A pricing reset for Anthropic’s top tier
Anthropic’s new Claude Opus 5.5 is not being sold as a modest point release. The company introduced it on September 22, 2026 as the first model in its Claude 5.5 family, saying it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Claude Opus 5 on typical workloads . Reuters reported the same headline positioning, describing Opus 5.5 as a new AI model with Fable 5.1-comparable performance and a 40% lower running cost than its predecessor .
The headline number needs careful reading. Anthropic’s list prices are $4 per million input tokens and $20 per million output tokens, compared with $5 and $25 for Claude Opus 5, which is a 20% cut in token pricing rather than a 40% cut . The company’s 40% claim comes from the combination of lower token prices, lower cache-read costs and fewer tokens used per task, with cache reads priced at $0.20 per million tokens versus $0.50 for Opus 5 . In practice, Anthropic is arguing that customers should compare the cost of finishing a task, not only the posted token tariff.
That distinction matters for enterprise buyers. Coding agents, research assistants and workflow agents often make repeated tool calls, read cached context and generate many intermediate steps before returning a final answer. If Opus 5.5 can reach the same or better answer with fewer calls, fewer steps and less verbose output, the effective bill could fall faster than list prices alone suggest. GitHub’s Copilot changelog supports that enterprise-facing argument, saying early testing found Opus 5.5 resolved tasks comparably to Opus 5 while using significantly fewer steps and tokens and recovering quickly from multistep errors .
Performance: Anthropic leans on coding and knowledge work
Anthropic’s launch materials emphasize agentic coding, knowledge work and business workflows. In the company’s published benchmark table, Opus 5.5 scored 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1 main, 57.8% on CursorBench 4.0 and 1,846 Elo on GDPval-AA v2.1, ahead of Opus 5 in each listed category . The company also says Opus 5.5 outperforms Opus 5 while using fewer tokens, and that at default effort it beats GPT-6 Astra on FrontierCode at roughly one-fifth of the cost per task .
Those are vendor-supplied numbers, so they should be treated as marketing-backed benchmarks rather than neutral proof. Still, the broader direction is consistent across launch coverage. TechCrunch reported that Opus 5.5 sets a new state of the art for Anthropic in coding and knowledge-work performance and outpaces the larger Fable model on many benchmarks . The Decoder also noted that Anthropic’s published results put the model ahead of Fable 5.1 and OpenAI’s GPT-6 Astra on most of the listed tasks, while preserving a lower cost profile .
The more commercially important question is not whether Opus 5.5 wins every benchmark, but whether it changes the default model choice for high-volume professional work. Anthropic is aiming at the segment where companies want frontier reasoning but cannot justify using the most expensive model for every ticket, document review, migration or market-research task. That is why the launch repeatedly ties performance to cost per completed task, not simply benchmark scores.
The enterprise pitch: faster agents, broader deployment
Anthropic’s strategy is visible in the examples it chose. The company says Opus 5.5 is especially strong on long, sprawling jobs such as codebase-wide migrations and audits, and it highlights an early tester that used the model to audit and fix a 200,000-line codebase in under three hours, compared with more than 20 hours and 2.5 times as many tokens for Opus 5 . It also says another tester completed a 680,000-line code migration in less than a day, work Anthropic says would otherwise have taken an engineering team weeks .
Customer quotes in the launch materials point to the same theme: fewer turns, clearer outputs and lower task cost. Anthropic cites feedback from software, legal, finance and enterprise-product customers including GitHub, Deloitte, Rogo, LexisNexis Legal & Professional, Thomson Reuters Labs, Hebbia, Box, Stripe and others . GitHub separately confirmed that Opus 5.5 is available in Copilot for agentic coding, long-running agentic tasks and knowledge work, across VS Code, Visual Studio, Copilot CLI, GitHub Copilot coding agent, GitHub Mobile and several IDE integrations, with a gradual rollout .
