CLAUDE OPUS 5
Released July 24, 2026. Approaches Fable 5 performance at half the price — same $5/$25 per million tokens as Opus 4.8. 1M context, 128K output, thinking on by default with five effort levels.
Claude Opus 5 is the flagship Opus of the Claude 5 generation, released July 24, 2026, and it supersedes Opus 4.8 as the recommended Opus-class default. Anthropic's positioning is unusually direct: performance approaching Claude Fable 5 — the $10/$50 top of the range — at exactly half the price, because Opus 5 keeps the Opus line's $5 per million input tokens and $25 per million output, unchanged from Opus 4.8. It ships a 1M-token context window (both default and maximum), a 128K-token output limit matching Fable 5's, vision input, and native web search. The thinking model changed with this generation: thinking is on by default (adaptive), tunable across five effort levels from low to max, and — a breaking API change from Opus 4.8 — disabling thinking is only allowed at effort high or below; requests that disable it at xhigh or max return a 400 error. Prompt cache reads cost $0.50 per million tokens and the Batch API halves list price to $2.50/$12.50. The case for Opus 5: it is the easy upgrade — same price as 4.8, a step-change in capability. The case against: the very hardest reasoning, where Fable 5 still leads, and routine agentic volume, where Sonnet 5 ($3/$15 standard) stays cheaper.
vs Opus 4.8: the easy upgrade. Identical $5/$25 pricing, same 1M context, and a generational step in capability — there is no cost reason to stay on 4.8, only the API-behavior review below. vs Fable 5: the value question. Fable 5 ($10/$50) keeps the crown on the hardest reasoning and the longest agent horizons, but Opus 5 approaches its performance at half the price — for most workloads that used to justify Fable, Opus 5 is now the rational default and Fable the escalation. vs Sonnet 5: the budget question. Sonnet 5 ($3/$15 standard) remains the pick for routine agentic work and high-volume coding; Opus 5 earns its premium when patch quality, hard reasoning, or output length (128K vs Sonnet's tier) decide the outcome.
Opus 4.8 was adaptive-thinking-only with no user knobs. Opus 5 keeps thinking on by default but adds five effort levels (low, medium, high, xhigh, max) — and changes the rules: disabling thinking is only permitted at effort high or below. A request that turns thinking off at xhigh or max effort returns a 400 error. If you're migrating integrations from Opus 4.8, audit any code path that disables thinking before swapping the model ID. On Council AI this is handled for you — the platform's thinking-mode toggle maps to valid effort/thinking combinations automatically.
Claude Opus 5 is available on every paid Council AI plan (Plus, Pro, Ultra) as a premium-tier model, and it slots in as the new default Anthropic anchor seat — the quality of Fable 5 councils at closer to Opus-council spend. One prompt fans out to Opus 5 and models from up to 8 other labs in parallel, and the moderator synthesizes a consensus answer with an agreement score. A high-signal frontier council after this release is Opus 5 + GPT-5.6 Sol + Gemini 3.5 Flash: three labs, three independently trained perspectives, with Opus 5 covering the deep-reasoning seat that previously cost Fable 5 prices.
$5.00 per million input tokens and $25.00 per million output tokens — identical to Opus 4.8. Prompt cache reads are $0.50/1M, and the Batch API runs at $2.50/$12.50. Council AI bundles it into monthly plan budgets.
1,000,000 tokens — both the default and the maximum — with a 128,000-token output limit, matching Fable 5's output budget.
Yes for capability — same price, step-change improvement, and it supersedes 4.8 as the recommended Opus-class default. One caution for API users: the thinking/effort contract changed (disabling thinking above effort high returns a 400 error), so review integrations before switching model IDs.
Anthropic positions Opus 5 as approaching Fable 5's performance at half the price ($5/$25 vs $10/$50). Fable 5 still leads on the hardest reasoning and longest agent horizons; Opus 5 is the rational default for most formerly-Fable workloads.
Five levels from low to max, with thinking on by default (adaptive). Thinking can only be disabled at effort high or below — disabling it at xhigh or max effort is rejected with a 400 error, a breaking change from Opus 4.8's adaptive-only behavior.
Yes — on all paid plans (Plus, Pro, Ultra) as a premium-tier model, running in parallel councils with models from 8 other labs and consensus scoring on every answer.