Quick answer
Anthropic released Claude Sonnet 5.5 on September 28, 2026, six days after Opus 5.5. Pricing is unchanged at $2 input / $10 output per million tokens, with cache reads at $0.20, batch at $1/$5, and a 1M-token context with 128K output. Model ID: claude-sonnet-5-5. Anthropic says it is more than 30% faster than Sonnet 5 and uses up to 30% fewer tokens per task. On Anthropic's own benchmarks it scores 70.6% on Terminal-Bench 4.0, against 10.3% for Sonnet 5 and 66.4% for Opus 5.5. Haiku 5.5 is promised "in the coming weeks."
The mid-tier model is the one most products actually run, which makes a Sonnet release more consequential for most developers than an Opus one. This one is unusual: Anthropic is claiming that on agentic coding, the cheaper model now edges the flagship.
What changed
- Speed: more than 30% faster output than Sonnet 5, per Anthropic
- Efficiency: up to 30% fewer tokens per task, which lowers cost even at the same price per token
- Context: 1M input tokens and 128K output, matching Opus 5.5
- Cache: reads at $0.20 per million, with a 512-token minimum for caching
- Availability: Claude Platform, AWS Bedrock, Google Cloud Vertex AI, and Microsoft Azure
The benchmarks, as Anthropic reports them
- Terminal-Bench 4.0: 70.6% (Sonnet 5: 10.3%; Opus 5.5: 66.4%)
- FrontierCode: 46.2%, rising to 52.1% at maximum effort
- CursorBench: 55.5%
- GDPval: 1844
- Humanity's Last Exam: 64.5%
- OSWorld 2.1 (computer use): 80.1%
Every number above is vendor-reported and none has independent confirmation yet. The Terminal-Bench jump from 10.3% to 70.6% within one model generation is large enough that the benchmark itself changed between versions; treat the direction as meaningful and the magnitude as provisional. Anthropic's own framing is that Opus 5.5 remains "clearly stronger at complex open-ended work," so this is a coding and agent story, not a replacement for the flagship.
Effort levels and what they cost
Like Opus 5.5, Sonnet 5.5 exposes an effort setting that controls how much it thinks. Anthropic published the average cost of one FrontierCode attempt at each level: about $12.54 at maximum effort down to about $0.76 at low. That sixteen-fold spread is the practical point. For most tasks, medium or low effort on Sonnet 5.5 will cost a fraction of what Sonnet 5 cost for equivalent results; for hard problems, maximum effort buys benchmark points at a price closer to Opus.
Breaking API changes
- Requests with explicit thinking budgets, sampling parameters (temperature, top-p, top-k), assistant prefill, forced tool choice, or thinking disabled now return 400 errors
- Thinking cannot be turned off; the documented workaround for tool-heavy loops is the between_tools setting
- Code written against Sonnet 5 that uses any of the above needs changes before switching the model ID
Safeguards
Sonnet 5.5 is the first Sonnet with cyber safeguards that fall back to Sonnet 5 for certain requests, and it ships with classifiers intended to detect distillation attempts. Anthropic described both as additions rather than changes in policy.
Early customer claims
Anthropic quoted early testers: Base44 reported tasks completing in 3.6 iterations against 7.7 previously, Slack reported 14% fewer tokens, Box reported 2.4 times faster processing, and CodeRabbit reported improved review quality. These are customer statements published by the vendor, not independent measurements.
If you run Sonnet 5 in production: the price is the same, the speed is better, and the API is stricter. Fix the parameters first, then test effort levels on your own workload before believing any benchmark.
Bottom line
Sonnet 5.5 is the release most developers were waiting for: a faster, more efficient default model at an unchanged price, with a vendor-claimed lead in agentic coding that is worth testing. Read the breaking changes before you flip the model ID.
