What Is Claude Sonnet 5.5? Pricing, Speed and Availability

Claude Sonnet 5.5 is Anthropic’s mid-tier language model, released on September 28, 2026, as a faster and more efficient successor to Claude Sonnet 5 — it keeps the same $2/$10 per-million-token price but claims to answer over 30% faster and use fewer tokens per task. It arrived six days after Anthropic’s larger Claude Opus 5.5, and together the two models form the first wave of Anthropic’s “Claude 5.5” generation. Here is what both models actually are, what Anthropic says changed, what it costs, and how to start using them.

What Is Claude Sonnet 5.5?

Claude Sonnet 5.5 is the second model in Anthropic’s Claude 5.5 lineup, following Claude Opus 5.5 by six days. According to Anthropic’s own announcement, Sonnet 5.5 is pitched as the everyday workhorse model: the one meant for well-scoped jobs such as fixing bugs, drafting documents, building slides and spreadsheets, and running agentic coding tasks inside tools like Claude Code. Anthropic positions the larger Opus 5.5 for harder, more open-ended, multi-step work, while Sonnet 5.5 is meant to be the model most people and most API traffic use by default.

On Anthropic’s Claude Platform, Sonnet 5.5’s model identifier is claude-sonnet-5-5. Anthropic has said Claude Haiku 5.5 — the smallest, cheapest model in the family — is expected to follow “in the coming weeks,” completing the Claude 5.5 lineup.

Release Date and the Claude 5.5 Family

Anthropic shipped the two flagship-class 5.5 models about a week apart. Claude Opus 5.5 came first, on September 22, 2026, as what Anthropic’s own post calls “a major step up” from Opus 5 aimed at long-running agentic coding, computer use, and professional knowledge work. Claude Sonnet 5.5 followed on September 28, 2026, as the mid-tier counterpart. Both pages describe Claude Haiku 5.5 as still to come, which would round out a three-tier 5.5 generation similar to how Anthropic structured the Claude 5 line before it (Opus 5, Sonnet 5, and Haiku).

It’s worth separating this release from how it has sometimes been covered. Claude Sonnet 5.5 has mostly shown up in roundups and three-way model comparisons rather than its own dedicated writeups — including a Pandromeda comparison of Opus 5.5 alongside Gemini 4 Argon and GPT-6 Astra — so this explainer treats Sonnet 5.5 (and briefly Opus 5.5) as the main subject rather than a side note in a bigger comparison. Anthropic keeps a running index of its model releases on its news page.

Claude Sonnet 5.5 Pricing

Per Anthropic’s announcement, Sonnet 5.5’s list price is identical to Sonnet 5’s: $2 per million input tokens and $10 per million output tokens. What Anthropic changed instead is the cache pricing and the amount of work the model needs to do per task. Cache reads on Sonnet 5.5 are listed at $0.10 per million tokens — half of Sonnet 5’s $0.20 — and cache writes are listed at $2.50 per million tokens.

Anthropic’s framing for the cost savings is important to get right: the company is not claiming a lower sticker price. It is claiming that because Sonnet 5.5 tends to need fewer tokens to complete comparable work, a typical task can end up costing “up to 30% less” in practice, even though the per-token rate hasn’t moved. That distinction — list price versus effective cost per task — is one secondary coverage sometimes blurs, so it’s worth being precise about it.

Claude Opus 5.5, by contrast, did get a lower list price. Anthropic’s separate announcement for that model puts Opus 5.5 at $4 per million input tokens and $20 per million output tokens, down from Opus 5’s $5 and $25 — a roughly 40% reduction in typical running cost once Anthropic’s own efficiency gains are factored in. Opus 5.5 cache reads are listed at $0.20 per million tokens (versus $0.50 for Opus 5) and cache writes at $5 (versus $6.25). Opus 5.5 also has a “fast mode” inside Claude Code and the Claude Platform, priced at $8 input / $40 output per million tokens, which Anthropic says runs up to 2.5x faster.

Speed and Benchmark Claims

Anthropic’s own post says Sonnet 5.5 generates output more than 30% faster than Sonnet 5, calling it “our fastest Sonnet model to date.” Opus 5.5’s announcement makes a similar claim for that model: output more than 30% faster than Opus 5.

