Anthropic has released Claude Opus 5, positioning it as an everyday frontier model that approaches Claude Fable 5's performance at roughly half the cost.
Opus 5 costs $5 per million input tokens and $25 per million output tokens—the same pricing as Opus 4.8. It is available across Claude products and the API, becomes the default for Claude Max, and includes a one-million-token context window.
Its defining control is effort. Developers can spend less computation on routine work or raise the setting for difficult coding, research, and long-running agent tasks.
Performance claims
Anthropic reports that Opus 5 more than doubles Opus 4.8's Frontier-Bench result while lowering the cost of a completed task. On CursorBench, maximum effort comes within 0.5 percentage points of Fable 5 while costing about half as much.
Independent ARC Prize testing provides one of the strongest early signals. Opus 5 scored 30.16% on ARC-AGI-3 at high effort—the highest verified result at launch—and solved five environments no previously tested model had completed. It also scored 97.5% on ARC-AGI-1 and 90.4% on the semi-private ARC-AGI-2 evaluation at its highest setting.
Anthropic also highlights stronger long-horizon software work, computer use, scientific reasoning, and self-verification. The model was not optimized to advance cyber capability, where the specialized Claude Mythos remains stronger.
What changes in practice
Opus 5 is designed to hold more of a complex job in one run: understanding a large repository, planning across files, making coordinated edits, checking its work, and preserving constraints over a long context.
Anthropic recommends low or medium effort for everyday tasks and higher settings when a better answer justifies added latency and tokens. A fast mode offers about 2.5 times the speed at twice the standard price.
First reactions: strong, but not universal
Early reactions describe Opus 5 as deliberate and consistent across large tasks. Developers have highlighted root-cause debugging, repository-scale changes, frontend implementation, and its tendency to verify work before declaring it complete.
Independent CodeRabbit testing adds useful balance. At maximum effort, actionable code-review precision improved from a 35.2% baseline to 39.3%, but coverage fell from 61.1% to 55.2%. It produced roughly four times as many minor comments while consuming about 50% more input and 65% more output tokens per review.
That makes Opus 5 look promising as a builder and precision-oriented specialist, but not an automatic replacement for every review setup. More reasoning can improve some failures while creating additional verbosity, latency, or overchecking in others.
Why it matters
The frontier-model contest is moving beyond which system tops the most charts. For teams buying AI work, the useful questions are how much accepted work a model completes per dollar, how reliably it performs across long tasks, and how much human cleanup remains afterward.
Opus 5 makes that economic framing explicit. Anthropic still reserves Fable 5 for the hardest workloads, but now offers a less expensive model that can approach it on many tasks and lets developers tune the reasoning budget.
The early evidence suggests a meaningful efficiency improvement, not a universal capability leap. Teams should evaluate Opus 5 on their own repositories, review standards, latency targets, and acceptance criteria.
The product lesson
Model selection should be based on accepted outcomes, not raw token price or headline benchmark rank. A cheaper run is not cheaper if it creates more rework, and a smarter run is not better if its extra reasoning adds output users do not value.
The durable advantage may be controllability: one model can serve routine production work and difficult frontier tasks at different effort levels, provided the product measures quality, cost, latency, and correction effort together.
Relevant links
- Anthropic: Introducing Claude Opus 5
- Anthropic: Prompting Claude Opus 5
- ARC Prize: Claude Opus 5 results
- CodeRabbit: Opus 5 code-review evaluation
- Axios: Anthropic releases Claude Opus 5
- SunMarc archive: Anthropic Introduces Claude Fable 5 and Mythos
- SunMarc archive: AI's Next Benchmark Is Useful Intelligence per Dollar