Anthropic has released Opus 5.5 only two months after Opus 5, combining a lower serving price with claims of stronger coding and knowledge-work performance. The company presents the update as both a capability release and an example of how it intends to manage increasingly powerful models.
Opus remains the most capable and expensive tier in Anthropic’s three-level Claude lineup, above Sonnet and Haiku. The company says the new version surpasses the larger Fable model on many benchmarks and succeeds at several informal tasks that Fable could not complete. TechCrunch does not provide the underlying scores, so those comparisons should be read as Anthropic’s assessment rather than an independent ranking.
Lower Prices Change the Opus Calculation
Output from Opus 5.5 costs $20 per million tokens, down from $25 for the previous version. TechCrunch reports similar reductions in other pricing measures, while Anthropic says the model is faster and requires less compute to serve.
The update also changes how Opus communicates. Anthropic says it is less likely to rely on jargon and more likely to place important information at the beginning of an answer. Those behavioral adjustments may matter in knowledge work, where a technically capable response can still be difficult to use if its conclusion is buried or unnecessarily complex.
The release arrives just two months after Opus 5 became available on July 24. Anthropic says comparable improvements will reach Sonnet 5.5 and Haiku 5.5 in the coming weeks, extending the pricing and performance strategy across its lineup.
Faster releases and lower inference costs can widen practical use, but the source does not disclose adoption figures, latency measurements or an independent cost-performance comparison. Buyers therefore have a clear list price and a set of vendor claims, not a complete picture of how the model behaves across production workloads.
Capability Claims Bring Stronger Safeguards
Anthropic says Opus 5.5 has biology and cybersecurity capabilities comparable with its Mythos model. For that reason, the company is applying safeguards similar to those used for Fable. The restrictions are intended to limit assistance with tasks such as finding exploits in compiled software or developing recognizable biological weapons.
The company reports that safety training broadly followed the approach used for earlier models. It included alignment testing and pre-release evaluation by outside organizations including METR and Frontier Design. TechCrunch does not detail the evaluations or their results, so the report confirms that external testing occurred without establishing how Opus 5.5 performed in every risk category.
This distinction is important because Anthropic’s performance statements and its safety position come from the same organization. External evaluation can add scrutiny, but the available article does not turn those safeguards into a measurable guarantee against misuse or unexpected behavior.
A Model Release Tests Anthropic’s Pacing Promise
Opus 5.5 is Anthropic’s first model launch since chief executive Dario Amodei endorsed deliberately pacing frontier capability advances so that alignment work could keep up. The rapid interval since Opus 5 makes the new release an early test of what that principle means in practice.
Anthropic says it is preparing more advanced training, security and monitoring systems for future models. It also argues that public policy should play a larger role as AI systems become more capable, because institutions need time to build oversight capacity.
The release does not by itself resolve the tension between commercial competition and deliberate pacing. Opus 5.5 is cheaper, faster and, according to Anthropic, more capable, while its safety program remains evolutionary rather than visibly transformed in the account TechCrunch provides.
For customers, the immediate decision is concrete: whether the reduced price and claimed gains justify moving workloads to the new model. For Anthropic, the longer test is whether frequent improvements can coexist with the slower institutional and technical work its leadership says advanced AI requires.
