Anthropic launches Opus 5.5 with lower prices and claimed Fable-level performance

Anthropic has released Claude Opus 5.5, a new flagship model that the company says matches or exceeds its more capable Fable 5.1 system on many tasks while costing less to operate. The launch gives developers access to a faster, cheaper top-tier model at a time when AI buyers are weighing capability against inference costs.

Opus 5.5 is available immediately through Anthropic’s services and major cloud platforms. The company also says it has improved the model’s writing style, reducing jargon and placing important information earlier in responses. Those changes target a recurring complaint about formulaic or overly elaborate output from earlier Claude models.

Lower prices, faster output

Anthropic lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. The company also says cache-read costs have fallen by 60 percent. TechCrunch reported the output price as a reduction from $25 to $20 per million tokens and noted that the new model requires less compute to serve.

Anthropic estimates that total operating costs should be about 40 percent lower than Opus 5 when token usage is included. That broader estimate reflects the model’s claimed ability to use fewer tokens, as well as output generated more than 30 percent faster. The distinction matters for production teams: lower list prices do not automatically guarantee lower bills if a model uses substantially more tokens on difficult tasks.

Subscriber five-hour usage limits will rise by 20 percent, according to The Decoder’s account of Anthropic’s announcement. Anthropic says the combination of higher limits and lower usage costs should extend practical capacity by about 25 percent.

The model is available on the Claude Platform with the model ID claude-opus-5-5, as well as through Amazon Web Services, Google Cloud, and Microsoft Azure. That distribution gives enterprise teams more options for integrating the model into existing cloud, compliance, and procurement systems.

Performance claims need context

Anthropic describes Opus 5.5 as its strongest-performing model to date and says it matches Claude Fable 5.1 on most tasks. Its comparisons also place the new model ahead of OpenAI’s GPT-6 Astra on several evaluations, despite a lower listed price. These are vendor-reported claims and should not be treated as independent confirmation of general superiority.

The Decoder reported Anthropic’s coding comparisons in more detail. On FrontierCode, the company says Opus 5.5 beats GPT-6 Astra at roughly 20 percent of the cost per task. On Terminal-Bench 4.0, Anthropic claims comparable performance at 40 percent of Astra’s cost, while on CursorBench it reports an 11-point advantage over GPT-5.6 Sol at one-third of the cost.

There is some external support for the model’s benchmark position. Artificial Analysis, an independent platform cited by The Decoder, gave Opus 5.5 a score of 58 on its Intelligence Index at maximum effort, the highest score recorded on that index at the time of reporting. It also reported leading results on six of ten evaluations, including Humanity’s Last Exam and SciCode.

However, the same analysis identified an efficiency tradeoff. At maximum effort, Opus 5.5 reportedly generated about 119,000 output tokens per task, compared with 73,000 for Opus 5, 78,000 for Fable 5.1, and 27,000 for GPT-6 Astra. Lower token prices can offset that usage in some workloads, but the model is not necessarily cheaper on every individual task.

Communication and safety changes

Anthropic says Opus 5.5 is less likely to use jargon, more likely to lead with the most important information, and better at following writing instructions. Early testers described by The Decoder found the output clearer, but those impressions remain limited and are not a substitute for controlled evaluation across real workflows.

The release also expands the safety controls applied to Anthropic’s highest-capability models. Opus 5.5 receives safeguards comparable to Fable 5.1 for cybersecurity, biology, and frontier AI development. Anthropic says requests that trigger those controls may be routed to another model. Most flagged cybersecurity work is reportedly directed to Opus 4.8, while certain biology and frontier-development requests go to Opus 5.

Verified organizations can apply for biological research access through Anthropic’s Life Sciences Verification Program. The company plans to extend its Cyber Verification Program to Opus 5.5. The model also includes “Preserved Thinking,” which restricts API users from editing prior context to extract reasoning, and watermarking measures described as supporting compliance with the EU AI Act.

Anthropic says Opus 5.5 underwent pre-release evaluation by METR and Frontier Design. The company is also preparing stricter screening for reinforcement-learning environments, stronger monitoring, and automated safety-training scenarios. These plans are important, but they describe ongoing work rather than evidence that all alignment or misuse risks have been solved.

What the launch means for builders

For software teams, the main opportunity is not simply access to a higher benchmark score. A model that combines strong coding and knowledge-work performance with lower published prices could make longer-running AI agents more practical, particularly for code review, debugging, documentation, research, and internal analysis.

The cost picture will depend heavily on effort settings, prompt design, context reuse, and output length. The Artificial Analysis figures suggest that maximum-effort Opus 5.5 can consume far more tokens than competing systems. Teams should therefore test cost per completed workflow rather than compare token prices alone.

For enterprise buyers, availability across the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure may reduce deployment friction. But buyers will still need to validate latency, data-retention options, routing behavior when safeguards activate, and the model’s reliability on their own documents and tools. Anthropic’s claims about clearer writing could matter in long sessions, but should be measured against existing models rather than assumed.

Anthropic’s release also intensifies competition among premium models. The company is positioning Opus 5.5 against Fable 5.1 and GPT-6 Astra while responding to cheaper alternatives, including Chinese models that The Decoder described as offering lower performance at much lower prices. The next phase of competition may be decided as much by cost per successful task and deployment controls as by headline benchmark leadership.

What to watch next

The immediate signal will be whether independent evaluations reproduce Anthropic’s reported gains on coding, agentic work, and business tasks. Actual customer costs will be another important test, especially for maximum-effort workloads that generate large outputs.

Anthropic says Sonnet 5.5 and Haiku 5.5 will arrive in the coming weeks with similar improvements in performance, efficiency, and safety. Their pricing and capabilities will show whether the Opus changes represent a broader family redesign or mainly a flagship upgrade.

Developers should also watch the rollout of the Cyber Verification Program, changes to model routing under safety controls, and further details on Anthropic’s monitoring and anti-distillation measures. Those policies could affect which research and security workflows can run directly on Opus 5.5.

Creati.ai perspective

Opus 5.5 is significant because Anthropic is selling efficiency as part of capability, not as a compromise against it. The company’s reported results suggest that premium AI products are moving toward a more practical metric: the cost and reliability of completing an entire workflow.

The evidence is promising but still heavily shaped by Anthropic’s own benchmarks and testing choices. For builders and buyers, the sensible response is targeted evaluation: measure successful task completion, token consumption, latency, safety routing, and human editing requirements on representative work before treating the claimed 40 percent operating-cost reduction as a production fact.