Claude Opus 5: effort, long agents, what SMEs should change

Claude Opus 5 in facts (24 Jul 2026)

Anthropic shipped Claude Opus 5 on 24 July 2026, under the API ID claude-opus-5. The announced price stays identical to Opus 4.8: $5 input and $25 output per million tokens. The model offers a 1M-token context window and up to 128k output tokens, with thinking on by default.

The model is available on the Claude API, Amazon Bedrock, Google Cloud (Vertex) and Microsoft Foundry. Fast mode reaches about 2.5 times the default speed, at twice the base price, on the same logic as Opus 4.8.

Anthropic's product pitch fits in one sentence: an everyday Opus that approaches Fable 5 on coding and knowledge work, at about half the cost per task in their internal evals (Frontier-Bench and CursorBench per their post). We do not treat their leaderboards as absolute truth; we look at what you must freeze as a team.

Effort ladder: the real cost / quality lever

On Opus 5, model choice steps back behind a new dial: effort, with its five levels low, medium, high, xhigh and max. The max level stays reserved for capability-critical work. The release notes state it clearly: effort becomes the primary control for steering the model.

One breaking change deserves a note: disabling thinking is only allowed at effort high or below. A call with thinking: disabled and an xhigh or max effort returns a 400 error, which changes from Opus 4.8.

In SME practice, one page of model policy is enough:

Without that page, you pay the ceiling by default on the whole backlog.

Long agents: what changes, what does not

Anthropic positions Opus 5 for long-running agents: the model checks its own work better, iterates and holds a multi-step thread. Early-access feedback cites debugging, root cause analysis and end-to-end business workflows. That is useful, but it does not justify opening full access in prod.

A stronger model changes nothing about the approval contract, the HITL on auth, payment, PII and pushes to main, nor about measuring tokens and the percentage of accepted deliverables. An agent transcript still does not make team documentation.

Two platform betas deserve a follow-up if you build custom agents. Mid-conversation tool changes let you add or remove tools without busting the prompt cache, behind the beta header mid-conversation-tool-changes-2026-07-01. Automatic fallbacks route to another model when a safety classifier blocks, instead of failing hard.

If your issue is rather "Claude Code sessions rot", the evergreen Claude Code guide stays the reference. This news post dates the Opus 5 delta and the effort policy.

SME migration checklist (1 page)

The migration fits on one page:

1. Inventory the calls that hardcode claude-opus-4-8 (API, Bedrock, Vertex, Foundry, Claude Code default)
2. Map effort by task type, in a 4-5 row table
3. Set caps: output tokens, tool turns, weekly budget with an owner
4. Run regression tests on 5-10 critical workflows, same prompts, comparing cost and acceptance
5. Document safety fallbacks: who gets what if a cyber classifier fires
6. Set a cutover date and a rollback model ID
7. Keep one weekly cost alert, not a vanity dashboard

The sources are the Anthropic Claude Opus 5 announcement and the platform release notes of 24 July 2026. Re-read the system card and the Opus 5 prompting guide before flipping sensitive agents.

If you want to freeze model and effort policy on a real process, not a demo, we can scope in 20-40 minutes.

Frequently asked questions