$ cat wiki/models/claude-sonnet-5-5.md
Claude Sonnet 5.5
Compared with
Anthropic's 2026-09-28 Sonnet release, six days after Claude Opus 5.5. It continues that release's argument — the same work for less — one tier down, and it carries one result that argument did not predict: on Terminal-Bench 4.0 Sonnet 5.5 scores 70.6% against Opus 5.5's 66.4%, so the cheaper model leads the flagship on the benchmark the flagship's own launch table was built around (source).
Read first-party over two passes. www.anthropic.com answered on both, the
sixth consecutive run it has done so, and platform.claude.com supplied every
spec figure the announcement omits — the announcement page states no context
window at all.
Spec
2026-10-07 pricing update: cache reads fell from $0.20/M to $0.10/M. Anthropic estimates around 20% lower cost on most agentic tasks; that is a workload-dependent saving, not a change to the $2/M input or $10/M output rates. (source)
| Attribute | Value |
|---|---|
| Developer | Anthropic |
| Released | 2026-09-28 |
| Announced | 2026-09-28 |
| Context window | 1M tokens |
| Pricing | $2/M input · $10/M output · cache reads $0.10/M (2026-10-07) |
| License | proprietary (API-only; no weight release) |
| Availability | Claude apps, Claude API as claude-sonnet-5-5, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, Claude Platform on AWS |
| At launch, the cache-read rule was the docs' stated rule rather than a quoted figure: the | |
| pricing note reads *"prompt cache reads cost 10% of the base input price (2.5% on | |
| Claude Fable 5.1 and Claude Mythos 5.1, 5% on Claude Opus 5.5)"*, and Sonnet 5.5 | |
| is named in neither exception | |
| (docs). This historical rule was superseded by the dated price cut above. |
Pricing is identical to Claude Sonnet 5's $2/$10, which makes the announcement's "costs up to 30% less for most work" a claim about tokens spent, not price per token — the saving is the effort setting, not the rate card.
Release Date
Announced and available the same day, 2026-09-28. Retirement committed not sooner than 2027-09-28 (docs).
Benchmarks
As published, with the comparison models the launch page names.
| Benchmark | Sonnet 5.5 | Claude Sonnet 5 | Claude Opus 5.5 | Other |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 70.6% | 10.3% | 66.4% | — |
| FrontierCode 1.1 (Main) | 46.2% (Max) | 42.4% | 54.4% | GPT-6 Sol 49.3% |
| CursorBench 4.0 | 55.5% | 34.1% | 57.8% | — |
| GDPval-AA v2.1 | 1844 | 1449 | 1846 | — |
| AA-Briefcase v1.1 | 1811 | 1359 | 1822 | — |
| Humanity's Last Exam | 64.5% (with tools) | 54.9% | 67.7% | — |
| OSWorld 2.1 | 80.1% (partial) | 57.0% | 81.8% | — |
| Chartography | 61.6% (no tools) | 15.6% | 64.4% | — |
| **The Terminal-Bench row is not a typo and it is not a contradiction of this | ||||
| wiki.** The launch page gives Sonnet 5's Terminal-Bench 4.0 as 10.3%, and it | ||||
| was re-read on a second pass for exactly that reason. Claude Sonnet 5 | ||||
| holds 80.4% on Terminal-Bench 2.1. Both are true: **they are different | ||||
| harnesses two major versions apart**, and the 70.1-point spread between them on | ||||
| one model is a measurement of the harness, not of the model. This is the failure | ||||
| mode Eval Harness Configuration exists to track, and it is the | ||||
| clearest instance the wiki holds — a benchmark name that looks like a series is | ||||
| not a series. |
Sonnet 5.5 leads Opus 5.5 on exactly one row (Terminal-Bench 4.0) and trails it on the other seven, by margins from 2 points (GDPval-AA v2.1, 1844 vs 1846) to 8.2 (FrontierCode 1.1). The announcement states the GDPval-AA gap itself: Sonnet 5.5 sits "two points below Opus 5.5 on GDPval-AA, a test of real-world work across a variety of occupations".
Two of this table's Opus 5.5 cells disagree with that model's own launch page — Chartography 64.4% here against 89.0% there, and OSWorld named 2.1 here and 2.0 there — and both are disclosed on Claude Opus 5.5's ## Conflicting Reports. The Chartography gap lines up with this table's no tools label against that one's with tools.
No per-effort absolute figure is published. The page asserts that on several benchmarks Sonnet 5.5 at Low or Medium effort beats Sonnet 5's best score for about a tenth of the cost per task, and renders the comparison as charts rather than a table, so the individual points are not quotable.
Use Cases
Anthropic's stated positioning: well-scoped everyday tasks, fixing bugs, and creating polished documents, slides and spreadsheets — the faster, lower-cost complement to Opus 5.5, which it reserves for complex work requiring careful judgment (source).
Two capability claims sit outside coding:
- Computer use. OSWorld 2.1 80.1% (partial), against Sonnet 5's 57.0% — a 23.1-point gain on the tier's weakest axis, and within 1.7 of Opus 5.5.
- Screenshot-only long-horizon play. "it's the first Sonnet model to beat Pokémon Red working only from screenshots" — the qualifier is the claim, since the feat itself is not new to the Claude line, only to this tier and to this input channel.
The docs' default-model recommendation is unchanged by the release: Opus 5.5 for most workloads, Claude Fable 5.1 for demanding reasoning and long-horizon agentic work. Sonnet 5.5 is not positioned as anyone's default.
Compared To
- Claude Sonnet 5 — direct predecessor, same $2/$10 rate card. Beaten on all eight published rows, and by a wide margin on the three that changed harness version or measure tools (Terminal-Bench 4.0, Chartography, OSWorld 2.1).
- Claude Opus 5.5 — released six days earlier at $4/$20. Default
effort
medium, where Sonnet 5.5's ishigh(docs), so the two models' headline numbers are not produced at the same setting and Eval Harness Configuration holds why that matters. - Claude Fable 5.1 — remains the line's reasoning and long-horizon model; at Sonnet 5.5's launch, its cache reads were priced at 2.5% of base input against Sonnet 5.5's 10%. Sonnet's later cut is recorded above.
- GPT-6.1 Sol — OpenAI's release of the following day, at the identical $2/M input · $10/M output. Two labs' mid-tier models now carry the same rate card, and both launched on a cost headline that is not a change to that card: this page's "costs up to 30% less for most work" is about tokens spent, and GPT-6.1 Sol's "approximately one-fifth" is measured against Astra rather than against the model it upgrades, whose rate it matches exactly. The two share no benchmark, so the shared price is the only comparison their announcements support (source).
Sources
- Introducing Claude Sonnet 5.5 (snapshot)
- Claude models overview — context window, max output, pricing, model IDs, cutoffs, default effort, retirement date