Verified July 27, 2026
Claude Opus 5
Anthropic's July 24, 2026 flagship: it leads most of the published agentic benchmarks at $5/$25 per million tokens, half of Fable 5's list price. Here are the specs, the numbers, and the parts that are not settled yet.
The short answer
Opus 5 shipped on July 24, 2026 at $5 per million input tokens and $25 per million output tokens — the same price as Opus 4.8, half of Fable 5. It leads Anthropic's published agentic benchmarks (43.3% on Frontier-Bench v0.1 against Fable 5's 33.7%), carries no data retention requirement for general access, and became the default model on Claude Max at unchanged subscription pricing. One caveat before you migrate: in the days after launch several experienced developers publicly reported the opposite experience and switched back to Fable 5. Those are individual public statements, not controlled tests. All figures verified against Anthropic's published materials as of July 27, 2026.
Specs and pricing
API list prices as published by Anthropic, verified July 27, 2026. The model is three days old at the time of writing — re-check the pricing page before you build a budget on these numbers.
| Claude Opus 5 | |
|---|---|
| Released | 2026-07-24 |
| API model ID | claude-opus-5 |
| Input / MTok | $5.00 |
| Output / MTok | $25.00 |
| Cache read / MTok | $0.50 |
| Cache write / MTok (5 min / 1 hr) | $6.25 |
| Batch (input / output) | $2.50 / $12.50 |
| Fast mode (speed / price) | $10.00 / $50.00 |
| Context window | 1M |
| Max output | 128K |
| Effort levels | low · medium · high · xhigh · max |
| Data retention | None |
Benchmarks
Every figure below comes from Anthropic's own launch materials or from Artificial Analysis. As of July 27, 2026 we have found no independent reproduction of the Opus 5 results — treat them as vendor-reported and directional, not settled.
Frontier-Bench v0.1 — 43.3%
Opus 5 scores 43.3%, against 33.7% for Fable 5 and 18.7% for Opus 4.8. This is the number Anthropic led with. It is an agentic coding benchmark, which makes it the most relevant of the set for most readers — and also the one a vendor has the most incentive to optimise for.
CursorBench 3.2 — within 0.5% of Fable 5
At max effort Opus 5 lands within 0.5% of Fable 5's peak score, at roughly half the cost per task. Fable 5 still holds the higher absolute number. Whether half a percentage point justifies double the price is a question about your error budget, not about the model.
ARC-AGI 3 — about 3x the next-best model
Anthropic reports roughly triple the score of the second-place model. ARC-AGI measures abstract pattern induction rather than software engineering, so a large lead here does not transfer cleanly to your codebase — and we have not seen a third party reproduce it.
Zapier AutomationBench — around 1.5x the pass rate
At the same cost per task, Opus 5's pass rate is about 1.5x that of the next-best model. Of everything Anthropic published, this is the closest proxy for tool-calling reliability in production automation workflows.
OSWorld 2.0 — beats Fable 5 at a third of the cost
Opus 5 exceeds Fable 5's best computer-use result at slightly more than a third of the cost. For screen-driving agents this is the widest reported gap between the two models.
Cybersecurity and biology — still behind Mythos 5
Anthropic states plainly that Opus 5 remains behind Mythos 5 on cybersecurity and biology. If those domains are your workload, neither Opus 5 nor Fable 5 is the ceiling — and Fable 5's safety classifiers route much of that traffic away in any case.
Effort levels and speed
Effort is the main quality-versus-price dial. Low, medium and high are the generally available tiers; xhigh and max are API-only. Anthropic's headline CursorBench and Frontier-Bench results were produced at the top of the ladder, so a default-effort deployment will not reproduce the launch charts. Fast mode is a separate axis: roughly 2.5x the speed at $10/$50 per million tokens — that is, Opus 5 output at Fable 5's list price. Verified July 27, 2026.
How to get it
Available on the API, in the Claude apps and in Claude Code from launch day, July 24, 2026.
API — claude-opus-5
Point your client at the model ID claude-opus-5. One million tokens of context, 128K maximum output. Two related changes shipped alongside it: automatic fallback routing on the API, and prompt caches that survive a tool-set change mid-conversation, where previously swapping tools invalidated the cache.
Claude Max — now the default
Opus 5 became the default model on Claude Max on launch day, with no change to subscription pricing. If you are on Max and have not touched the model picker since July 24, you have been using Opus 5 without selecting it.
Claude Pro — the strongest model on the plan
Opus 5 is the most capable model available on Claude Pro, again at unchanged pricing. Context worth knowing: Fable 5 was removed from Pro and Team Standard on July 20, 2026, replaced by a one-time $100 credit.
Claude Code
Select the model in-session or set the model ID in your configuration. Because effort is exposed here, this is the cheapest place to run your own A/B against Fable 5 before you change anything in production.
Decide with your own numbers
The published benchmarks and the first week of field reports point in different directions, which means neither one is a substitute for testing on your own tasks. Compare the two models side by side, then price your real monthly volume.