

Claude Sonnet 5.5 for chemical distributors and producers.

Where its results point. No public test used a chemical company's own documents, so treat each of these as a job to test, not a result.

Research a Regulation

Find what a rule requires before you ship a grade to a new market, with the source text quoted beside the answer.

What to check

Make it quote the dated official text, and check that text yourself. The test is US law, not REACH or GHS.

See this job in the guide

Write a Customer Report

Turn your sales data, account notes, and market news into a quarterly review a customer will read.

What to check

Check every figure against its source. Opus 5.5 rates level on this work, and in one test an Opus task cost the same as a Sonnet task.

See this job in the guide

Read a Price List

Pull a supplier's multi-page price list into rows you can quote from.

What to check

It is 16th of 157 on tables. A free 1.2B model is first on tables in the independent document test, so try both on your longest lists.

See this job in the guide

Where Sonnet 5.5 costs more than it looks.

Its price per token is low. These are the places that price misleads.

Questions about Claude Sonnet 5.5.

Short answers, from the same sources as the rest of this page.

What is Claude Sonnet 5.5?

A general model from Anthropic, released on 28 September 2026. It reads text and images and costs $2 for every million tokens sent and $10 for every million written.

Is Claude Sonnet 5.5 the best model for regulatory research?

It is tied at the top of vals.ai's test of 66 US regulatory questions with Gemini 4 Hydrogen and Gemini 4 Neon. Sonnet and Hydrogen score 69.70% and Neon 71.21%, inside the test's margin. The test cannot separate the three from 29 other models, including open models at under 20 cents a question and Meta's Muse Ember at under 6 cents.

Is Claude Sonnet 5.5 cheaper than Claude Opus 5.5?

Per token, yes: half the price. Per task the gap disappears, because it uses more tokens. One index task cost $10.02 on Sonnet 5.5 and $10.04 on Opus 5.5. On regulatory questions Sonnet cost $11.67 a question against $20.86.

Can Claude Sonnet 5.5 read chemical price lists and supplier documents?

On ParseBench's table test it scores 91.07, 16th of 157, at 2.74 cents a page. A free 1.2B model, TeleOCR, is first on tables in the independent document test, so compare both on your own documents.

Does the regulatory result cover REACH or GHS?

No. The test is US law. For EU or GHS work, make it quote the dated official text, and check that text yourself.

Sources.

  1. vals.ai Legal Research, regulatory questions vals.ai, an independent testing company. Lawyers wrote and checked the questions; an answer counts only if it meets every point on the lawyers' list.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 cost per index task, and its professional work rating.States no date · Vailent checked 16 Nov 2026
  3. Artificial Analysis: Claude Sonnet 5.5 reaches #2 on the Intelligence Index Artificial Analysis, an independent testing company, on its own tests of the model.Published 28 Sep 2026 · Vailent checked 30 Sep 2026
  4. ParseBench leaderboard LlamaIndex, which sells LlamaParse, a product it ranks.Published 14 Nov 2026 · Vailent checked 15 Nov 2026
  5. Anthropic: Claude Sonnet The maker's own page.States no date · Vailent checked 30 Sep 2026
  6. Anthropic: pricing The maker's own page.States no date · Vailent checked 30 Sep 2026