Anthropic released the first model of the Claude 5.5 series, Opus 5.5, at 12:00 AM on September 23rd. The official said that the performance of most tasks is comparable to Fable 5.1. Under default settings, the cost of typical tasks is reduced by 40% compared to the previous generation Opus 5, and the output speed is improved by more than 30%. It has been launched on Claude as well as Amazon Web Services, Google Cloud, and Microsoft Azure.
In the 9 tests officially announced, Opus 5.5 leads Fable 5.1, Opus 5, and OpenAI's GPT-6 Astra, GPT-5.6 Sol in 7 tests. In three programming tests, it achieved the highest scores: Terminal-Bench 4.0 reached 66.4%, FrontierCode v1.1 was 54.4%, and CursorBench 4.0 was 57.8%. However, it lags behind GPT-6 Astra in business process and research agent tests. On the Artificial Analysis Intelligence ranking list, it ranked first with 58 points, leading Fable 5.1 and GPT-6 Astra with 53 points each.
Long task performance was a major focus of the upgrade: in internal testing, Opus 5.5 completed the rewrite of the load balancing software HAProxy from C language to Rust in 9.5 hours, with a cost 51% lower than Fable 5.1; in a financial report research test requiring checking numbers and citations one by one, it passed 16 out of 18 reports, while Fable 5.1 and Opus 5 did not pass; an early tester completed a migration involving 680,000 lines of code in less than a day.
In terms of pricing, the cost for one million input and output tokens is $4 and $20 respectively, a 20% reduction from Opus 5, and cache read costs are reduced by 60%; the five-hour usage quota for Pro, Max, and other packages is also increased, and subscribed users get an additional opportunity to reset the quota.
In terms of security, Opus 5.5 attempts to bypass isolation boundaries about 85% less frequently than Opus 5 and reports such behaviors on its own. Its success rate in defending against prompt injection is among the lowest for tested models, matching Fable 5.1. However, most network security tasks are still handed over to Opus 4.8. The official admitted that pre-deployment assessments cannot identify all failure scenarios.
Less than two hours after the release, OpenAI launched GPT-6 Sol and Luna; Sonnet 5.5 and Haiku 5.5 are planned to be launched in the coming weeks. Trading efficiency and cost compression for the large-scale deployment of agents is becoming the main trend in competition among top models.
