Sonnet 5.5 release — Frontier Now

Anthropic released its mid-tier Claude Sonnet 5.5 on September 28, positioned for everyday coding and office work. The company says it is 30% faster than Sonnet 5 (released about three months ago) with a significantly slower token burn, and that its own benchmarks show Sonnet 5.5 beating Opus 5.5 on agentic coding — likely, Anthropic says, because it can run multiple agents in parallel without blowing cost limits. A new Haiku arrives in the coming weeks.

The facts

Pricing and availability: API name claude-sonnet-5-5 at $2/M input, $10/M output, $2.50 cache writes and $0.20 cache reads — level with Sonnet 5 and half of Opus 5.5 ($4/$20/$5) except cache reads; live on AWS, Google Cloud and Azure.
Performance: 30% faster than the previous Sonnet 5 with a significantly slower token burn; Anthropic says per-task cost drops by up to 30% in most jobs. Terminal-Bench 4.0 scores 70.6% (Sonnet 5: 10.3%), and FrontierCode at High effort scores 10 points above Sonnet 5 at about one-fifteenth the per-task cost.
Agentic coding: beats Opus 5.5 on Anthropic’s own benchmarks, attributed to parallel multi-agent runs finally being cost-feasible; for the competitive frame, see GPT-6 Sol and Luna.
Cyber safeguards: cyber capabilities comparable to Opus 5; the first Sonnet under the same Fable/Opus-tier cyber safeguards.
Sourcing note: all claims are Anthropic’s own; pricing verified by 机器之心 against the official page.

Our take

A mid-tier model taking over agentic coding is this month’s most important cost curve: Anthropic’s own explanation — parallel agents are finally affordable — declares the split-long-tasks-and-parallelize playbook a default rather than a luxury. Teams running coding agents or batch evals should benchmark Sonnet 5.5 against Opus 5.5 on their own workloads now. The price list has landed: the sticker matches Sonnet 5, and the savings come from fewer tokens per task — redo your math on measured token counts, not the per-token price alone.