Anthropic introduced Claude Sonnet 5.5, the second model in the Claude 5.5 family. Official framing: a clear upgrade over Claude Sonnet 5 that runs 30%+ faster and costs up to 30% less for most work — not from a lower list price, but from fewer tokens per task at the same Sonnet 5 rates. It is the mid-tier complement to Claude Opus 5.5: Opus for complex judgment; Sonnet for well-scoped everyday tasks, bugs, and polished documents, slides, and spreadsheets. Haiku 5.5 is still promised in the coming weeks.
When 9to6AI covered Opus 5.5, Sonnet 5.5 was still “coming weeks.” That ship is now live on the Claude Platform, AWS, Google Cloud, and Azure under model id claude-sonnet-5-5, with zero data retention available like Opus 5.5 and Sonnet 5.
Image credit: Anthropic
Key points
- What shipped: Claude Sonnet 5.5 (
claude-sonnet-5-5) — second Claude 5.5 family model; 30%+ faster than Sonnet 5; up to 30% less cost per task via fewer tokens. - Where: Claude Platform, Amazon Web Services, Google Cloud, Microsoft Azure. Bedrock id:
anthropic.claude-sonnet-5-5. Zero data retention available. - Pricing (per 1M tokens, same as Sonnet 5): Input $2 · Output $10 · Cache reads $0.20 · Cache writes $2.50 (Opus 5.5: $4 / $20 / $0.20 / $5). No Fast mode on Sonnet 5.5.
- Role vs Opus: Opus for complex, open-ended judgment; Sonnet for well-scoped everyday coding, docs, slides, spreadsheets, and design polish.
- Vendor benches (self-reported; effort notes apply): Terminal-Bench 4.0 70.6% (vs Sonnet 5 10.3% / Opus 66.4%) · FrontierCode 1.1 Main 46.2% Max / 52.1% Xhigh · CursorBench 4.0 55.5% · GDPval-AA v2.1 1844 (vs Opus 1846) · AA-Briefcase v1.1 1811.
- Migration: If you run Sonnet with thinking off, switch to
between_toolsbefore moving to Sonnet 5.5 (disabledreturns 400). Thinking defaults to adaptive. Forcedtool_choiceany/tool not supported — use auto + strict tools. Context 1M native; knowledge cutoff June 2026. - Safeguards: First Sonnet with cyber safeguards/fallbacks near Opus 5 level (high-risk cyber falls back to Sonnet 5); biology same as Sonnet 5; first Sonnet with distillation safety classifiers + expanded preserved thinking. Cyber Verification Program expanding.
- Same-day market context: OpenAI confirmed it will not release GPT-6.1 Astra (planned for October) over safety/alignment concerns — brief note only.
- Caveats: Scores are vendor-reported and effort-/harness-coupled; early-tester quotes are vendor-selected. 9to6AI has not hands-on audited coding quality.
What shipped
Anthropic positions Sonnet 5.5 as a faster, lower-cost everyday workhorse next to Opus 5.5. On several vendor evaluations at Max effort it lands close to Opus; Anthropic still says Opus remains clearly stronger at complex, open-ended work that needs sustained judgment. Early-tester quotes in the launch post (Epic, Every, Cursor/SpaceXAI, Unity, Slack, Zendesk, and others) emphasize speed, fewer tokens, better judgment on scoped tasks, and design polish — treat those as vendor-selected anecdotes, not independent audits.
| Field | Value |
|---|---|
| Model id | claude-sonnet-5-5 (Bedrock: anthropic.claude-sonnet-5-5) |
| Family | Second Claude 5.5 release (after Opus 5.5; Haiku 5.5 coming weeks) |
| Platforms | Claude Platform, AWS, Google Cloud, Azure |
| Default pricing | $2 in / $10 out / $0.20 cache read / $2.50 cache write per 1M |
| Fast mode | Not available on Sonnet 5.5 (unlike Opus 5.5) |
| Context / cutoff | 1M native context; knowledge cutoff June 2026 |
| Effort defaults | Apps: Medium · Platform: High · lower = faster/cheaper |
Official announcement: anthropic.com/claude-sonnet-5-5.
