Anthropic ships Sonnet 5.5 at Sonnet 5's price but 30% cheaper per job, with watermarking baked in

The sticker price did not move. Sonnet 5.5 costs the same as Sonnet 5: $2 per million input tokens, $10 per million output tokens, and $0.20 per million cache reads ^[4]. The saving comes from the model using far fewer tokens to do the same work, which Anthropic says makes it about 30% cheaper than the previous generation ^[5]. That distinction matters on Monday. Your rate card is unchanged, so a like-for-like invoice comparison will not show the saving. It shows up only in the token counts on the jobs you actually run, which means someone has to measure Sonnet 5 and Sonnet 5.5 on the same workload before the 30% is real for you.

The speed claim is separate and simpler. Anthropic says Sonnet 5.5 is more than 30% faster in output speed than the previous generation, making it the fastest Sonnet to date ^[3]. One early tester put a number on what that does to a real workflow. "We fed Claude Sonnet 5.5 hundreds of real support use cases across replies and escalation requests," said Zendesk Director of AI Abhinay Kathuria. "It made fewer wrong decisions and resolved tickets faster than the Claude models we use in production today. Tickets were processed 20% faster, getting our customers the help they need without the wait." ^[6]

The change that a compliance officer will care about is the watermarking. Every text Sonnet 5.5 writes now carries invisible watermarking ^[7]. Anthropic said it does not affect text quality or readability, but increases the chance that AI-generated text can be detected, and that it added the capability to comply with global regulations including the EU AI Act ^[8]. For a team that has spent a year telling customers its Claude-drafted copy is indistinguishable, that is a design goal reversing. The output is now built to be identifiable as machine-written, and there is no toggle described to turn it off.

On benchmarks, Anthropic reported Sonnet 5.5 scored 70.6% on Terminal-Bench 4.0, an agentic coding test, against the previous model's 10.3% ^[9]. It scored two points below the flagship Opus 5.5 on GDPval-AA, a test of real-world occupational work ^[10]. TechCrunch notes Sonnet's appeal over Opus is that it can spawn multiple agents without hitting cost limits ^[11], which is why a cheaper-per-job mid-tier matters more than the leaderboard gap suggests.

Sonnet 5.5 is also the first Sonnet to launch with the cyber safeguards applied to more capable models, and Anthropic says those may trigger on prompts touching cybersecurity, biology, or other sensitive content, while most software and life-sciences work goes unaffected ^[12]. TechCrunch reports the company claims Sonnet 5.5 has cyber capabilities comparable to Opus 5 ^[13]. If your product legitimately handles security tooling, that is a new refusal surface to test against your own prompts.

Sonnet 5 was announced about three months ago, with efficient agentic deployment as its pitch ^[14]. Anthropic plans to release a new Haiku, its smallest and most cost-sensitive model, in the coming weeks, without a firm date ^[15]. The forcing function for a team on Claude today is narrow: run your top three workloads on both versions, compare token counts and error rates, and decide whether the watermark is a feature you can sell or a fact you now have to disclose.