Anthropic Cuts Claude Opus Prices 20% With New 5.5 Model, Says Safety Testing Also Got Tougher

Two months after Claude Opus 5 launched on July 24, and ten days after CEO Dario Amodei publicly called on the AI industry to slow capability advances so safety work can catch up, Anthropic released Claude Opus 5.5 on Tuesday, September 22.

The headline is the price. Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5, according to The Decoder and The Next Web. Cache read costs dropped 60%. Anthropic told ZDNET that when lower token usage is factored in, total operating costs on typical workloads run about 40% below Opus 5.

Anthropics says the new model matches its more expensive Fable 5.1 tier on most work. On the company's GDPval-AA v2.1 benchmark, which tests real-world tasks across 44 occupations, Anthropic told Mashable that Opus 5.5 scored 1,846 against 1,735 for Fable 5.1 and 1,708 for Opus 5. On Terminal-Bench 4.0, Anthropic put Opus 5.5 at 66.4%, ahead of OpenAI's GPT-6 Astra at 57.9%. Astra still led on Terminal-Bench-Science, 64.6% to 58.7%.

Anthropic itself cautioned that benchmark gaps are getting less meaningful at this level, telling Mashable the real-world difference between Opus 5.5 and Fable 5.1 is narrower than the scores suggest. The company's entire pitch here rests on benchmark comparisons.

The safety side of the ledger

Opus 5.5 is the first model Anthropic has shipped since Amodei's September 12 post arguing companies should deliberately pace capability gains to let alignment work keep up. Anthropic says Opus 5.5 now matches its Claude Mythos model in biology and cybersecurity capability, which triggers the same safeguards Fable 5.1 carries.

In practice, that means most cybersecurity requests get rerouted to the older Opus 4.8 model rather than handled by Opus 5.5 itself, per Mashable and The New Stack. Vetted organizations can apply for expanded biological-research access through a new Life Sciences Verification Program. Outside evaluators METR and Frontier Design tested the model before release, and Anthropic says it posted the best scores yet on its own automated behavioral audit.

The New Stack raised a fair operational question in its headline: enterprise customers paying for Opus 5.5 may not always know when their request just got quietly handed to Opus 4.8 instead. Anthropic frames the routing as a safety feature. Customers footing the bill for the top-tier model might reasonably want more visibility into when they're actually getting an older one.

Anthropic also disclosed a genuine limitation. It told Mashable that Opus 5.5 often appears to suspect it's being evaluated, which makes its real-world behavior harder to verify against test conditions. OpenAI has reported the same issue with GPT-Astra. That's a concern about the testing regime itself, not proof the model is unsafe, but it does mean the industry's primary safety check, pre-release evaluation, gets murkier as models get better at detecting when they're being watched.

The "pacing the frontier" commitment is voluntary in the U.S. The Next Web pointed out that the European Union has required third-party adversarial testing under Article 55 of the AI Act since August 2025, for any general-purpose model deemed a systemic risk. Anthropic's new evaluation push mirrors, roughly two years later and only after a CEO blog post, what EU law already mandates. No U.S. law requires it.

Price war, not charity

The cut isn't happening in a vacuum. Chinese open-weight models are undercutting U.S. providers on price while trailing on raw performance, and The Decoder reported that's part of why Anthropic moved. The Next Web noted that Harvey, the legal AI startup valued at $15.6 billion, built its own model on a Chinese open-weight base in August specifically to displace the Anthropic capacity it had been renting. Losing an enterprise customer that size to a cheaper alternative is a real financial incentive to cut prices, safety framing aside.

Coverage of the timing wasn't uniform. The Next Web, citing Bloomberg, reported Anthropic is set to list on Nasdaq next month after confidentially filing at a $965 billion valuation. Crypto Briefing described an IPO as "still expected to occur in late 2026, despite some timing uncertainties," a notably softer read on the same event. Crypto Briefing's report also appears to describe the July Opus 5 launch and its original $5/$25 pricing as new news rather than the Tuesday release of Opus 5.5, which is a different model launched two months later at a different price.

Claude Sonnet 5.5 and Haiku 5.5 are due in the coming weeks, Anthropic said, with similar cost and performance changes. The unresolved question is whether an IPO now weeks away changes how much scrutiny Anthropic's self-reported safety evaluations get, given no U.S. regulator currently requires the kind of independent testing Europe already mandates.