Anthropic Releases Claude Opus 5 — Frontier-Class Performance at Half the Price

On 24 July 2026 Anthropic released a new AI model, Claude Opus 5. Two things matter. Performance approaches the flagship (Fable 5) and exceeds it on some benchmarks, and pricing is unchanged from the previous generation (Opus 4.8).

1. Pricing — unchanged

• Claude Opus 5 — $5 per million input tokens / $25 per million output tokens
• Fable 5 (flagship) — $10 input / $50 output

That opens room to run the same work at half the cost. Because the unit price matches Opus 4.8, teams already on Opus effectively keep their costs and gain performance.

2. Benchmarks — a large generational jump

• FrontierBench v0.1 — Opus 5 43.3% / Opus 4.8 18.7% / Fable 5 33.7%
• ARC-AGI-3 (reasoning) — Opus 5 30.16% / Opus 4.8 1.52%
• OSWorld 2.0 (computer use) — Opus 5 70.57% / Opus 4.8 55.7%
• SWE-bench Verified (real code fixes) — 96.0%
• Zapier AutomationBench (work automation) — Opus 5 26.0% / Opus 4.8 17.0%

The biggest gains are in agentic tasks: fixing code, operating a screen, chaining several steps on its own. These figures come from the announcement and press coverage, so it is safer to verify against your own data before relying on them.

3. Specifications

• 1M-token context — long documents or large codebases in a single pass
• 128K max output (300K via the Batches API)
• Knowledge cutoff May 2026 — the most recent of any Claude model
• Model ID claude-opus-5 · default model on Claude Max, top model on Pro

4. How to read it for real work

When performance rises and the unit price holds, automation you postponed for cost reasons starts to clear the break-even line. Three things stand out from our own service operations.

• AI support handling — raise answer quality while holding or lowering cost per inquiry.
• Multilingual content — running 13 languages exposes the gap between word-swap translation and context-aware translation.
• Dev and ops automation — more of the manual work (code fixes, log checks) can be delegated.

That said, swapping the model alone does not improve results. What material you feed it, and how — knowledge and context design — still decides quality.

Summary

Claude Opus 5 unsettles the assumption that better results require a more expensive model. The practical significance is simple: you can now try flagship-class results at half the price.

Note: performance and pricing figures are based on the 24 July 2026 announcement and related coverage.