Anthropic has announced the release of Claude Opus 5.5, the first model in its 5.5 series. The company states that Opus 5.5 matches the performance of Fable 5.1 on most tasks but at a usage cost 40 percent lower than the previous version. It is described as an improved model for long tasks and knowledge work, and it positions itself at the top of the Opus range, opening the 5.5 generation. The model is equipped with the capabilities of Fable 5.1, its most advanced model to date, on most tasks. The cost for a million tokens is now 4 dollars for input and 20 dollars for output, compared to 5 dollars and 25 dollars previously. With a more efficient token consumption, Anthropic estimates a 40 percent saving on common usage. This reduction, however, does not come at the expense of performance: on Terminal-Bench 4.0, Opus 5.5 achieves 66.4 percent, compared to 55.8 percent for Fable 5.1. The model leads on most benchmarks published by the company, which however warns that at this level of capability, these tests do not accurately reflect the real differences between models. In its blog post, Anthropic also highlights work on how the model expresses itself, one of the most frequent criticisms addressed to Opus 5. Two settings have changed: the default effort level is now set to "medium," and the reflection mode can no longer be disabled. The model is accessible on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure, under the identifier claude-opus-5-5. For Claude subscribers, the usage limits on five hours for the Pro, Max, Team, and Enterprise plans per seat are increased, with an offer to reset them that they can trigger whenever they want. Safety measures modeled on those of Fable 5.1. With Opus 5.5, Anthropic extends to its Opus range the protections that previously surrounded its most powerful models. These include cybersecurity and biology, two areas where the model reaches a level comparable to Claude Mythos 5.1. In practice, cybersecurity requests largely shift to Opus 4.8, and biology work shifts to Opus 5. The same system was already in place for Fable 5.1, whose cybersecurity safeguards had been relaxed at the beginning of September to reduce false positives. On its automated behavioral audit, which covers nearly 2,000 scenarios, Opus 5.5 achieves the best results ever measured by the company. A dedicated evaluation on crossing containment limits reports about 85 percent fewer attempts compared to Opus 5 and Mythos 5.1, all of low severity. A blind spot remains: the model often suspects it is being evaluated, which complicates the interpretation of its real behavior. It is also the first model in the Opus range to include content labeling required by the AI Act. A release that follows closely on the heels of Anthropic's call to slow down. Anthropic presents Opus 5.5 as its first launch since the experiment in which Dario Amodei called on labs to collectively moderate the pace of their progress, published in mid-September. Before its release, the model was tested by external organizations, including METR and Frontier Design. Meanwhile, OpenAI revealed the same day GPT-6 Sol and GPT-6 Luna, two lightweight versions of GPT-6 Astra, priced at 2 dollars and 10 dollars per million tokens for Sol, and 0.10 dollars and 0.50 dollars for Luna. Note that Claude Sonnet 5.5 and Claude Haiku 5.5 are expected in the coming weeks, before a possible initial public offering by the company, which would be postponed to November.