Anthropic released Claude Opus 5.5 on Tuesday, dropping output token pricing to $20 per million from $25 while claiming state-of-the-art results in coding and knowledge work. The release lands two months after Opus 5 shipped on July 24, an unusually tight cadence for the top tier of Anthropic's lineup. Opus sits above Sonnet and Haiku as the most capable and most expensive Claude tier, and the price cut narrows the gap with mid-market rivals without moving Opus off the frontier bench.
The headline claim is that Opus 5.5 outperforms Anthropic's larger Fable model on many benchmarks and cleared several informal tasks Fable could not. That is the more interesting fact than the price drop: a smaller, cheaper model beating a bigger one at Anthropic's own evaluations suggests the company is squeezing meaningful gains out of training and inference optimization rather than raw parameter count. Opus 5.5 also runs faster, reflecting a broader drop in the compute required to serve it.
Pricing across other metrics fell in similar proportion to the output-token cut, though Anthropic did not itemize each line. The company is holding the Opus brand at the top of the stack while making it materially cheaper to call at scale — a posture that reads as a direct response to the aggressive price-per-token positioning coming from OpenAI's GPT-6 Sol and Luna releases earlier this month.
Key facts
- 01Anthropic priced Opus 5.5 output tokens at $20 per million, down from $25 for the prior Opus model.
- 02The new model beats the larger Fable on several benchmarks and completed informal tasks Fable failed.
- 03Opus 5.5 ships two months after Opus 5 landed on July 24.
- 04Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks with similar improvements.
- 05Biology and cybersecurity capabilities match Mythos, triggering the same safeguards Anthropic applies to Fable.
Anthropic also retrained Opus 5.5 on how it communicates. The model is less likely to lean on jargon and more likely to front-load important information at the start of a response. Small on paper, meaningful for anyone piping Opus into an agent loop or a customer-facing product where the first sentence of the output determines whether a downstream step fires correctly.
Sonnet 5.5 and Haiku 5.5 are slated for the coming weeks with similar performance improvements, according to Anthropic's announcement. That would refresh the full three-tier lineup within a single quarter, a pace Anthropic has not previously sustained across all three tiers simultaneously.
The release is the first from Anthropic since CEO Dario Amodei publicly endorsed pacing the frontier — deliberately slowing capability progress to give alignment work time to catch up. Amodei wrote earlier this month that he had come to believe more prudence was necessary, not just more investment in safety research.
Opus 5.5's safety training was broadly similar to its predecessors, with pre-release evaluation from outside organizations including METR and Frontier Design. Anthropic classified the model as comparable to Mythos in biology and cybersecurity capability, which means it ships under the same safeguards applied to Fable — limits on generating exploits for compiled programs and on producing recognizable biological weapons, among other categories. The company said more advanced training and evaluation systems, including improved security and monitoring, are being prepared for future models.
The tension inside this release is worth naming. Amodei is publicly arguing for slower capability progress, and the same company just shipped a faster, cheaper, more capable Opus two months after the last one. Anthropic's framing is that pacing the frontier means investing in the alignment and policy infrastructure to keep up, not freezing model releases — but the two-month cadence will invite scrutiny from the same safety community Amodei has been courting.
Anthropic is running the pricing playbook the rest of the market ran a year ago: keep the flagship name, cut the token cost, ship faster. With Opus 5.5 undercutting its own predecessor by 20% on output while claiming benchmark wins over a larger internal model, the read is that Anthropic thinks it can compete on cost-per-intelligence without ceding the top of the stack. If Sonnet 5.5 and Haiku 5.5 land on schedule with proportional gains, the pressure moves back onto OpenAI and Google to answer with either sharper pricing or a clear capability lead — and the frontier-pacing rhetoric will get tested against what shows up in the API next quarter.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
