Secret Chat Logo

Claude Sonnet 5.5 Is Here:
Faster Everyday Claude Answers on Secret Chat AI

Launch Private Chat

Published September 29, 2026

Claude Sonnet 5.5 replaced Claude Sonnet 5 on Secret Chat AI today. Anthropic released the model yesterday, September 28. If you already use Claude Web Search, there is nothing to install or select: open an existing chat and continue. Under the next reply, the app names the model that wrote it, “Claude Sonnet 5.5.”

This is the Claude tier you meet when a question needs a quick web lookup or a substantial generator result. Anthropic calls it “the best combination of speed and intelligence.” Its own measurements show a large gain over Sonnet 5, while Anthropic still recommends Opus 5.5 for complex, open-ended work that needs sustained judgment.

What Changed From Claude Sonnet 5

Anthropic says Sonnet 5.5 writes output more than 30% faster than Sonnet 5. In its early testing, it also completed tasks in fewer steps and batched tool calls more effectively. That matters when a web answer needs several pages or a coding task needs several checks; the model spends less time moving between them.

The model takes text and images, uses adaptive thinking, has a one-million-token context window and supports up to 128K output tokens. Its reliable knowledge cutoff is June 2026. It uses the same tokenizer as Sonnet 5, so the same text produces the same token count. On Secret Chat, Sonnet 5.5 runs at medium thinking effort, as Sonnet 5 did, to keep Web Search reasoning useful without a long wait.

Anthropic points to bug fixing and polished documents, slides and spreadsheets as strengths. The company also made this the first Sonnet with classifiers against attempts to extract its reasoning, and deployed cyber safeguards similar to those on its most capable models. These are Anthropic's descriptions of the release, not Secret Chat tests.

What Anthropic's Benchmarks Show

On Anthropic's published Terminal-Bench 4.0 results, Sonnet 5.5 scored 70.6%, Sonnet 5 scored 10.3%, and Opus 5.5 scored 66.4%. That is the one listed benchmark where Sonnet 5.5 leads Opus 5.5; the Opus result is reported at Xhigh effort, its highest score. Anthropic's table has no GPT 6 Sol score for Terminal-Bench 4.0.

On FrontierCode 1.1 (Main), Sonnet 5.5 scored 46.2% at Max effort, behind Opus 5.5 at 54.4% and GPT 6 Sol at 52.1% at Xhigh effort. Anthropic says Sonnet 5.5 scored lower at Max than at Xhigh because it more often ran a multi-agent code review, leading to timeouts or edits beyond the task's scope; FrontierCode penalizes those extra edits.

The picture changes with the task. On CursorBench 4.0, Anthropic reports 55.5% for Sonnet 5.5, 34.1% for Sonnet 5 and 57.8% for Opus 5.5. On GDPval-AA v2.1, the scores are 1844, 1449, 1846 and 1487 for Sonnet 5.5, Sonnet 5, Opus 5.5 and GPT 6 Sol respectively. On AA-Briefcase v1.1, Sonnet 5.5 scored 1811, Opus 5.5 1822 and GPT 6 Sol 1483. Artificial Analysis ran GDPval-AA and AA-Briefcase on a pre-release Sonnet 5.5 deployment with a structured-outputs bug, since fixed, that Anthropic expects may have slightly understated its scores. Anthropic's table has no GPT 6 Sol score for CursorBench 4.0.

On Humanity's Last Exam with tools, Sonnet 5.5 scored 64.5% against Opus 5.5's 67.7%. For screenshot-driven computer use, Anthropic reports 80.1% for Sonnet 5.5, 57.0% for Sonnet 5 and 81.8% for Opus 5.5 on OSWorld 2.1; all three results are marked partial. On Chartography without tools, the figures are 61.6% for Sonnet 5.5, 15.6% for Sonnet 5, 64.4% for Opus 5.5 and 53.6% for GPT 6 Sol. Anthropic's table has no GPT 6 Sol score for Humanity's Last Exam or OSWorld 2.1.

OpenAI recently fixed a bug that degraded GPT 6 Sol's image understanding, so the published GDPval-AA, AA-Briefcase and Chartography scores may not reflect the fix yet; Artificial Analysis expects no major impact on its two scores, and Anthropic's internal Chartography testing suggested no impact. All these scores are from Anthropic's published tables, not evaluations run by Secret Chat; a benchmark score does not settle which model will answer your particular question best.

Where Sonnet 5.5 Fits Among the Current Models

Within Claude, Sonnet 5.5 is the middle choice for everyday work. Claude Haiku 4.5 is faster and cheaper for short answers. Claude Opus 5.5 is stronger when a task needs sustained judgment, and costs more to run. Claude Fable 5.1 remains the separate manual choice for especially demanding reasoning and long work; it is much more expensive to run than Sonnet 5.5. Sonnet 5.5 costs the same as Sonnet 5 did. Anthropic says Claude Haiku 5.5 will follow in the coming weeks.

Against GPT 6 Sol in Anthropic's published table, Sonnet 5.5 leads on GDPval-AA, AA-Briefcase and Chartography, but trails on FrontierCode. Anthropic reports no GPT 6 Sol score for Terminal-Bench 4.0, CursorBench 4.0, Humanity's Last Exam or OSWorld 2.1. It provides no like-for-like published comparison with Gemini 3.8 Flash or Grok 4.7.

