Anthropic Releases Opus 5.5, Price Cuts and Enhanced Safeguards

Key Takeaways

1Opus 5.5 is launched with security restrictions equivalent to Fable and an external evaluation by METR and Frontier Design

2The model is offered at $20 per million output tokens, down from $25 previously, and runs faster

3Anthropic claims superior performance to Fable on many benchmarks and comparability to Mythos in biology and cybersecurity

4Sonnet 5.5 and Haiku 5.5 are announced for the coming weeks with similar improvements

Why it matters : Opus 5.5 marks a technical advancement while strengthening security measures, in the context of increased scrutiny on AI development.

Anthropic launches Opus 5.5 with security restrictions aligned with Fable and an external evaluation by METR and Frontier Design. The model is touted as faster and cheaper than its predecessor, priced at $20 per million output tokens. The company claims superior performance to Fable across many tests and has refined its communication. Sonnet 5.5 and Haiku 5.5 versions are expected to follow in the coming weeks.

Equivalent Controls to Fable and a Rise in Regulation

The release of Opus 5.5 comes with the same safeguards applied to the Fable model. These measures notably limit the system's use for seeking vulnerabilities in compiled programs or developing recognizable biological weapons, and cover other sensitive tasks. Security training has remained close to that of previous models, with alignment tests and a pre-launch evaluation conducted by external organizations, including METR and Frontier Design. Anthropic is also preparing more advanced training and evaluation mechanisms for its future models, including enhanced security and monitoring systems. The company argues that public policy will need to weigh more heavily as systems gain capability, indicating that this competence is built over time and asserting that it has begun to establish the corresponding infrastructure. It plans to detail this work soon. This launch follows Dario Amodei's decision to deliberately slow the increase in capabilities to keep pace with advancements in alignment, a move he justified this month by calling for more caution and an adjustment in the pace of progress as well as investment in risk prevention.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Claimed Capabilities and Comparisons with Fable and Mythos

Anthropic presents Opus 5.5 as establishing a new state of the art for coding and knowledge work, claiming it surpasses Fable on many benchmarks while successfully handling informal tasks that the latter struggled with. The company states it is comparable to Mythos in biology and cybersecurity. In terms of usability, Opus 5.5 is described as less jargon-heavy and more inclined to prioritize key messages.

Reduced Pricing and Faster Execution

Opus 5.5 is offered at $20 per million output tokens, down from $25 for the previous generation, and other pricing indicators follow the same trend. The model executes faster, reflecting a general decrease in the computational resources required for its operation.

Tight Schedule and Upcoming Wave of the Claude Range

The launch took place on a Tuesday, two months after the release of Opus 5 on July 24, confirming a close cadence. Within the Claude range, Opus remains the most capable and expensive level, Sonnet is the intermediate, and Haiku is the fastest and most economical. Anthropic announces that Sonnet 5.5 and Haiku 5.5 will arrive in the coming weeks, with performance improvements of a similar order.