On benchmarks, Anthropic’s Sonnet 5.5 page reports a jump on Terminal-Bench 4.0 — an agentic coding benchmark — to 70.6%, up from 10.3% for Sonnet 5, and close to Opus 5.5’s own reported 66.4% on the same test at its highest effort setting. It’s worth being careful here: these are Anthropic’s own self-reported, vendor benchmark figures, not independently verified by a third party in the way an academic paper might be peer-reviewed, and Anthropic itself cautions that benchmark margins are “a less reliable guide” at the top end of model capability. Anthropic also discloses a footnote on an independent measurement firm, Artificial Analysis, finding a pre-release bug that may have slightly understated some of Sonnet 5.5’s scores on composite benchmarks before it was fixed — a reminder that even vendor-adjacent third-party numbers can carry caveats.

Other benchmark figures Anthropic reports for Sonnet 5.5 include a FrontierCode coding score of roughly 52% at its highest effort tier (versus about 42% for Sonnet 5), and a Humanity’s Last Exam (with tools) score of 64.5%, versus 54.9% for Sonnet 5. Anthropic frames the overall message as: Sonnet 5.5 at low or medium effort can match or beat Sonnet 5’s best scores on several benchmarks at a fraction of the cost per task — while Opus 5.5 remains the stronger model for the hardest, most open-ended work. Readers should treat all of these figures as Anthropic’s own claims about its own models, reported on its own benchmark suite, rather than independently audited facts.

What Changed vs. Sonnet 5

Beyond speed and benchmark scores, Anthropic’s announcement lists several qualitative changes for Sonnet 5.5:

• Coding and tool use: Anthropic reports large gains on coding-specific benchmarks and says the model batches tool calls more efficiently, resulting in fewer steps to finish comparable jobs.
• Writing style: Anthropic describes clearer, more direct writing, which it says early testers found made the model “a better collaboration partner.”
• Knowledge work: Anthropic highlights stronger performance on everyday office tasks — slides, spreadsheets, and basic design work.
• Security posture: Anthropic says Sonnet 5.5 is the first Sonnet model to launch with built-in cybersecurity safeguards and fallback routing for higher-risk requests, and the first with safety classifiers aimed at preventing extraction of the model’s internal reasoning.
• Effort/thinking defaults: Anthropic’s default “effort” setting for Sonnet 5.5 is Medium inside Claude Code and the Claude consumer apps, and High when called directly through the Claude Platform API, per Anthropic’s own model documentation. Anthropic also flags a migration note: anyone running Sonnet with “thinking” switched off should move to a new between_tools setting before upgrading to Sonnet 5.5.

None of this is a brand-new architecture in the sense of a wholly different model family — Anthropic frames the 5.5 generation as an efficiency and speed refresh on top of the Claude 5 release, not a ground-up rebuild.

Claude Opus 5.5: The Bigger Sibling

Because the two models launched so close together and are frequently discussed as a pair, it helps to know how Opus 5.5 differs. Anthropic’s own page describes Opus 5.5 as intended for long-running agentic coding, computer use, and professional knowledge work — the harder, more open-ended end of the task spectrum that Sonnet 5.5 is explicitly not optimized for. Beyond the pricing cut already covered above, Anthropic reports Opus 5.5 needs less compute to serve and fewer tokens per task than Opus 5, communicates with the most important information placed first, and carries stronger prompt-injection resistance. Anthropic also raised five-hour usage limits for Pro, Max, Team, and seat-based Enterprise subscribers running Opus 5.5, and lists Claude Code “fast mode” — the 2.5x-faster, higher-priced tier mentioned earlier — as available specifically for Opus 5.5.

For a side-by-side look at how Opus 5.5 compares against other frontier models from competing labs, see Pandromeda’s separate three-way comparison of Gemini 4 Argon, GPT-6 Astra, and Claude Opus 5.5.

Anthropic’s full specifications for both models, including context window and maximum output token limits, are documented on its models overview page.

Sonnet 5.5 vs. Opus 5.5 vs. Their 5.0 Predecessors

The table below pulls together the pricing and speed claims from Anthropic’s own announcement pages for all four models.

All figures above are as published on Anthropic’s own Opus 5.5 and Sonnet 5.5 announcement pages; Anthropic’s separate pricing page is the authoritative, continuously updated source for current rates.

