Anthropic rolled out Claude Opus 5 on July 24, expanding the context window to one million tokens while keeping Opus 4.8’s price. Developers building automated workflows should care: the update adds effort-level controls, a mandatory “thinking” stage, prompt caching for short prompts, and the ability to swap tools mid-conversation—features that could reshape integration patterns and cost structures.

Why the upgrade matters

Opus 5 claims intelligence comparable to the “Fable 5” benchmark at half the cost. In practice, the model handles inputs up to a million tokens and spits out up to 128 k tokens without raising fees.

New developer-centric knobs

Effort-level controls

The model defaults to a “high” effort setting but offers five tiers: low, medium, high, xhigh and max. Each tier trades compute time and cost for answer quality. Developers can start a job at high effort, drop to a lower tier if the task proves cheap, or bump to xhigh/max when the problem is tough.

Mandatory thinking mode

Claude now runs a “thinking” phase before emitting any output. The phase is mandatory for the two highest effort tiers (xhigh, max) and cannot be turned off.

Prompt caching for 512-token prompts

Prompt caching now works on prompts as small as 512 tokens.

Mid-conversation tool switching (beta)

A beta feature lets an ongoing Claude session change the external tool it calls—say, swapping a web-search module for a database query—without restarting the conversation.

Stakes for integration workflows

The added controls give engineers finer granularity over cost and performance, but they also introduce new failure modes. Mandatory thinking at xhigh/max means any existing code that disables internal reasoning will now hit errors. Mid-conversation tool switching is still in beta, so stability isn’t guaranteed.

Counter-point

Anthropic kept pricing flat, but the higher effort tiers consume more compute.

What to watch next

  • Updates to the “fast mode” preview, which promises lower latency for research workloads but is not yet generally available.

For developers whose pipelines run Claude unattended, Opus 5 offers a toolbox that can squeeze more intelligence out of each dollar—provided they adjust their code to accommodate thinking phases and test the new caching and tool-switching pathways.