What it is
Claude Sonnet 5.5 is Anthropic's mid-priced model, released on September 28, 2026. Anthropic describes it as the best mix of speed and intelligence in its lineup. It reads text and images, writes text, and has a 1M-token context with 128K output.
The API price is $2 per million input tokens and $10 per million output tokens, the same as Sonnet 5. Cache reads were cut to $0.10. Unlike Opus, you can turn off up-front thinking with a setting called between_tools.
It is available through the Claude API, the Claude apps, Amazon Bedrock, Google Cloud and Microsoft Foundry.
Our review
For most production work Sonnet 5.5 makes Opus 5.5 hard to justify. Anthropic's launch table has it at 1844 against Opus's 1846 on GDPval-AA, 80.1% against 81.8% on OSWorld 2.1, and ahead on Terminal-Bench 4.0 at 70.6% to 66.4%, though that Opus figure was run at xhigh effort. Artificial Analysis's independent index puts it second overall at 56 on max effort, two points behind Opus 5.5 and three above GPT-6 Astra.
The catch is that the headline scores come from max effort, and Sonnet at max is not cheap. Artificial Analysis measured $5.46 a task at max against $0.48 at medium, where its score falls to 41. GPT-6.1 Sol scored 48 at medium for $0.21. Both list at $2/$10, yet the bills differ because the models spend tokens differently. Anthropic's API default is high effort, so teams that never touch the setting pay for more thinking than many jobs need.
Unlike Opus, Sonnet can drop up-front thinking with the between_tools setting at high effort or below, which suits chat and support where latency matters. The migration is heavier than Opus's, with five breaking changes. Setting temperature, top_p or top_k at all returns a 400 error, which will break older pipelines that pin temperature to zero.
It is also the first Sonnet with cyber safeguards. Higher-risk security requests visibly fall back to Sonnet 5, and Anthropic warns that some microbiology and virology requests may be flagged in error. A security vendor or a biotech should test those paths before signing, because the fallback model scores roughly 400 points lower on GDPval-AA.
Where it earns its place
A support operation answering a million tickets a month
Draft or send first replies without hiring for peak season
Use between_tools with medium or low effort to keep replies quick. At 3,000 input tokens per ticket (2,000 cached) and 400 output, 1M replies is about $6,200 a month before thinking tokens. Zendesk told Anthropic tickets were processed 20% faster (vendor-published).
A product team building code generation into its own developer tool
Get Opus-grade code edits without Opus pricing
CursorBench 4.0 puts Sonnet 5.5 at 55.5% against Opus 5.5's 57.8%. Route everything to Sonnet and escalate only failed tasks to Opus; Anthropic says Sonnet 5.5 costs up to 30% less per task than Sonnet 5 because it uses fewer tokens.
An engineering manager with a $10,000 monthly Opus 5.5 bill
Halve spend without rewriting the agent
Changing the model ID halves list price, but budget for migration: forced tool use, sampling parameters and the old computer-use tool all fail. Re-run your eval set at medium and high effort, since the API defaults to high and the cost gap between them is large.
Before you commit
- Setup
- Model ID claude-sonnet-5-5 on the Claude API (anthropic.claude-sonnet-5-5 on Bedrock), Google Cloud and Microsoft Foundry. Sonnet 5 code must move thinking-off configs to between_tools and drop forced tool use and sampling parameters.
- Integrations
- Same Messages API, SDKs, Batch API and prompt caching as the rest of the Claude line. Fast mode is not offered; it is Opus-only.
- Data and privacy
- Anthropic says retained API data is not used for training without express permission; standard retention is 30 days, zero data retention is available by agreement, and US-only inference costs 1.1x. Classifiers that block reasoning extraction ship with this model.
- Cost at scale
- 1M support replies a month at 3,000 input tokens (2,000 cached) and 400 output costs about $6,200 before thinking tokens; Batch halves that for offline work. Artificial Analysis measured $0.48 a task at medium effort and $5.46 at max.
- Limits
- 1M context, 128K output (300K on Batch, beta). Rate limits match Opus 5.5: 2M input and 400K output tokens a minute on the Start tier, 10M/2M on Scale. Retirement not before 28 September 2027.
What's good
- Second on Artificial Analysis's index (56), above GPT-6 Astra
- $2/$10 with $0.10 cache reads, half Opus 5.5's rates
- between_tools turns off up-front thinking for low-latency replies
- Output over 30% faster than Sonnet 5, per Anthropic
What's not
- Top scores need max effort, which cost $5.46 a task on Artificial Analysis's run
- Any temperature, top_p or top_k value returns a 400 error
- Five breaking changes from Sonnet 5
- Higher-risk security tasks fall back to Sonnet 5
Best for
Teams running coding agents, support automation or document work at volume who want near-Opus quality at half the token price.
Alternatives
Claude Sonnet 5.5: common questions
- Is Claude Sonnet 5.5 worth it?
- Within two points of Opus 5.5 on Anthropic's knowledge-work test at half the price, which turns Opus into a deliberate upgrade rather than the default. We rate it 4.6 out of 5.
- How much does Claude Sonnet 5.5 cost?
- API: $2 per 1M input tokens, $10 per 1M output tokens; cache reads $0.10; Batch API 50% off.
- What are the downsides of Claude Sonnet 5.5?
- Top scores need max effort, which cost $5.46 a task on Artificial Analysis's run; any temperature, top_p or top_k value returns a 400 error; five breaking changes from Sonnet 5; higher-risk security tasks fall back to Sonnet 5.
- Who is Claude Sonnet 5.5 best for?
- Teams running coding agents, support automation or document work at volume who want near-Opus quality at half the token price.
- How much does Claude Sonnet 5.5 cost for a team or in production?
- 1M support replies a month at 3,000 input tokens (2,000 cached) and 400 output costs about $6,200 before thinking tokens; Batch halves that for offline work. Artificial Analysis measured $0.48 a task at medium effort and $5.46 at max.
- Is Claude Sonnet 5.5 safe for company data?
- Anthropic says retained API data is not used for training without express permission; standard retention is 30 days, zero data retention is available by agreement, and US-only inference costs 1.1x. Classifiers that block reasoning extraction ship with this model.
- What are the best Claude Sonnet 5.5 alternatives?
- The closest alternatives we have reviewed are Claude Opus 5.5, GPT-6.1 Sol, Gemini 3.8 Flash and Claude Haiku 5.5.
Answers checked Oct 22, 2026.
Make Claude Sonnet 5.5? Show this rating on your site
