An agent is only as good as its worst hour. Plenty of models can take the first three steps of a task correctly; the ones worth building on are those that are still coherent at step ninety. Claude Opus 5 is built for that shape of work, with one caveat that will show up on your invoice if you ignore it.
This page is written for engineers building agents and automated workflows. If you want the model overview first, start with the complete Claude Opus 5 guide.
| Best effort level | xhigh |
| Context available | 1M tokens — enough for real working material, not just snippets |
| Cost | $5 per million input tokens, $25 per million output |
Why Claude Opus 5 suits autonomous agents
- Long-horizon coherence: it sustains complex work over long autonomous runs rather than losing the plot.
- It coordinates teams of subagents effectively, with few cases of agents overwriting each other's work.
- It verifies its own output before handing off, so writer-verifier patterns work well.
- Task budgets let you give it a token ceiling it is aware of, so it paces itself and finishes gracefully instead of being cut off.
A workflow that works
- Specify the whole goal in the first turn. Long-horizon quality depends heavily on a complete opening brief rather than incremental clarification.
- Define what "done" means in checkable terms. An agent without a success criterion runs until it runs out of budget.
- Set a task budget so the model paces itself against a ceiling it can see.
- Cap subagent fan-out explicitly — this is the single biggest cost lever.
- Stream and log. A single turn can run for many minutes; you want visibility, not a spinner.
Settings to start from
Set the effort level to xhigh. Long-horizon agentic work is the canonical case for the top of the ladder. Counter-intuitively it often costs less overall, because higher effort means fewer wandering turns.
Leave adaptive thinking on — it is the default on Claude Opus 5, and turning it off introduces failure modes that are not worth the saving. See effort levels explained if you want to tune this properly.
What to watch out for
- It delegates more readily than Opus 4.8 did. If your prompt library still says "delegate more" — a fix written for the previous model — remove it and add a ceiling instead.
- Scope expansion compounds in agents. One extra self-assigned step becomes ten by the end of a run.
- Priority Tier does not cover Opus 5, and it draws on a separate rate-limit bucket from the Opus 4.x models. Check your limits before shifting production volume.
The verdict
The strongest agent foundation in the Opus line — with an explicit delegation cap and a task budget as non-optional configuration.
If cost is the constraint, read how to cut Claude Opus 5 costs before downgrading — routing and caching usually beat switching models.
More on Claude Opus 5
Start with the complete Claude Opus 5 guide for the overview, or go deeper:
Ready to go deeper?
Read the full Claude Opus 5 guideFrequently Asked Questions
How long can a Claude Opus 5 agent run unattended?
A single request on a hard task can run for many minutes, and agent loops can run far longer. Plan for it in your architecture — streaming, generous timeouts, and progress visibility — rather than assuming a request completes in seconds.
How do I stop an agent from spawning too many subagents?
Say so explicitly in the system prompt: a hard ceiling on parallel agents, plus guidance to delegate only for genuinely independent, sizeable tracks of work. Opus 5 reaches for delegation more readily than Opus 4.8, so guidance written for the older model now points in the wrong direction.
What is a task budget?
A token ceiling for a whole agentic loop that the model can actually see, so it paces itself and wraps up gracefully rather than being cut off mid-task. It is different from a max-token limit, which the model is unaware of and which simply truncates. It is currently a beta feature with a 20,000-token minimum.
