Upgrade

Claude Opus 5.5: Anthropic Launches New Top Model Despite Calling for AI Slowdown

co-written by newsrooms.ai22. September 2026, 19:49

Anthropic has just released Claude Opus 5.5, the first model in its new Claude 5.5 family. According to independent benchmarking firm Artificial Analysis, the model takes the top spot on its Intelligence Index while also becoming significantly cheaper than its predecessor. The timing stands out: only days ago, Anthropic CEO Dario Amodei published a widely discussed essay arguing that the entire industry should slow the pace of AI development.

Top Spot on the Intelligence Index

At its maximum effort setting, Opus 5.5 scores 58 on the Artificial Analysis Intelligence Index. According to the testers, this is the highest score they have measured so far, ahead of the next model by several points. Opus 5.5 leads six of the ten evaluations that make up the index, including Humanity’s Last Exam (61.4 percent, up from the previous best of 59.1 percent by Claude Fable 5.1) and SciCode (66.9 percent versus 63.1 percent).

On the coding benchmark Terminal-Bench 4.0, Artificial Analysis measured Opus 5.5 at 59.6 percent. That puts Anthropic level with OpenAI’s GPT-6 Astra and marks an 11-point gain over the previous model, Opus 5. In its own announcement, Anthropic reports 66.4 percent on the same test, though with a different setup and effort level. Independent measurements therefore come out somewhat more modest than the company’s figures.

The model is particularly strong at so-called agentic knowledge work, tasks in which the AI independently produces professional outputs such as analyses or presentations. On Artificial Analysis’s private AA-Briefcase evaluation, Opus 5.5 reaches an Elo rating of 1,822, which is 143 points ahead of Claude Fable 5.1. It is also the first time an Anthropic model has beaten OpenAI’s GPT-5.6 Sol on presentation quality. On GDPval-AA v2.1, which covers real-world work across 44 occupations, Opus 5.5 scores 1,846 Elo.

Cheaper per Token, but More Talkative

Anthropic has cut prices substantially. One million input tokens now cost $4 (about €3.50), and one million output tokens cost $20 (about €17.40), a 20 percent reduction compared to Opus 5. The discount on cache reads is even steeper. Cache reads account for most of the cost in coding and agent workloads, and they now cost $0.20 (about €0.17) per million tokens instead of $0.50, a 60 percent drop. Anthropic says typical workloads will become 40 percent cheaper overall.

Artificial Analysis puts that claim into perspective. At maximum effort, Opus 5.5 uses around 119,000 output tokens per Intelligence Index task, 1.6 times as many as Opus 5 (around 73,000). GPT-6 Astra gets by with roughly 27,000. As a result, Opus 5.5 ends up at about the same cost per task as Opus 5. On the positive side, the testers note that four of its five effort levels (medium, high, xhigh and max) sit on the Pareto frontier: each is either cheaper or more capable than every other model scoring above 50 on the index.

The context window remains at one million tokens, with support for both text and image input.

Where Opus 5.5 Falls Short

Despite its top ranking, the model does not lead everywhere. According to Artificial Analysis, it trails competitors on CritPt (physics research tasks), AA-LCR (long-context reasoning) and GDP.pdf. Anthropic’s own table also shows GPT-6 Astra ahead on Terminal-Bench-Science, 64.6 to 58.7 percent, and narrowly ahead on the workflow benchmark AutomationBench.

Anthropic itself concedes that benchmark margins at this level have become a less reliable guide. In its own use, the company says, the gap to Fable 5.1 is narrower than the numbers suggest.

Some users will also notice new restrictions. Because Opus 5.5 is roughly as capable in cybersecurity and biology as Claude Mythos 5.1, a model Anthropic keeps under tight access, many security tasks are automatically routed to the older Opus 4.8 and sensitive biology requests to Opus 5. Anyone who wants the full model for this work has to apply to a verification program. In addition, thinking mode can no longer be switched off, and newly created API accounts can no longer edit the model’s prior reasoning due to a safeguard against distillation.

One open question concerns safety testing itself. Anthropic says Opus 5.5 outperforms every previous model on its most comprehensive internal alignment audit and attempted to circumvent sandbox boundaries around 85 percent less often than Opus 5. At the same time, the company writes that there are signs the model often recognizes when it is being tested, which makes it harder to predict how it will behave in real-world use.

A Fast Launch After a Call to Slow Down

This is where criticism of the timing comes in. In his essay “We Must Pace the Frontier,” published only days ago, Amodei argued that the AI industry should slow down so that safety practices can keep up with model capabilities. According to TechCrunch, his three-step plan calls for external evaluators embedded inside AI companies, shared safety standards and “limits on the rate of unchecked AI progress” brokered by the U.S. government, and international coordination that would include China. Anthropic is implementing the first step on its own.

Now comes Opus 5.5, a model that takes the top spot on the index and lowers prices. Anthropic addresses the apparent contradiction directly. It calls Opus 5.5 “our first release since we called for pacing the frontier” and says the model was tested before release by external evaluators including METR and Frontier Design. The slowdown, the company argues, is aimed mainly at future models, such as systems that could fully automate AI research itself. By Anthropic’s own assessment, today’s safeguards would not be sufficient for those systems, and public policy would need to play a larger role.

Whether that distinction is convincing remains a matter of debate. Critics, including a commentary in Forbes, argue that Amodei’s proposal does not go far enough in the first place. Supporters see the pacing concept as an operating model for responsible releases. What is clear is that the race with OpenAI remains tight, and Anthropic intends to keep pace.

Availability

Claude Opus 5.5 is available now in the Claude apps, on the Claude Platform, and through Amazon Web Services, Google Cloud and Microsoft Azure. Anthropic is also raising usage limits for subscribers on its Pro, Max, Team and Enterprise plans. A fast mode offering up to 2.5 times the speed costs $8 and $40 (about €7 and €35) per million input and output tokens respectively. The smaller models Sonnet 5.5 and Haiku 5.5 are expected to follow in the coming weeks.