Anthropic released Claude Sonnet 5.5, a mid-tier update the company says runs more than 30 percent faster than its predecessor and, despite unchanged per-token pricing, costs up to 30 percent less to complete the same task. The claim rests entirely on Anthropic’s own benchmark testing, published in the same announcement that introduced the model.
Sonnet 5.5 sits below Claude Opus 5.5 in Anthropic’s lineup, a tier meant for well-scoped everyday work: fixing bugs, drafting documents, building slides and spreadsheets. Opus 5.5 remains the pick for open-ended tasks that need sustained judgment, according to Anthropic. A third model, Claude Haiku 5.5, is expected within weeks and is aimed at high-volume, cost-sensitive applications.
The most striking number in Anthropic’s release is a jump on Terminal-Bench 4.0, a test of multi-step command-line tasks: Sonnet 5.5 scored 70.6 percent, up from Sonnet 5’s 10.3 percent. On GDPval-AA, a benchmark spanning 44 occupations, Anthropic says the new model lands within two points of the more expensive Opus 5.5. None of these figures have been verified outside Anthropic’s own testing, and the company did not publish head-to-head results from an independent lab.
The per-token rate has not moved: $2 buys a million input tokens and $10 buys a million output tokens, the same as Sonnet 5. The savings Anthropic advertises come from token efficiency rather than a price cut: the model reportedly batches tool calls more aggressively, so it needs fewer steps to finish comparable work. That is a meaningful distinction for teams billed by usage, since a rate cut and an efficiency gain behave differently as workloads scale.
Anthropic also flagged safety changes tied to the jump in capability. Because Sonnet 5.5’s cybersecurity skills now approach Opus 5’s, it is the first Sonnet model to ship with the cyber safeguards and fallback routing Anthropic built for its top-tier models: routine coding work proceeds normally, while higher-risk security tasks get rerouted to Sonnet 5 instead. The model also adds classifiers meant to block distillation, the practice of extracting a model’s capabilities through mass fake-account queries. Anthropic separately expanded what it calls preserved thinking, a safeguard that keeps a model’s internal reasoning locked to the account that produced it, so that reasoning cannot migrate to a different account.
Two customers quoted in the announcement, Epic Games and Slack, described gains in code review depth and fewer output tokens per task, respectively. Both are companies with existing commercial relationships with Anthropic, which does not make the endorsements false but does mean they are not neutral evaluations.
Anthropic’s release did not disclose adoption figures, pricing changes relative to competitors, or independent benchmark verification, three things a reader would need to judge whether this update changes competitive standings against OpenAI’s GPT-6 Sol or Google’s frontier models. The company’s own comparison table shows GPT-5.6 Sol (used because OpenAI has not published GPT-6 Sol scores for these tests) trailing Sonnet 5.5 on the benchmarks Anthropic chose to run.
For teams already spending meaningfully on Claude API calls, the practical test is a re-run of existing production workloads against Sonnet 5.5’s default effort setting before assuming the advertised 30 percent cost reduction applies to their own token patterns, since Anthropic’s savings figure is a testing-average, not a guarantee.
Reported by Anthropic in its own product announcement, “Introducing Claude Sonnet 5.5,” published on anthropic.com; no dateline was listed on the post.
