Anthropic Introduces Claude Opus 5: Near-Fable 5 Capability at Opus 4.8 Pricing, With No Retention Requirement

Anthropic launched Claude Opus 5 (claude-opus-5) on July 24, 2026. Per the Claude Platform release notes, it supports a 1M token context window (both default and maximum), 128k max output tokens, and thinking on by default, at $5 per million input tokens and $25 per million output tokens, the same as Claude Opus 4.8. It launched on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Anthropic also made it the default model on Claude Max and the strongest model on Claude Pro.

Anthropic positions Opus 5 as coming close to Claude Fable 5's frontier intelligence at half the price. The more practical difference for enterprises is governance: Anthropic states that, consistent with prior Opus models, Opus 5 has no data retention requirements for general access, whereas Fable 5 requires 30-day retention. Its safeguards are similar to Opus 4.8's, with stronger guardrails on a narrow range of cyber tasks; in Claude.ai, Claude Code, and Claude Cowork, flagged requests fall back to Opus 4.8 by default, and automatic fallback can be enabled on the API.

There are breaking API changes. Disabling thinking at effort xhigh or max returns a 400 error (Opus 4.8 allowed it), and effort becomes the primary control, with the full low-to-max ladder. The same release notes removed fast mode for Claude Opus 4.7, so requests with speed "fast" now error out. Opus 5 itself offers fast mode at roughly 2.5x speed and twice the base price.

Leadership claims on Frontier-Bench, ARC-AGI 3, Zapier AutomationBench, and OSWorld 2.0 come from Anthropic's post, some from partners' own evaluations, and have not been independently reproduced. Enterprises should compare cost and quality of Opus 5 against Opus 4.8 and Fable 5 on their own task sets rather than relying on vendor charts.

A practical migration checklist from Opus 4.8 to Opus 5: after switching the model ID to claude-opus-5, find and fix every request that sends thinking disabled at effort xhigh or max; replace thinking-budget tuning with effort levels (low, medium, high, xhigh, max); if you relied on fast mode for Claude Opus 4.7, move to Opus 5 or Opus 4.8, because speed "fast" on 4.7 now returns an error; and decide whether to enable automatic fallback on the API, while separating in billing and logs which responses were actually produced by Opus 4.8 so quality evaluations are not muddied.

Things to watch: whether Opus 5 becomes the default frontier model for most enterprises instead of retention-bound Fable 5, how effort levels translate into real cost curves, and how far the thinking-parameter restrictions ripple into existing code.