Claude Sonnet 5.5

Claude Sonnet 5.5 is Anthropic\u2019s mid-tier Claude, released on 28 September 2026. The official API id is claude-sonnet-5-5. It keeps Sonnet 5\u2019s $2 / $10 list and the same 1M context / 128K output windows, and Anthropic positions it as the everyday model next to Opus 5.5.

The product claim is speed and token efficiency, not a new price tier. Anthropic says output is 30%+ faster than Sonnet 5, and that typical tasks can cost up to 30% less because the model uses fewer tokens at the same rates. On their announcement table, the jump that stands out is Terminal-Bench 4.0: 70.6% for Sonnet 5.5 against 10.3% for Sonnet 5.

Anthropic\u2019s own positioning: strongest at well-scoped everyday tasks, bugfix, and polished docs, slides, and spreadsheets. Opus 5.5 remains stronger at complex open-ended judgment. Sonnet 5.5 is the first Sonnet with cyber safeguards like Opus; high-risk cyber work falls back to Sonnet 5. Haiku 5.5 is described as coming in weeks. Zero data retention is the same policy as Opus 5.5 and Sonnet 5.

Same list price, same context and output caps. The difference Anthropic published is speed, token use, and the benchmark table below. These are Anthropic\u2019s numbers from the 28 September 2026 announcement — not independent verification.

The practical read of Anthropic\u2019s table: Sonnet 5.5 is not a cheaper Sonnet 5, it is a faster one at the same rates, with a much larger published gap on terminal and chart work than on FrontierCode. If a workload already fits Sonnet 5, the reason to look at 5.5 is latency and tokens-per-task, not a new list price.

Anthropic\u2019s table: $2 per million input tokens and $10 per million output tokens. Cache reads are $0.20; cache writes are $2.50. That is the same list as Sonnet 5.

Anthropic\u2019s cost claim on top of the rate card: up to 30% less per task because the model uses fewer tokens, plus 30%+ faster output. Those two are vendor claims about token use and speed, not a different published rate.

Opus 5.5 on the same Anthropic table is $4 / $20 input/output, cache reads $0.20, cache writes $5. Sonnet 5.5 is half Opus 5.5 on input and output at list.

Prices are Anthropic\u2019s published list. llmrelay does not serve Sonnet 5.5 yet, so no relay price is quoted on this page. The Claude mid-tier we can call today is Claude Sonnet 5.

Yes on Anthropic\u2019s own platforms. No on llmrelay, as of 2 October 2026.

• +Claude Platform, AWS, GCP, Azure — official API id claude-sonnet-5-5, per Anthropic\u2019s announcement and migration guide.
• +claude.ai free tier — Anthropic and Simon Willison (28 September 2026) both describe Sonnet 5.5 as the free-tier model there.
• +llmrelay — not served. Our live Claude mid-tier is still claude-sonnet-5.

Two API notes from Anthropic\u2019s docs that matter if you call it on their platform: thinking: {type:disabled} errors — use between_tools instead. Default effort is Medium in Claude apps and High on Claude Platform.

When llmrelay can route claude-sonnet-5-5, this status box and the page update the same day. Until then, any copy that implies a relay endpoint for 5.5 is ahead of our inventory.

Anthropic\u2019s answer is: for well-scoped everyday coding, often yes; for complex open-ended judgment, Opus stays ahead. The announcement table is mixed, which matches that split rather than a single winner.

Sonnet 5.5 leads Terminal-Bench 4.0 on that table; Opus 5.5 leads FrontierCode, CursorBench, HLE with tools, OSWorld, Chartography, and the two knowledge-work scores by a small margin. Combined with Anthropic\u2019s copy — well-scoped bugs and polish versus open-ended judgment — the coding default they are selling is Sonnet 5.5, with Opus as the upgrade when the task is not well-scoped.

Anthropic positions Sonnet 5.5 for Claude Code and Cursor everyday work: well-scoped coding, bugfix, docs, slides, and spreadsheets. On Claude Platform the model id is claude-sonnet-5-5. Default effort is Medium in Claude apps and High on Claude Platform.

That is Anthropic\u2019s path. It is not llmrelay\u2019s path yet. Pointing Claude Code or Cursor at llmrelay and requesting claude-sonnet-5-5 will not get this model — our abilities list still only has claude-sonnet-5 on the live Claude mid-tier channel.

If you need a Sonnet-class model through llmrelay today, use Claude Sonnet 5 (id claude-sonnet-5). If you need Sonnet 5.5 specifically, call Anthropic, AWS, GCP, or Azure until this page\u2019s status box changes.

Anthropic\u2019s models comparison table (Sonnet 5.5 column): 1,000,000 input tokens and 128,000 output tokens.

Those caps match Sonnet 5. Long-context work is not a reason to move from 5 to 5.5 on paper — the published difference is speed, tokens-per-task, and the benchmark table, not a larger window.

Anthropic published a GPT-6 Sol column on the same announcement table. The rows below are that column only — we are not adding OpenAI prices or other benches.

On Anthropic\u2019s table, GPT-6 Sol is ahead on FrontierCode 1.1 Main Max (49.3% vs 46.2%). Sonnet 5.5 is ahead on GDPval-AA v2.1 (1844 vs 1487), AA-Briefcase v1.1 (1811 vs 1483), and Chartography with no tools (61.6% vs 53.6%). Anthropic did not publish GPT-6 Sol scores for Terminal-Bench 4.0, CursorBench 4.0, HLE with tools, or OSWorld 2.1 in that table, so those rows are omitted here.

Vendor table, vendor numbers. Independent verification is a separate question, and this page does not claim it.