Anthropic releases Claude Opus 5.5, claims benchmark lead over GPT-6 Astra

By Splitfeed EditorialLast updated Sep 26, 2026, 20:10 UTC

On September 22, 2026, Anthropic released Claude Opus 5.5, making it available on AWS the same day. According to the Artificial Analysis Intelligence Index, the model opened a five-point lead over OpenAI's GPT-6 Astra, which Anthropic is citing as evidence of a benchmark lead on most standard evaluations. The release was accompanied by reported efficiency improvements, with the pro-AI camp noting the model is 30 percent faster and 40 percent cheaper than its predecessor, with cache read costs also reduced. Within minutes of Anthropic's announcement, OpenAI launched GPT-6 Sol and Luna, a timing that underscored how closely the two companies are tracking each other's release schedules.

The benchmark claim is the central point of contention. The Artificial Analysis Intelligence Index score is a real and specific figure, but benchmark leadership in AI is rarely straightforward, since different evaluations measure different things and model developers have incentives to foreground the tests where they perform best. The near-simultaneous OpenAI launch complicates the competitive picture further, because GPT-6 Sol and Luna were not the models being directly compared in Anthropic's claims. Forkast flagged efficiency gains and strategic consolidation as the broader themes of the release, suggesting Anthropic is positioning Opus 5.5 as both a capability and a cost story rather than raw performance alone.

Pro-AI voices reacted with enthusiasm, with accounts like sneptech.bsky.social noting the high score on Andalite-bench and Nadya Voynich highlighting that the model outperforms Fable 5.1 while costing less than its predecessor. Critics pushed back on different grounds. Noah Weinberger reported that Opus 5.5 refused to assist with a Libreboot project for AMD machines on safety grounds, framing that as an example of overcautious content filtering. Ellie Lockhart raised a more pointed methodological concern about how Anthropic reports political bias in the model card, arguing that the headline figure obscures differences visible in the underlying data. Ilja offered a blunter dismissal. Middle-ground observers, including Dev Ops Briefly and Cyborg Tribe, acknowledged the cost and efficiency gains but added qualifications, with Cyborg Tribe noting the claims come from Anthropic itself. The New Stack raised a structural concern worth watching, reporting that safety classifiers in the model can silently reroute requests to older models mid-workflow without the user being notified.

The argument that would settle the benchmark dispute is an independent third-party evaluation that tests Opus 5.5 and GPT-6 Astra on the same tasks under controlled conditions. The safety rerouting behavior flagged by The New Stack is the kind of technical detail that developers and enterprise users are likely to press Anthropic on in the weeks ahead.