# Claude Sonnet 5.5: Pricing, API Specs, and Scores

URL: https://3ammarketer.com/news/claude-sonnet-5-5-release-specs-pricing-benchmarks
Published: 2026-10-06

> Anthropic released Claude Sonnet 5.5 on September 28, 2026, offering 1M context tokens at $2 input and $10 output per million tokens.

Anthropic released Claude Sonnet 5.5 on September 28, 2026, pricing base API tokens at $2 per million input and $10 per million output. The release marks the second model in the Claude 5.5 family, operating more than 30 percent faster than Claude Sonnet 5. Due to reduced token overhead per task, Sonnet 5.5 costs up to 30% less for standard operational workloads compared to its predecessor, as detailed in our analysis of [Claude tokens pricing and cost guide](/blog/claude-tokens-pricing-and-cost-guide). [1][2]

## Claude Sonnet 5.5 specifications, model IDs, and support timelines

Claude Sonnet 5.5 features a context window of 1M tokens and a maximum output limit of 128K tokens. Developers accessing the Claude API use the identifier claude-sonnet-5-5, while Amazon Bedrock deployments run under anthropic.claude-sonnet-5-5. [3]

![Anthropic Claude platform documentation comparing specifications, pricing, and model IDs across Claude models.](https://dygolixznokmekyvajbb.supabase.co/storage/v1/object/public/images/content/8f9e2cc8243ec9033a88c765.png)

The model has a reliable knowledge cutoff and training data cutoff of June 2026. Under Anthropic's formal retirement commitments, the model will not be retired sooner than September 28, 2027. [3]

## $2 input and $10 output: API pricing, prompt caching, and plan tiers

Base developer token costs mirror previous generation pricing: $2 per million input tokens and $10 per million output tokens. For prompt caching, 5-minute cache writes cost $2.50 per million tokens, 1-hour cache writes cost $4 per million tokens, and cache reads run at $0.20 per million tokens. Default cache entries hold a 5-minute lifetime, and cache read tokens apply a standard 0.1x multiplier against base input prices. Uncached input tokens and cache creation tokens count toward ITPM rate limits, while cache read tokens do not count toward ITPM for most models. [1][4]

Platform seat access captured across subscriptions lists Claude Free at $0 monthly, Claude Pro at $20 billed monthly ($17 billed annually), and Claude Max at $100 billed monthly. Team accounts price Standard seats at $25 billed monthly ($20 annually) and Premium seats at $125 billed monthly ($100 annually). [6]

## Benchmark evaluation: where Sonnet 5.5 rivals Opus 5.5 and where it trails

On agentic coding evaluations, Sonnet 5.5 reached 70.6% on Terminal-Bench 4.0, up from 10.3% recorded by Sonnet 5. [1]

Terminal-Bench 4.0 agentic coding evaluation scores across Claude models, September 2026. [1]

- Sonnet 5: 10.3 %
- Opus 5.5: 66.4 %
- Sonnet 5.5: 70.6 %

CursorBench 4.0 yielded a 55.5% score, while FrontierCode 1.1 (Main) registered 46.2% at Max effort and 52.1% at Xhigh effort. On knowledge benchmarks, Sonnet 5.5 achieved 1844 on GDPval-AA v2.1, trailing the 1846 score of Opus 5.5 by two points, while recording 1811 on AA-Briefcase v1.1. [1]

Across multi-discipline reasoning, Sonnet 5.5 scored 64.5% with tools on Humanity's Last Exam, 80.1% partial on OSWorld 2.1 computer use, 61.6% without tools on Chartography, and 56.0 on the AA Intelligence Index. It also became the first Sonnet variant to beat Pokémon Red using screenshots exclusively. Despite these gains, teams comparing architectures via [Claude Opus 5.5 vs Sonnet 5.5](/blog/claude-opus-5-5-vs-sonnet-5-5) will find Sonnet 5.5 trails Opus 5.5 on SWE-Bench Pro, scoring 81.3% compared to 89.9%. [1][7]

## Adaptive thinking and agentic tooling performance

Sonnet 5.5 implements adaptive thinking configured with a default effort level of high. API implementations must update legacy logic: Claude 4.7 and later models reject manual extended thinking requests using budget_tokens, returning an explicit 400 error. [3]

In tester workflows, the model batches tool calls together more aggressively than Sonnet 5, completing actions in fewer steps. Across application builds evaluated at Base44, Sonnet 5.5 averaged 3.6 iterations per build compared to 7.7 iterations required by Opus 5. In an internal enterprise test, Sonnet 5.5 synthesized earnings materials and call transcripts into a 10-slide operating review judged ready to send without edits by two experts. [1]

## Sources

1. [Introducing Claude Sonnet 5.5 \ Anthropic](https://www.anthropic.com/claude-sonnet-5-5)
2. [LLM API pricing timeline (3AM Marketer data set)](https://3ammarketer.com/data/model-pricing)
3. [Models overview - Claude Platform Docs](https://docs.anthropic.com/en/docs/about-claude/models)
4. [Prompt caching - Claude Platform Docs](https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching)
5. [Rate limits - Claude Platform Docs](https://docs.anthropic.com/en/api/rate-limits)
6. [AI tool pricing tracker (3AM Marketer data set)](https://3ammarketer.com/data/ai-tool-pricing)
7. [Sonnet 5.5 Specs, Benchmarks & Token Pricing: Anthropic | ARMES Docs](https://armes.ai/docs/models/anthropic/claude-sonnet-5.5)
8. [Claude Sonnet 5.5 Review (October 2026) | Features, Pricing & Guide](https://aitoolsreview.co.uk/tool/claude-sonnet-5-5)
9. [Extended thinking - Claude Platform Docs](https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking)
10. [⚡ Anthropic Launches Claude Sonnet 5.5](https://newsletter.evolvingai.io/p/anthropic-launches-claude-sonnet-5-5)