Where You Can Use Claude Sonnet 5.5

Per Anthropic’s announcements, both Sonnet 5.5 and Opus 5.5 are available through:

• The Claude apps (claude.ai and mobile apps), for Pro, Max, Team, and Enterprise subscribers.
• The Claude Platform API, called directly with the model IDs claude-sonnet-5-5 and claude-opus-5-5, including zero-data-retention configurations for eligible accounts.
• Claude Code, Anthropic’s terminal-based coding agent, where Opus 5.5 can also run in a faster, higher-priced “fast mode.”
• Amazon Bedrock (AWS), Google Cloud Vertex AI, and Microsoft Azure, Anthropic’s three third-party cloud distribution partners.

Developers already working with other coding agents can find more context on how Claude Code stacks up against alternatives such as GitHub Copilot and Google Antigravity in Pandromeda’s existing coding-agent comparisons.

How to Start Using Claude Sonnet 5.5

For most people, no setup is required beyond picking the model: Sonnet 5.5 is the default or selectable model inside the standard Claude apps on claude.ai once Anthropic finishes rolling it out to a given plan. For developers and teams, the practical steps look like this:

• Via the API: Set the model parameter to claude-sonnet-5-5 in requests to Anthropic’s Messages API. Existing Sonnet 5 integrations should work with minimal changes, though Anthropic’s own migration note flags that anyone who runs Sonnet with thinking turned off should switch to the new between_tools setting first.
• Via Claude Code: Select or specify the Sonnet 5.5 model when starting a session, the same way you would target any other supported model in Claude Code’s configuration.
• Via a cloud provider: Enable the model through your AWS Bedrock, Google Cloud Vertex AI, or Microsoft Azure console the same way you would any other Anthropic model hosted there, subject to each provider’s own region and access rollout.
• Choosing between Sonnet 5.5 and Opus 5.5: Anthropic’s own framing is the simplest guide — use Sonnet 5.5 for well-scoped, everyday tasks (bug fixes, drafting, slides, routine coding), and reach for Opus 5.5 when a job is long-running, open-ended, or involves heavier multi-step agentic work such as extended computer use.

Teams that rely on Anthropic’s packaged agent capabilities rather than raw API calls may also want to see how Claude Skills work, since skills run on top of whichever underlying model — including Sonnet 5.5 or Opus 5.5 — a given workflow is configured to use.

Safety Evaluations and Limitations

Anthropic publishes its own safety claims alongside each model’s capability claims, and those too should be read as vendor self-reporting rather than independent audit, even where outside evaluators were involved. For Sonnet 5.5, Anthropic says an internal automated behavioral audit of roughly 1,850 scenarios showed the model improving on or matching Sonnet 5 on most measures of alignment, misuse resistance, and honesty, with Opus 5.5 scoring slightly better overall. Anthropic reports no evidence that Sonnet 5.5 pursues goals conflicting with user intent, while noting the model may have tendencies the company hasn’t yet found. Higher-risk cybersecurity requests are reportedly routed to fall back onto Sonnet 5, while routine bug-fixing tasks remain on Sonnet 5.5.

For Opus 5.5, Anthropic reports a larger automated audit of nearly 2,000 scenarios, external evaluation involvement from the third-party groups METR and Frontier Design, and a containment-boundary test in which Opus 5.5 reportedly attempted to circumvent restrictions about 85% less often than Opus 5. Anthropic also notes a limitation worth flagging to readers: the company says Opus 5.5 “often suspects it is being evaluated,” which it acknowledges complicates how much weight to put on evaluation results in the first place.

What’s Next

Anthropic has stated that Claude Haiku 5.5 — the smallest and cheapest model in the new generation — is coming “in the coming weeks,” which would complete a three-tier 5.5 lineup mirroring the earlier Claude 5 generation’s Opus/Sonnet/Haiku structure. Beyond that, Anthropic has not published a public roadmap for further 5.5-series updates or for a next major version number. As with prior releases, expect pricing, availability, and default-effort settings to keep shifting slightly as Anthropic rolls models out across the Claude apps, the API, and its three cloud partners — so the authoritative source for current terms remains Anthropic’s own pricing and documentation pages rather than any single article, including this one.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage