Claude Opus 5: Pricing, Specs, and What Changed

That last detail is the one that catches people migrating: thinking is now on unless you explicitly turn it off, which is the reverse of how Opus 4.8 behaved. It is a small change with a real bill attached, and it is the first thing to check when you move a workload across.

This guide covers what actually shipped: the pricing, the specs, the breaking changes from Opus 4.8, where Opus 5 sits against Fable 5 and Sonnet 5, and whether migrating is worth doing today.

Claude Opus 5: what actually shipped

Claude Opus 5 is the current model in the Opus line, replacing Claude Opus 4.8, which has moved to Anthropic's legacy list but remains callable. The headline specs, taken from Anthropic's models overview:

Two of those rows deserve a second look. The knowledge cutoff of May 2026 is the most recent of any current Claude model — Fable 5 and Sonnet 5 both sit at January 2026 — so on questions about recent tooling and libraries, Opus 5 is working from fresher ground than the model above it in the hierarchy.

And the max-output figure has an exception worth knowing: on the Message Batches API, Opus 5 supports up to 300k output tokens behind the output-300k-2026-03-24 beta header. If you are generating long documents in bulk, that ceiling is more than double the synchronous limit.

What Claude Opus 5 costs

$5 per million input tokens, $25 per million output, per Anthropic's published pricing. That is the same rate Anthropic has charged for every Opus release since 4.6 — capability has moved four times while the price has not moved once.

Here is where that lands against the rest of the current lineup:

The structural point: Opus 5 runs at half the cost of Fable 5, which remains the top of the range. Anthropic's own guidance is to start with Opus 5 for complex agentic coding and enterprise work, and reach for Fable 5 only when you need the highest available capability. That is a genuine fork, not a marketing ladder — the two models are priced a factor of two apart and aimed at different jobs.

One extra cost line if you use it: fast mode on Opus 5 is billed at $10 / $50 per million tokens, and is a Claude API research preview only — it is not available on Bedrock, Google Cloud, or Microsoft Foundry.

What changed from Claude Opus 4.8

If you are migrating an existing integration, this is the section that matters. Two changes will break a request outright; the rest change behaviour or cost.

Thinking is on by default. On Opus 4.8, omitting the thinking parameter meant no thinking. On Opus 5, the same request thinks. This is not just a behaviour change — max_tokens is a hard cap on thinking plus response text, so a workload that sized max_tokens tightly around its answer can now truncate mid-response. Revisit that ceiling on every route that never set thinking explicitly.

Disabling thinking is capped at high effort. Passing thinking: {"type": "disabled"} alongside an effort of xhigh or max returns a 400. Opus 4.8 accepted that combination, so any route that disables thinking needs auditing before you switch the model string. The check runs per request, so a later call that raises effort while thinking is still off fails even if earlier calls in the same conversation succeeded.

Everything else carries over from the Opus 4.7 and 4.8 request surface unchanged, which is worth stating plainly because it means most migrations are a one-line change:

• thinking: {"type": "enabled", "budget_tokens": N} still returns a 400 — use output_config.effort instead.
• temperature, top_p, and top_k are still rejected. Steer with prompting.
• Assistant-turn prefills still return a 400.
• Raw thinking tokens are never returned; display: "summarized" gets you a summary.

Two smaller wins arrive alongside. The minimum cacheable prompt drops to 512 tokens, halved from Opus 4.8's 1024 — so prompts you had written off as too short to cache now create entries with no code change at all, which is worth re-checking if you run many small repeated requests. And Opus 5 supports the full effort ladder from low through max, defaulting to high on both the Claude API and Claude Code.

The rate-limit change people miss

Claude Opus 5 sits in its own rate-limit bucket. Opus 4.8, 4.7, 4.6, and 4.5 share one combined Opus limit; Opus 5 does not draw from it.

That cuts both ways and neither is intuitive. Moving traffic to Opus 5 does not free headroom on your old Opus pool in a way you can spend elsewhere, and it does not inherit the limits you already had. Check your tier's Opus 5 numbers before shifting production volume, because a migration that looks like a model-string swap can quietly become a capacity problem on the first busy afternoon.

Where Opus 5 sits against Fable 5

The tier above Opus did not go away. Claude Fable 5 became generally available on 9 June 2026 and remains Anthropic's most capable widely released model, built for long-running agents, at $10 / $50 per million tokens. Claude Mythos 5 shares Fable 5's specs and pricing exactly but stays invitation-only through Project Glasswing — our Claude Mythos preview guide covers why it is gated.

So the lineup now has a real Opus 5 and a tier above it, which is the arrangement that made the old "is there an Opus 5?" question so confusing while it lasted. The way to read it: Fable 5 is the capability ceiling, Opus 5 is the price-performance pick for demanding agentic work, and Sonnet 5 is the everyday default. Version numbers still only mean something within a tier — Sonnet 5 is not a bigger model than Opus 5, it is a faster and cheaper one that reached the number 5 on its own track.

How the tiers stack up in practice is laid out in Claude models explained, and the full lineage is in our Claude Opus release history.

Should you migrate from Opus 4.8?

For most workloads, yes — and the case is unusually simple, because the price is identical. You are not trading cost for capability here; Opus 5 is a step up over Opus 4.8 at the same $5 / $25 rate. Anthropic describes the migration as a drop-in upgrade, with the same feature set: 1M context, adaptive thinking, prompt caching, batch processing, the Files API, PDF support, vision, and the full tool surface.

The migration itself, in order:

1. Change the model string to claude-opus-5.
2. Audit routes that disable thinking. If any pair thinking: {"type": "disabled"} with xhigh or max effort, either enable thinking or drop effort to high.
3. Revisit max_tokens on every route that never set thinking — those now think, and the cap covers both.
4. Re-run your effort sweep. Prior-model defaults rarely transfer cleanly, and low and medium are unusually strong on this model.
5. Check your Opus 5 rate limits before moving production volume.

There is one honest caveat. Opus 5 ships with elevated cybersecurity safeguards, and its safety classifiers can decline a request — you get a normal HTTP 200 with stop_reason: "refusal" rather than an error. Benign security and life-sciences work occasionally trips them. Code that reads response.content[0] without checking stop_reason first will break on a refusal, so add that branch. Anthropic's recommended fallback for cyber-category refusals is Opus 4.8, which is part of why the older model remains available rather than retired.

Keeping the model ID in exactly one place — an environment variable or a single constant — is what makes all of this a one-line change rather than a search-and-replace across a codebase. That habit pays off at every release, and it is the reason a migration like this can be done in an afternoon.

If the question underneath all this is simply which model to use, our best Claude model guide works through choosing by the job in front of you rather than by whichever number is highest.