Anthropic released Claude Sonnet 5.5 today, the second model in the 5.5 family after Opus 5.5 launched last week. Three facts define the release:
- Output is generated 30%+ faster than Sonnet 5.
- It costs up to 30% less per task in Anthropic's testing, even though the per token price did not change.
- On several benchmarks it scores close to Opus 5.5, the flagship that costs twice as much per token.
What actually changed
Speed. This is the fastest Sonnet model Anthropic has shipped. For interactive use in the Claude apps and Claude Code, where output latency is the bottleneck, the difference is immediately noticeable.
Cost per task, not per token. The rate card is identical to Sonnet 5: $2 per million input tokens, $10 per million output tokens, $0.20 per million cache reads, and $2.50 per million cache writes. The savings come from efficiency: the model needs fewer tokens and fewer tool calls to complete the same job. In concrete terms, a task that cost $1.00 of API usage on Sonnet 5 should land around $0.70 on Sonnet 5.5 in Anthropic's testing. Enterprise budgets stay comparable while covering more work, which is part of why Anthropic now frames pricing around completed tasks rather than tokens.
Capability. The jump is largest in agentic coding. On Terminal Bench 4.0, Sonnet 5.5 scores 70.6% against Sonnet 5's 10.3% and Opus 5.5's 66.4% under Anthropic's settings. On knowledge work it nearly matches the flagship: GDPval AA gives Sonnet 5.5 1844 against Opus 5.5's 1846, up from 1449 on Sonnet 5.
Sonnet 5.5 vs Opus 5.5
| Metric | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| API price per 1M input / output | $2 / $10 | $4 / $20 |
| Cache writes per 1M | $2.50 | $5 |
| Artificial Analysis Intelligence Index (max effort) | 56 | 58 |
| Terminal Bench 4.0, agentic coding (Anthropic) | 70.6% | 66.4% |
| Terminal Bench 4.0 (Artificial Analysis, independent) | 63.6% | 59.6% |
| CursorBench 4.0, real coding sessions | 55.5% | 57.8% |
| FrontierCode 1.1 (Max effort) | 46.2% | 54.4% |
| GDPval AA v2.1, knowledge work | 1844 | 1846 |
| AA Briefcase v1.1, agentic knowledge work | 1811 | 1822 |
| AutomationBench AA | 71% | 70% |
| Humanity's Last Exam (with tools) | 64.5% | 67.7% |
| OSWorld 2.1, computer use | 80.1% | 81.8% |
| Chartography, chart reading (no tools) | 61.6% | 64.4% |
| Best suited for | Well scoped everyday tasks: bug fixes, documents, decks, spreadsheets, fast iteration | Complex open ended work requiring sustained judgment |
Two notes on the numbers. Artificial Analysis's independent Terminal Bench 4.0 run (63.6%) differs from Anthropic's (70.6%); harnesses differ, and Sonnet 5.5 leads Opus 5.5 in both. On the Terminal Bench Science benchmark for agentic research, Sonnet 5.5 scores 53.3%, behind GPT 6 Astra at 63.3% and Opus 5.5 at 59.0%.
How the new effort setting works
Anthropic added an effort parameter with five levels: Low, Medium, High, XHigh, and Max. It controls how long the model works before answering, trading speed and token cost against quality.
- The default is Medium in the Claude apps and Claude Code, and High on the API platform. The same model behaves differently depending on where you call it.
- At Low or Medium effort, Sonnet 5.5 beats Sonnet 5's best score on several benchmarks at roughly one tenth the cost per task. Run it low for routine work; raise effort only where quality actually changes the outcome.
- At Max effort it can match Opus 5.5 on some evaluations, but there is a known wrinkle: extra sub agent code reviews at Max can trigger timeouts. If long runs stall at Max, drop to XHigh first.
Migration detail: if you currently run Sonnet with thinking off, you must switch to the new between_tools setting before moving to Sonnet 5.5. Anthropic's migration guide covers the change.
Things to know before switching
- Your thinking is now tied to your account. Sonnet 5.5 is the first Sonnet to ship with anti distillation classifiers that block reasoning extraction, plus expanded preserved thinking: the model's thinking cannot be separated from the account that created it. Moving a conversation between accounts mid session, including in Claude Code, will hit friction.
- Cyber safeguards now match the flagship tier. Sonnet 5.5's cyber capabilities are comparable to Opus 5, so it launches with the same style of safeguards. Routine software development is unaffected, but higher risk cybersecurity requests will visibly fall back to Sonnet 5. Cyber defenders can apply to the expanded Cyber Verification Program for tiered access.
- Biology safeguards are unchanged from Sonnet 5. Most research, education, and clinical work is unaffected.
- Visual and long horizon work improved. Sonnet 5.5 is the first Sonnet to beat Pokémon Red working only from screenshots, which is a reasonable proxy for sustained visual reasoning, not just game playing. Chart reading jumped from 15.6% to 61.6% on Chartography.
- Enterprise signal is real. Epic Games ran it across tens of thousands of lines of gameplay code and multi hour tasks with less prescriptive prompting. Slack dropped it into Slackbot evals with no prompt changes and measured better results with about 14% fewer output tokens.
- Availability. Everywhere today: the Claude apps, the API under the model ID
claude-sonnet-5-5, AWS, Google Cloud, and Azure, all with zero data retention. Claude Haiku 5.5, built for high volume work, arrives in the coming weeks.
Who should switch
Use Sonnet 5.5 as the default for well defined everyday work: debugging, documentation, building decks and spreadsheets, iterating on interfaces, and agentic coding where cost per task matters. The benchmarks put it within a few points of Opus 5.5 on those workloads at half the per token price.
Stay on Opus 5.5 for ambiguous, open ended work that requires sustained judgment. Anthropic's own assessment and external testers agree the flagship is still clearly stronger there. The gap narrowed a lot, but it did not close.