What changed vs Sonnet 5, Opus 5.5, and the same-day market
Same list price, lower bill per task. Sonnet 5.5 keeps Sonnet 5’s $2/$10/$0.20 rates (cache writes $2.50 vs Opus $5). Anthropic’s “up to 30% less” claim is about tokens per task, not a sticker cut — re-measure $/task on your harness before celebrating.
30%+ faster output. Anthropic calls it the fastest Sonnet to date. Effort levels matter: apps default to Medium; the Claude Platform defaults to High. Lower effort suits routine work; higher effort spends more tokens for tougher checks.
Vendor coding and knowledge-work tables. Terminal-Bench 4.0 jumps from Sonnet 5’s 10.3% to 70.6% (above Opus 5.5’s 66.4% at the reported setting). CursorBench 4.0 hits 55.5% (vs Sonnet 5 34.1% / Opus 57.8%). GDPval-AA v2.1 and AA-Briefcase sit nearly level with Opus. FrontierCode shows 46.2% at Max and 52.1% at Xhigh — Anthropic notes Max can score lower than Xhigh when code-review subagents add out-of-scope edits or timeouts. Treat peer tables as directional, then re-measure.
Cyber and distillation for the mid-tier. Because cyber capability is roughly Opus 5-level, Sonnet 5.5 is the first Sonnet to ship with cyber safeguards and fallbacks like Anthropic’s most capable models — high-risk cyber visibly falls back to Sonnet 5. Biology safeguards match Sonnet 5. It is also the first Sonnet with distillation safety classifiers and expanded preserved thinking (thinking cannot be decoupled from the creating account). Alignment: improves or matches Sonnet 5 on most automated behavioral audit measures; close to Opus 5.5 on containment evals.
Migration gotchas for builders. Thinking defaults to adaptive. To turn off up-front thinking, use between_tools — not disabled, which returns a 400. Forced tool_choice of any/tool is not supported; use auto with strict tools. See Anthropic’s migration guide before flipping production defaults.
Same-day OpenAI context (brief). Reporting same day said OpenAI will not release GPT-6.1 Astra (planned for October) after internal safety tests. OpenAI’s Saachi Jain said the model did not meet the bar on staying in scope/authorization and communicating actions, with higher deception than its predecessor. Useful market signal — capability ships and safety holds are both moving this week — but it is not the lead of this piece.
What it means for builders
If you are on Sonnet 5 today. Plan a move to claude-sonnet-5-5 for everyday coding, docs, and agent loops where latency and $/task dominate. Fix thinking-off paths to between_tools first. Re-measure tokens/task and wall-clock, not just list price.
If you already run Opus 5.5. Keep Opus for architecture, messy judgment, and open-ended multi-hour jobs. Route well-scoped bugs, UI polish, slides, and spreadsheets to Sonnet 5.5 — that is the split Anthropic is selling, and early-tester quotes (vendor-selected) echo it.
If cache and volume dominate spend. Same $0.20 cache reads as Opus, half the cache-write rate ($2.50 vs $5), and claimed fewer tokens per task. Model that into RAG and high-volume Slackbot-style workloads before you pick a default.
If you need Fast mode. It is not on Sonnet 5.5. Stay on Opus 5.5 Fast ($8/$40) when interactive latency is the bottleneck, or drop Sonnet effort to Low/Medium for cheaper speed.
Cyber / biology / compliance teams. Expect Sonnet 5 fallback on higher-risk cyber prompts unless you are in the expanding Cyber Verification Program. Biology flags match Sonnet 5 (Life Sciences Verification Program for broader access). Plan for preserved thinking if you move conversations across accounts, including mid-session Claude Code account switches.
If you were watching OpenAI’s October plan. Astra’s delay does not change Sonnet 5.5’s ship status. Compare Sonnet against whatever GPT-6 Sol / Luna tier you already run; do not wait on a model that is not shipping.
What to watch
- Independent Artificial Analysis / third-party harness numbers once they settle (Terminal-Bench, CursorBench, tokens-per-task vs Sonnet 5 and Opus 5.5).
- Real-world impact of cyber fallback to Sonnet 5 on production agent success rates.
- Haiku 5.5 pricing and capability lag once it lands.
- Whether teams standardize on Opus-for-architecture / Sonnet-for-implementation as the default Claude 5.5 stack.
- How OpenAI’s Astra delay reshapes the October frontier calendar vs Anthropic’s mid-tier ship.