The comparable routes in other Secret Chat families solve similar jobs with different models. ChatGPT Web Search currently uses GPT 6 Sol, while GPT 6 Luna handles its fast tier and routing. Gemini Web Search and Research both use Gemini 3.8 Flash. Grok Web Search and Research use Grok 4.7. These are choices of mode and provider, rather than a ranking from a shared Secret Chat benchmark. If you want a quick sourced Claude answer, choose Claude Web Search; for a harder Claude investigation, choose Research and get Opus 5.5.

Where Claude Sonnet 5.5 Runs on Secret Chat

  • Claude Web Search uses Sonnet 5.5 with Anthropic's built-in web search and web fetch tools.
  • Smart Agentic Claude uses Claude Haiku 4.5 to route a question. A quick web lookup goes to Sonnet 5.5; a harder question or Research goes to Claude Opus 5.5; a quick answer stays on Haiku 4.5.
  • AI Council puts Haiku 4.5 in the Claude seat on its lighter rung and Sonnet 5.5 on its heavier rung. Opus 5.5 is not a council seat.
  • Generators offer a “Claude Sonnet” row among six model choices in Email, Translator and Rewriter. The Claude row in Document, Resume, Table, Chart and Logo now runs Sonnet 5.5. In our own testing, Haiku 4.5 invented details in documents and drew weak logos, so these generators use the stronger tier.

Haiku 4.5 still handles Claude Fast Chat, routing, chat titles and summaries. Claude Opus 5.5 still handles Research and Smart Agentic escalations. Claude Fable 5.1 stays a separate manual mode. Answers written by Sonnet 5 before today's switch are still found and deleted the same way where the provider allows it, and the Session Privacy report lists them.

How Anonymous Access Works

Secret Chat sends your prompt to Anthropic without a personal account at Anthropic or a profile built around you. We store no archive of chats on our servers, and prompts are never used for training.

The prompt goes to Anthropic verbatim so Sonnet 5.5 can answer. Anonymity does not remove names, case numbers or other identifying details from what you type. Redact sensitive details before sending; that part is your responsibility.

Frequently Asked Questions

  1. When did Claude Sonnet 5.5 arrive?

    Anthropic released it on September 28, 2026. It replaced Claude Sonnet 5 on Secret Chat AI on September 29, 2026.

  2. Do I need to start a new chat?

    No. Existing chats continue on the new model in the modes that use Sonnet. Under each new reply, Secret Chat shows the model that wrote it: “Claude Sonnet 5.5.”

  3. What improved over Claude Sonnet 5?

    Anthropic says output is more than 30% faster, with fewer steps and better batching of tool calls. Its published Terminal-Bench 4.0 and CursorBench 4.0 results also show large gains over Sonnet 5; those are Anthropic's figures.

  4. Is Sonnet 5.5 stronger than Claude Opus 5.5?

    It leads on Anthropic's Terminal-Bench 4.0 table only. It trails Opus 5.5, mostly narrowly, on CursorBench 4.0, GDPval-AA v2.1, AA-Briefcase v1.1, Humanity's Last Exam, OSWorld 2.1 and Chartography, and by a wider margin on FrontierCode 1.1 (Main). Anthropic says Opus 5.5 remains clearly stronger for complex, open-ended work requiring sustained judgment.

  5. How does Claude Sonnet 5.5 compare with GPT 6 Sol?

    In Anthropic's published table, Sonnet 5.5 leads on GDPval-AA v2.1 (1844 vs 1487), AA-Briefcase v1.1 (1811 vs 1483) and Chartography without tools (61.6% vs 53.6%). It trails on FrontierCode 1.1 (Main): 46.2% at Max effort vs 52.1% for GPT 6 Sol at Xhigh effort. The table has no GPT 6 Sol score for Terminal-Bench 4.0, CursorBench 4.0, Humanity's Last Exam or OSWorld 2.1. OpenAI has since fixed an image-understanding bug, so the GDPval-AA, AA-Briefcase and Chartography figures may not reflect the fix; Artificial Analysis expects no major impact on its two scores, and Anthropic's internal Chartography testing suggested no impact. These are Anthropic's published tables, not Secret Chat tests.

  6. Which Secret Chat modes use Sonnet 5.5?

    Claude Web Search, Smart Agentic Claude's quick web lookup tier, the heavier Claude seat in AI Council, and the Claude rows in Document, Resume, Table, Chart and Logo generators. Email, Translator and Rewriter offer it through their “Claude Sonnet” model row.

  7. Does Claude Research use Sonnet 5.5?

    No. Claude Research and harder Smart Agentic escalations use Claude Opus 5.5. Fast Chat, routing, titles and summaries use Claude Haiku 4.5; Claude Fable 5.1 remains a separate manual mode.

  8. Does Sonnet 5.5 cost more than Sonnet 5?

    No. Sonnet 5.5 costs the same as Sonnet 5 did. It is cheaper to run than Claude Opus 5.5 and much cheaper than Claude Fable 5.1; Haiku 4.5 is cheaper still.

  9. What does anonymous access protect?

    You need no personal Anthropic account, Secret Chat builds no profile and stores no server chat archive, and prompts are never used for training. Your prompt still goes to Anthropic verbatim. Remove sensitive details before sending.

Sources

Related Reading