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. API model ID: claude-opus-5. Announced price: $5 / $25 per million tokens (input / output), same as Opus 4.8. 1M context window, up to 128k output tokens, thinking on by default.

Available on the Claude API, Amazon Bedrock, Google Cloud (Vertex) and Microsoft Foundry. Fast mode: about 2.5× default speed at 2× base price (same logic as Opus 4.8).

Anthropic's product pitch: 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, 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, the main control is no longer only the model. It is effort: low, medium, high, xhigh, max. max is for capability-critical work. The release notes state effort is the primary steering control.

Useful breaking change: disabling thinking is only allowed at effort high or below. thinking: disabled with xhigh or max returns a 400 (change vs Opus 4.8).

SME practice: one page of model policy. Routine work (lint, short summary, simple ticket) = low or medium effort. Sensitive review, migration, multi-file debug = high. Rarely max, and only with turn caps and an owner. Without that, 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: better self-checks, iteration, multi-step threads. Early-access notes cite debugging, root cause, end-to-end business workflows. Useful. Not a reason to open full-access in prod.

What a stronger model does not change: the approval contract, HITL on auth / payment / PII / push main, measuring tokens + % of accepted deliverables, the fact that an agent transcript is not team docs.

Two platform betas if you build custom agents: mid-conversation tool changes (add / remove tools without busting the prompt cache, beta header mid-conversation-tool-changes-2026-07-01) and automatic fallbacks (route to another model if a safety classifier blocks, instead of hard failing).

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

SME migration checklist (1 page)

1) Inventory calls that hardcode claude-opus-4-8 (API, Bedrock, Vertex, Foundry, Claude Code default). 2) Map effort by task type (4-5 row table). 3) Caps: max out tokens, max tool turns, weekly budget owner. 4) Regression tests on 5-10 critical workflows (same prompts, compare cost and acceptance). 5) Documented safety fallbacks (who gets what if a cyber classifier fires). 6) Cutover date + model ID rollback. 7) One weekly cost alert, not a vanity dashboard.

Sources: Anthropic Claude Opus 5 announcement, platform release notes 24 Jul 2026. Re-read the system card and Opus 5 prompting guide before flipping sensitive agents.

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