Harvey Cut Token Spend 26% Overnight. The Feature That Made It Possible Has Nothing to Do With Intelligence.

Harvey, the AI legal platform deployed at more than 40 Am Law 200 firms, ran its standard evaluation suite on Claude Opus 5 the day it launched. The model matched maximum-reasoning outputs while generating 26% fewer tokens on average. Nothing changed in the prompt. Nothing changed in the output quality. The change was a parameter: effort set to medium instead of high.

Anthropic released Claude Opus 5 on July 24 with an "effort dial" — a first-class API parameter that controls how much computation the model expends per inference. This isn't a variation on extended thinking or chain-of-thought toggling. It's the first time a frontier lab has made compute intensity a developer-configurable parameter at the model layer, not at the infrastructure layer. The model handles the allocation. The developer sets the budget.

Zapier's CEO offered the most precise illustration of what changes. Opus 5 completed a complex churn-prevention sequence on AutomationBench end-to-end — a test every previous model had failed. The sequence required sustained multi-step reasoning and tool use across dozens of calls. Previous models didn't fail because they lacked raw reasoning ability; they failed because they couldn't sustain the right level of effort across the right parts of a long task chain. Opus 5 doesn't just reason harder — it reasons appropriately. Calibrated effort across a workflow is a different capability than maximum effort at a single prompt.

The pricing hasn't changed: $5 per million input tokens, $25 per million output tokens — identical to Opus 4.8. But the effective cost per completed workflow dropped for every team that was previously running all agentic tasks at maximum effort by default. If Harvey's 26% reduction generalizes even partially across customer workloads, the implications are meaningful. A legal platform running 100,000 model calls per day in discovery review faces a different build-vs-bill calculation at 74% of the previous token volume than it did at 100%.

The investment implication isn't about the model — it's about the design surface the effort dial creates for application developers. An enterprise workflow tool now has to think about where in its pipeline each model call belongs on the effort spectrum. A credit underwriting agent can run high-effort reasoning on risk analysis at intake, then downshift to low-effort summarization for the memo draft. A compliance monitoring agent can run at medium effort on routine transactions and escalate to high effort only on flagged patterns. Developers who build these effort allocations intelligently will deliver better cost-to-outcome ratios than competitors who don't.

The simultaneous day-one launch across Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry signals that multi-cloud enterprise deployment is now a baseline requirement for Anthropic, not a strategy. That availability pattern — and the effort dial's natural role in differentiating cost-performance across cloud pricing tiers — will pressure every enterprise AI provider to expose similar controls. The next generation of enterprise AI won't be differentiated by who has the smartest model. It will be differentiated by who gives developers the most precise control over what that intelligence costs per outcome.