Anthropic officially released Claude Opus 5 on July 24, 2026, positioning it as a high-capability model priced at $5 per million input tokens and $25 per million output tokens — exactly half the cost of the company's top-tier Claude Fable 5. The model carries a one-million-token context window and up to 128,000 output tokens on the synchronous Messages API, making it one of the more capable mid-tier options in the current AI market.
Pricing Architecture and Market Position
The $5/$25 input/output pricing matches the previous Opus 4.8 rate, meaning developers who were already budgeting for Opus 4.8 can upgrade without touching their cost models. Anthropic has made Opus 5 the default model on Claude Max and the most powerful option available to Claude Pro subscribers. Fable 5 — priced at $10 per million input tokens and $50 per million output tokens — remains the top-of-range product for multi-day autonomous agents and the highest-complexity projects.
Beyond per-token rates, Anthropic continues to offer prompt caching that can cut input costs by up to 90% on reused content, and a Batch API that reduces costs by 50%. For long-running agents that repeatedly pass system prompts, project rules, repository maps, and task histories, these mechanisms can substantially compress the real cost per completed task.
Benchmark Results and Their Limits
On CursorBench 3.2 at maximum effort, Opus 5 scored roughly 0.5% below Fable 5's peak while costing half as much per task. Anthropic also reports that Opus 5 leads Frontier-Bench v0.1 and delivers more than double the performance of Opus 4.8 at lower per-task cost. Those Frontier-Bench runs used a mini-SWE-agent on a GKE backend, averaging results across five attempts per task.
One detail worth scrutinizing: Opus 4.8 was used as a fallback whenever Opus 5's or Fable 5's safety classifiers declined a request. This means some benchmark runs labeled as Opus 5 were actually completed by Opus 4.8. The benchmark still measures a real product system — model plus guardrails — but it is not a clean measurement of the model weights alone. Developers building on the API should account for this when projecting real-world performance.
ARC-AGI 3 results are more striking. Anthropic claims Opus 5 scored three times higher than the second-ranked model on that benchmark, which tests novel problem-solving rather than pattern recall. Independent replication of that figure has not yet been published.
Coding and Agentic Capabilities
The practical coding story is competitive. Early partner Lovable reported a 22% improvement over Opus 4.7 on their hardest agentic coding tasks, with lower variance between runs. Cursor's team described Opus 5 as exhibiting many of the same behaviors as Fable 5 in editor workflows. These are internal customer evaluations rather than independent benchmarks, but they signal real deployment gains rather than synthetic-only improvements.
Anthropica highlights two illustrative agent behaviors. In one case, Opus 5 received a mechanical engineering drawing without direct viewing tools; it wrote its own computer vision pipeline to extract geometry from pixels and reconstructed a FreeCAD 3D model. In another, it traced a bug to its root cause inside an open-source package manager and patched an edge case the community fix had missed. Both examples point to a model that generates its own verification methods rather than just producing a first-pass output.
Variance reduction matters as much as peak scores for teams running agents at scale. A model that sometimes hits high marks but fluctuates sharply forces more human review, more retries, and unpredictable cost curves. The lower variance reported by early partners is, if it holds across broader workloads, arguably more valuable than the headline benchmark gap with Fable 5.
Safety Profile
Anthropica's automated behavioral audits placed Opus 5 as its most constitutionally aligned model to date, reporting
