Claude Opus 5: Near-Fable Intelligence at Half the Price

Introduction

Anthropic released Claude Opus 5 on July 24, 2026, positioning it as "a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price." It ships at $5 per million input tokens and $25 per million output tokens — the same list price as Opus 4.8 — and was available on all platforms on launch day, including the Claude API, Claude apps, Claude Code, and Claude Cowork.

The API model ID is claude-opus-5. Anthropic's documentation now names it the recommended starting point for complex agentic coding and enterprise work, the role Opus 4.8 held since May.

What Anthropic Published

The launch post frames most of its results as comparisons rather than absolute scores:

• Frontier-Bench v0.1 — surpasses all other models and "more than doubles Opus 4.8's performance at a lower cost per task."
• CursorBench 3.2 (max effort) — within 0.5% of Fable 5's peak score at half the cost.
• ARC-AGI 3 — a score "three times as high as the next-best model."
• OSWorld 2.0 — surpasses Fable 5's best result "at just over a third of the cost."
• Zapier AutomationBench — pass rate around 1.5× the next-best model for the same cost per task.
• Scientific work — 10.2 percentage points higher than Opus 4.8 on organic chemistry tasks and 7.7 points higher on protein tasks.

Anthropic also claims Opus 5 uses 26% fewer tokens on average than Opus 4.8, which matters more than the sticker price on agentic workloads where output volume dominates the bill.

Note what these are and are not. Anthropic chose the benchmarks, the harnesses and the effort levels, and the launch post publishes no per-benchmark score table as text. Several of the most striking figures in the announcement — Box's 8% improvement, Goldman Sachs' trading benchmark using "a seventh of the reasoning tokens," Perplexity's "nearly double Opus 4.8" on first-turn redlines — come from customers' internal evaluations, not from Anthropic's own measurements. The system card, published the same day, carries the full safety and capability evaluation.

What Changes for Developers

Opus 5 keeps the Opus 4.7/4.8 request surface — no temperature, top_p or top_k, no budget_tokens, no last-assistant-turn prefills — but two defaults changed, and one of them is a hard error:

Adaptive thinking is now on by default. On Opus 4.8, a request that omitted the thinking field ran without thinking. On Opus 5 the same request runs with adaptive thinking. Since max_tokens caps total output — thinking plus visible text — workloads that ran thinking-off by omission need their max_tokens revisited.

Disabling thinking is capped at high effort. thinking: {"type": "disabled"} combined with an effort level of xhigh or max returns a 400 error. Either re-enable thinking or step effort down to high or below.

Three smaller changes are worth picking up:

• The minimum cacheable prompt drops from 1,024 tokens to 512, so shorter prefixes can now create prompt cache entries with no code change.
• Mid-conversation tool changes (beta, header mid-conversation-tool-changes-2026-07-01) let you add or remove tools between turns without invalidating cache hits on earlier turns.
• Automatic fallbacks (beta, header server-side-fallback-2026-07-01) accept fallbacks: "default", which picks a recommended fallback model based on the refusal category rather than making you name one.

Two capabilities present on Opus 4.8 are missing: Priority Tier is not supported on Opus 5, and the server-side web fetch tool is not available. Teams with a Priority Tier commitment need to plan capacity separately.

The specifications otherwise carry over: a 1M-token context window at standard pricing, 128K max output per request (300K on the Message Batches API with the output-300k-2026-03-24 beta header), and effort defaulting to high on the Claude API and Claude Code. The knowledge cutoff moves forward to May 2026, four months later than Opus 4.8's January 2026.

Behavior and Prompt Re-Tuning

Anthropic's migration guidance is unusually specific about what to remove from existing prompts:

• Delete carried-over verification instructions. Opus 5 verifies its own work without being told to, and leaving "double-check your work" scaffolding in place causes over-verification.
• Prompt explicitly for length. Default visible responses and written deliverables run longer than on Opus 4.8, and lowering effort reduces thinking volume without reliably shortening the visible answer.
• Constrain scope and subagents. For narrow tasks, say so explicitly; in multi-agent frameworks, give guidance on when delegation is warranted or cap the number of subagents.
• Re-run your effort sweep. Anthropic says low and medium are stronger on Opus 5 than on earlier Opus models and are worth testing as cost controls, while max is worth testing where capability matters more than token spend.

Safety Posture

Anthropic describes Opus 5 as its "most aligned model to date," reporting the lowest rates of deceptive behavior it has measured and a score of 2.3 on overall misaligned behavior in its automated behavioral audit. It also reports that safety classifiers intervene 85% less often than on Fable 5 — and that biology-related requests blocked on Fable 5 now route to Opus 5 rather than Opus 4.8.

The cybersecurity safeguards are drawn narrowly rather than broadly: vulnerability finding in source code is allowed, while binary-based scanning, penetration testing and exploit generation are blocked. Enterprises whose legitimate work falls inside those blocks can apply to the Cyber Verification Program. Anthropic is explicit that Opus 5 remains behind Claude Mythos 5 on both offensive cybersecurity and biology research.

Conclusion

Opus 5's pitch is not a new capability ceiling — Fable 5 still holds that — but a better price-performance point at the tier most production workloads actually run on. The 26% token reduction and the near-Fable results at half Fable's price are the claims worth validating first on your own evaluation set, because they are the ones that change a budget.

The migration itself is small: swap the model ID, audit any request that disables thinking, and re-check max_tokens on workloads that were implicitly running thinking-off. The larger job is prompt hygiene — the instructions you added to make Opus 4.8 verify itself and stay brief now work against you.

Sources

• Introducing Claude Opus 5 — Anthropic's launch announcement, July 24, 2026
• Claude Opus 5 System Card — safety and capability evaluation
• Models overview — specifications and pricing for all current Claude models
• Migrating to Claude Opus 5 — breaking changes and the migration checklist