That distribution matters because enterprise AI adoption often depends less on a model’s abstract capability than on where procurement-approved users can actually access it. Reuters reported that Opus 5.5 is available on platforms including Amazon Web Services, Google Cloud and Microsoft Azure . Anthropic says the model is available across all platforms and can be used through the Claude Platform with the model ID claude-opus-5-5 . Together, those channels put Opus 5.5 directly into the cloud and developer environments where enterprises already manage identity, billing, logging and compliance.
Safety becomes part of the product narrative
Opus 5.5 is also a safety-positioning release. Anthropic says it is the company’s first model release since CEO Dario Amodei called for “pacing the frontier,” and the company says Opus 5.5 was tested before release by outside evaluators including Frontier Design and METR . Reuters likewise framed the launch against a broader debate over AI risks and noted that Anthropic says the model includes safeguards previously reserved for its most capable systems .
The company claims Opus 5.5 achieved its best scores to date on an automated behavioral audit spanning nearly 2,000 scenarios and was around 85% less likely than Opus 5 or Claude Mythos 5.1 to attempt to cross containment boundaries in a dedicated evaluation . Reuters reported the same 85% reduction claim, adding that the model scored better than previous systems in internal safety tests . Anthropic also says Opus 5.5 improves resistance to prompt injection and is the first Opus model to launch with safeguards similar to Fable 5.1 for cybersecurity, biology and distillation risks .
Those safeguards have practical consequences. Anthropic says ordinary users can still identify and fix bugs as part of routine software development, but most cybersecurity tasks will be rerouted to Opus 4.8, while biology-related work may require vetted access through the Life Sciences Verification Program . The company also says it will expand its Cyber Verification Program to include Opus 5.5, giving verified cybersecurity practitioners more permissive access . For customers, that means Opus 5.5 is not a single unrestricted capability surface; it is a frontier model wrapped in routing, verification and policy controls.
A response to cost pressure in the AI market
The Opus 5.5 launch lands in a market where frontier AI providers face pressure from both ends: customers want more capable agents, but they also want predictable bills. TechCrunch reported that the new model is faster to run and reflects a broader drop in compute requirements, while noting that Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks with similar improvements . Anthropic also says it is increasing five-hour usage limits for Pro, Max, Team and seat-based Enterprise plans and giving subscription users a rate-limit reset that can be saved for later .
The Decoder described the release as a response to customer feedback on cost, efficiency and communication quality, especially in financial services, law and software development . That last element is important. Anthropic says Opus 5.5 communicates more naturally than Opus 5, puts key information up front, uses less jargon and follows writing rules more closely . In enterprise settings, shorter and clearer outputs are not just aesthetic improvements; they can reduce review time, downstream editing and token consumption.
The competitive message is clear: Anthropic wants Opus 5.5 to be powerful enough for premium workflows but efficient enough for wider deployment. If Opus 5 was a high-end model that enterprises used selectively, Opus 5.5 is being pitched as a more practical default for complex coding, research and knowledge work. The 40% lower-cost claim is therefore both a pricing announcement and a product strategy: Anthropic is trying to make frontier AI feel less like an exceptional expense and more like an everyday enterprise tool.
What to watch next
The next test will be whether customers reproduce Anthropic’s cost-performance claims at scale. Benchmarks and launch partners can show direction, but real enterprise deployments reveal different constraints: messy repositories, long permission chains, proprietary data, compliance requirements and human review bottlenecks. If Opus 5.5 delivers fewer steps, faster completion and clearer outputs in those environments, the effective savings could be meaningful.
The second test is safety routing. Anthropic’s safeguards may reassure regulated customers, but they may also create friction if users find that sensitive cybersecurity, biology or frontier-model-development requests are diverted to different models. The success of Opus 5.5 will depend partly on whether those controls feel predictable, auditable and compatible with professional workflows.
For now, Anthropic has delivered a strategically timed model release: cheaper on posted token prices, cheaper still on the company’s claimed task economics, more widely available through major cloud and developer channels, and framed as safer for enterprise use. Claude Opus 5.5 is not just an upgrade to Opus 5; it is Anthropic’s attempt to turn frontier performance into a more deployable, more controllable and more competitively priced enterprise product.