Anthropic Releases Claude Sonnet 5.5 with Major Coding Gains but High Cost at Max Reasoning

Anthropic launched Claude Sonnet 5.5 on September 28, the second model in the Claude 5.5 family, less than a week after the release of Claude Opus 5.5. The company also stated that Haiku 5.5 will arrive in the coming weeks.

From the published data, Sonnet 5.5's core improvements are concentrated in Coding Agent scenarios. On Terminal-Bench 4.0, it jumped from the previous generation's 10.3% to 70.6%, surpassing Opus 5.5's best published score; on CursorBench, which is closer to everyday development, it scored 55.5%, only about two percentage points behind Opus 5.5's 57.8%.

A mid-tier model approaching or even partially surpassing the flagship in coding Agent capability should have been a clear product narrative. But what was actually discussed repeatedly after launch was not how strong it is, but in what scenarios it should be used—and whether, when users choose it, it is actually the one ultimately processing the request.

Anthropic claims that Sonnet 5.5's output speed is over 30% faster, that it requires fewer tokens to complete most tasks, and therefore cost per task can be reduced by up to 30%. On pricing, it continues Sonnet 5's pricing: $2 per million input tokens, $10 per million output tokens, $0.2 for cache reads, and $2.5 for cache writes. Compared with Opus 5.5, input, output, and cache write prices are all half, while cache read pricing is the same.

But Artificial Analysis's testing produced a different set of numbers: when reasoning effort is set to Max, Sonnet 5.5 generates an average of about 193,000 output tokens per Intelligence Index task, the highest among models the organization has tested; Sonnet 5.5 Max's cost per task reaches $7.60, about 50% higher than Sonnet 5. In exchange, it scores 56 on the Artificial Analysis Intelligence Index, only two points below Opus 5.5 Max.

The two sets of numbers are not contradictory, but they point to different ways of using the model. The official claim of 'up to 30% lower cost per task' targets most tasks and does not mean it is still more economical after reasoning effort is set to Max. For API users billed by token, the choice of effort level directly determines whether Sonnet 5.5 is a 'cheap mid-tier model' or a 'high-consumption model close to flagship pricing.'

There is another anomaly worth noting: Sonnet 5.5 scored 46.2% on FrontierCode 1.1 when using Max settings, actually lower than Xhigh's 52.1%. Anthropic explained in a footnote that Max settings are more likely to trigger code review executed by multiple sub-Agents; in two cases examined by Cognition, timeouts or extra modifications outside the task scope appeared, ultimately resulting in FrontierCode deductions. The mechanism details have not yet been fully explained.