Claude Sonnet 5.5 Release Date, Price and What's New

Claude Sonnet 5.5 launched on September 28, 2026, priced exactly the same as its predecessor at $2 per million input tokens and $10 per million output tokens — but Anthropic says it runs more than 30% faster and, because it needs fewer tokens to finish the same task, costs up to 30% less per task in practice. It's available immediately on the Claude API (model string claude-sonnet-5-5), in Claude Code, on claude.ai and the Claude apps, and through Amazon Web Services, Google Cloud, and Microsoft Azure.

What's new in Claude Sonnet 5.5

Anthropic is positioning Sonnet 5.5 as a direct, same-price successor to Sonnet 5 rather than a new pricing tier: identical per-token rates, the same 1M-token context window on supported platforms, and the same general role as the mid-tier model for coding and agentic work. The headline change is efficiency. Anthropic writes that Sonnet 5.5 "typically needs far fewer tokens to do the same work" as Sonnet 5, which is what drives both the speed gain and the cost savings, since token count directly determines latency and, ultimately, the bill.

The efficiency claim shows up most clearly on agentic coding benchmarks. Anthropic says that at Medium effort — the default setting in the Claude apps and Claude Code — Sonnet 5.5 "far exceeds Sonnet 5's best score for less than a tenth of the cost per task" on Terminal-Bench 4.0, a benchmark that measures how well a model completes complex, multi-step tasks inside a command-line interface. On a separate coding benchmark, FrontierCode 1.1, Anthropic says that at High effort — the default on the Claude Platform — Sonnet 5.5 "matches GPT-6 Sol's best score for about a fifth of the cost per task."

Two customers quoted in Anthropic's announcement describe the same pattern from the outside. An Epic Games executive said Sonnet 5.5 "cleared the same quality bar you'd expect from a higher-tier model," while an engineering lead at AI app-building platform Base44 said the model completed builds in 3.6 iterations on average, compared with 7.7 iterations for Opus 5 on the same tasks — roughly half the back-and-forth for a comparable result.

Benchmark results: Sonnet 5.5 vs. Sonnet 5, Opus 5.5, and GPT-6 Sol

Anthropic published a full benchmark table alongside the release, comparing Sonnet 5.5 against Sonnet 5, Claude Opus 5.5, and OpenAI's GPT-6 Sol (Anthropic notes that OpenAI did not publicly report GPT-6 Sol's performance on some of these tests, and that Anthropic substituted GPT-5.6 Sol results where noted). The gap between Sonnet 5.5 and Sonnet 5 is largest on agentic and computer-use tasks; the gap between Sonnet 5.5 and the pricier Opus 5.5 is narrow on most tests.

A few things stand out in that table. On Terminal-Bench 4.0, Sonnet 5.5 doesn't just beat Sonnet 5 — at 70.6% it actually edges out Opus 5.5's 66.4% (Anthropic notes the Opus 5.5 figure is reported at Xhigh effort, its highest setting). On the knowledge-work tests, GDPval-AA and AA-Briefcase, Sonnet 5.5 lands within roughly 10-15 points of Opus 5.5 out of totals near 1,800-1,850, a small gap given the price difference between the two models. Anthropic also flags one caveat on the GDPval-AA and AA-Briefcase numbers: independent benchmark group Artificial Analysis ran those two tests on a pre-release deployment of Sonnet 5.5 that had a bug affecting structured-output requests, which Anthropic expects to have understated Sonnet 5.5's true score, if it had any effect at all.

Chartography, a visual-chart-reading test, shows the widest relative jump: Sonnet 5.5 answers correctly 61.6% of the time with no tool access, up from 15.6% for Sonnet 5 — a four-fold improvement that Anthropic attributes to broader gains in image understanding and long-horizon work.

Anthropic's table also includes GPT-6 Sol as a reference point where public data exists. On FrontierCode 1.1, GPT-6 Sol scores 49.3%, putting it between Sonnet 5.5's 46.2% (Max effort) and Opus 5.5's 54.4%; on Chartography, GPT-6 Sol scores 53.6%, below Sonnet 5.5's 61.6%. Anthropic notes that OpenAI hasn't published GPT-6 Sol results for Terminal-Bench 4.0, CursorBench 4.0, Humanity's Last Exam, or OSWorld 2.1, so no comparison is shown on those rows. Read individually, these numbers are Anthropic's own reported results rather than independently reproduced scores, which is worth keeping in mind when comparing across vendors.

Sonnet 5.5 vs. Sonnet 5: what actually changed

Because the sticker price per token is identical between the two models, the real-world savings Anthropic advertises come entirely from Sonnet 5.5 completing tasks in fewer tokens, not from a discounted rate card. That distinction matters for budgeting: a workload billed by the task (a code review, a customer ticket, an agent run) should get cheaper under Sonnet 5.5, but a workload billed by raw token volume at a fixed ratio of input to output won't see any change in the per-token math.

