Anthropic launches Claude Sonnet 5.5 with output more than 30% faster than Sonnet 5 and lower token use for common work. Released on September 28, the mid-tier model is available through Claude, Anthropic's API and major cloud platforms at the same headline token prices as its predecessor.

 

The cost claim needs a careful distinction: Anthropic did not cut the posted rate. It estimates that typical workloads can cost up to 30% less because Sonnet 5.5 completes tasks with fewer tokens, steps and tool calls. That efficiency will vary by prompt, application and thinking setting.

 

The Claude Sonnet 5.5 release has four practical changes:

  • Output speed rises by more than 30%.
  • Token prices remain $2 input and $10 output per million.
  • Agent, coding and office workflows use fewer steps.
  • Frontier-model safeguards now cover the Sonnet tier.

 

More on This Story

 

Claude Sonnet 5.5 Targets High-Volume AI Work

Sonnet is Anthropic's workhorse model line, positioned between the smaller Haiku family and the more capable Opus and frontier systems. The company describes Sonnet 5.5 as a faster, lower-cost complement to Opus 5.5 for well-scoped coding, agent and knowledge-work tasks that may run repeatedly at production scale.

 

Anthropic says the model can build features, fix bugs, review code, draft documents, analyze information and create structured slides. It also gives API users controls for thinking effort, allowing developers to trade response time and token consumption against deeper reasoning when a task demands it.

 

The release follows Claude Opus 5.5 by six days. That sequence gives customers a choice between higher-end performance and a model optimized for throughput. Anthropic says Sonnet 5.5 sits only two points behind Opus 5.5 on GDPval-AA, an evaluation of real occupational tasks, though benchmark results remain vendor-reported.

 

Sonnet 5.5 Pricing Stays Flat as Token Use Falls

Anthropic's model page lists Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, matching Sonnet 5. Prompt caching can reduce eligible costs by up to 90%, while batch processing offers a 50% discount under the company's published terms.

 

The model identifier is claude-sonnet-5-5. It is available in Anthropic's API and through Amazon Web Services, Google Cloud and Microsoft Foundry. Consumers can use it on Claude's web, iOS and Android applications, making this a general release rather than a limited research preview.

 

For United States-only inference, Anthropic charges 1.1 times the standard input and output rates. That option may matter to regulated organizations with data-location requirements, but it also means the cheapest advertised price is not the price for every deployment configuration.

 

Lower token use can be more important than a headline rate cut. An agent that needs fewer retries, shell commands or tool calls consumes less compute and finishes sooner. The benefit depends on real task completion, however; a fast first answer is not cheaper if an application must repeatedly correct it.

 

Anthropic Reports Faster Coding and Agent Performance

Anthropic reports that Sonnet 5.5 scored 70.6% on Terminal-Bench 4.0, an agentic coding test, and calls it the strongest Sonnet yet on long-horizon work. Those figures come from the company and should be read alongside independent testing, task-specific evaluations and production error rates.

 

Early customer results point in the same direction but are not controlled comparisons. Slack reported about 14% fewer output tokens on its offline assistant evaluations. Zendesk said support tickets were processed 20% faster, while Cursor reported a 55.5% score on its private CursorBench 4.0 coding evaluation.

 

One of the more useful signals is that testers measured completed work rather than response style alone. Unity checked whether changes worked after projects were reopened, and another application-building evaluation tracked failed tool calls and the number of iterations needed to finish. These tests still need outside replication.

 

SiliconANGLE's independent report confirms the pricing, availability and speed claims while noting that the release is aimed at everyday coding, documents, spreadsheets and agent workflows. The publication also reports that Anthropic plans to release Haiku 5.5 in the coming weeks.

 

Cyber Safeguards Move Into the Sonnet Tier

Sonnet 5.5 is built on the same foundation as Opus 5.5 and is the first Sonnet release with safeguards previously used for more capable models. Anthropic says sensitive cybersecurity, biology and other high-risk requests can trigger additional checks or route to models configured for narrower capabilities.

 

The company has published a system card covering safety, security and reliability testing. That documentation matters because higher speed and lower cost encourage more autonomous use, including agents that can operate browsers, terminals and business software. Greater action capacity increases the consequences of incorrect tool use or weak permissions.

 

For enterprises, the immediate question is whether the model's efficiency holds on their own workflows. Teams should compare task completion, correction rate, latency, token use and safety behavior under identical prompts. Sonnet 5.5's launch strengthens Anthropic's production offering, but customer-specific evidence will determine whether the advertised savings reach deployed systems.