Alignment and safety changes

Anthropic says its automated behavioral audit found Sonnet 5.5 improves on or matches Sonnet 5 on most measures of alignment. Because the new model's cybersecurity capabilities are comparable to Opus 5's, Anthropic says Sonnet 5.5 is the first Sonnet-tier model to launch with the same cyber safeguards and fallback behaviors built for Anthropic's most capable models. Its biology safeguards carry over unchanged from Sonnet 5. Anthropic describes both safeguards as targeting "a narrow set of high-risk requests," with routine software development and most life-sciences work unaffected.

Sonnet 5.5 also becomes the first Sonnet model to ship with safety classifiers specifically aimed at preventing reasoning extraction, and it expands Anthropic's "preserved thinking" mechanism, which binds a model's internal reasoning to the specific model and conversation that produced it (documented in Anthropic's preserved-thinking docs). In practice, this means reasoning (thinking) blocks generated by Sonnet 5.5 are not portable to other models and, in enforced accounts, can't be replayed after earlier turns in a conversation are edited.

How to access Claude Sonnet 5.5

Developers can call the model today on the Claude Platform using the model string claude-sonnet-5-5. It's also live in Claude Code and in the consumer and business Claude apps, and Anthropic says it has rolled out simultaneously across all three major clouds: Amazon Web Services (Bedrock), Google Cloud (Vertex AI), and Microsoft Azure (Foundry). As with Opus 5.5 and Sonnet 5, Sonnet 5.5 is available with zero data retention for organizations that need it.

There are two migration details worth flagging for anyone porting existing Sonnet 5 code over. First, Sonnet 5.5 no longer accepts the old way of disabling extended thinking outright; teams that previously ran Sonnet with thinking turned off need to switch to a new between_tools setting to get equivalent behavior, and that setting only works at High effort or below. Second, reasoning effort now defaults differently depending on where the model runs: Claude Code and the Claude apps default to Medium effort, while the Claude Platform (the raw API) defaults to High. Anthropic frames the effort dial as the primary way to trade cost and speed against thoroughness — lower settings answer faster and use fewer tokens for routine work, higher settings reason longer and check their own output more before answering, which suits harder or higher-stakes tasks.

For teams already comparing Claude against other assistants on price and plans rather than raw API access, Pandromeda's plan-by-plan comparison of ChatGPT, Claude, and Gemini is worth checking against Sonnet 5.5's numbers above.

Where this fits: Anthropic's September 2026 model refresh

Sonnet 5.5 is the third Claude release in under a month. Anthropic opened the cycle on September 1, 2026 with Claude Fable 5.1 and Claude Mythos 5.1, its most capable and most expensive models, aimed at long-running agentic and coding work and priced at $10 per million input tokens and $50 per million output tokens. On September 22, Anthropic followed with Claude Opus 5.5, which the company says performs at roughly the level of Fable 5.1 on most work while costing 40% less to run than the outgoing Opus 5 — $4 per million input tokens and $20 per million output tokens, down from Opus 5's $5/$25.

Sonnet 5.5 now closes out the refresh at the mid tier, undercutting Opus 5.5 on price by half while, according to Anthropic's own benchmark table above, narrowing the performance gap between the two models on several agentic coding tasks to single digits. Anthropic's stated logic across all three releases has been consistent: each new model in a given tier should do more per dollar than the model it replaces, rather than simply adding raw capability at a higher price. For a broader look at how Opus 5.5 stacks up against competing flagship models from OpenAI and Google, see Pandromeda's Gemini 3.8 Flash vs. GPT-6 vs. Claude Opus 5.5 comparison.

Taken together, the three releases give Anthropic a full current-generation lineup spanning roughly a 5x price range: Sonnet 5.5 at $2/$10 per million tokens for everyday coding and agent work, Opus 5.5 at $4/$20 for harder daily-driver tasks, and Fable 5.1 at $10/$50 for the most demanding long-horizon reasoning and research work. All three share the same 1M-token context window on supported platforms and the same general API surface, which should make it straightforward for developers to test a task across tiers and pick the cheapest model that clears their own quality bar.

What's next

Anthropic has not announced a release date for the next model in the Sonnet line, and its recent cadence — three model releases within a four-week span — suggests further updates could continue to arrive in quick succession rather than on a fixed annual schedule. For now, Sonnet 5.5 is Anthropic's clearest pitch yet for teams that want most of Opus-tier agentic coding performance without Opus-tier pricing, while Opus 5.5 and Fable 5.1 remain the options for work that demands the highest available accuracy regardless of cost. Developers evaluating all three models against competitors should watch Anthropic's newsroom and API documentation directly, since effort defaults, tool-use behavior, and safety classifier coverage have all shifted with each release in this cycle.

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