Choosing an artificial intelligence model used to look like a catalog decision: reserve the most capable option for hard problems and use a cheaper one for everything else. Anthropic’s Claude Opus 5 blurs that boundary. Its effort control lets users spend more reasoning on demanding tasks or conserve tokens when speed and cost matter more.
The important change is not another model reaching the top of a chart. Intelligence is becoming an operating variable that can be tuned per task. For teams running agents for hours, validating code, or processing long documents, the question shifts from “which model do we adopt?” to “how much cognitive work does this step justify?”.
One model, several operating modes
Anthropic prices Opus 5 at the same $5 per million input tokens and $25 per million output tokens as Opus 4.8. That tariff no longer tells the whole story. Effort level changes performance, consumption, and latency, while Fast mode runs about 2.5 times faster at twice the base price.
This is closer to choosing both an engine and a gear. A routine review may run at low effort. A delicate migration may justify max effort. An urgent repair may be worth the speed premium. The practical benefit is avoiding a policy that sends every request to the most expensive setting for caution, or to the cheapest one for budget reasons.
In Anthropic’s own published evaluations, Opus 5 more than doubled Opus 4.8’s Frontier-Bench v0.1 result at a lower cost per task. At max effort on CursorBench 3.2, it came within 0.5% of Fable 5’s peak score at half the cost per task. These figures are useful for forming hypotheses, but they do not replace local evaluation. They come from the vendor, depend on specific setups, and do not automatically represent an organization’s code, documents, or risks.
Verifying work matters as much as producing it
The release focuses on long, iterative tasks. Anthropic describes a model that persists in finding root causes, builds ways to test its own output, and corrects mistakes before stopping. GitHub, which has added Opus 5 to Copilot, highlights autonomous code changes, regression verification, and tool coordination. It also says rollout is gradual and that administrators must enable the model for Business and Enterprise plans.
This matters because an agent’s cost does not end when code is generated. An apparently cheap result becomes expensive when another person has to discover that the tests missed the defect or that the patch addressed a symptom instead of the cause. More model effort may pay for itself if it reduces rework. If it does not, it is simply extra spend.
The sensible approach is to measure the full cycle: accepted-task rate, regressions, time to review, tool calls, latency, and cost. An effort selector only becomes an advantage when it is tied to those outcomes.
Safety becomes part of routing
Opus 5 does not remove the tension between capability and restriction. Anthropic says it intentionally avoided training the model on cyber tasks, yet broader capability improvements brought it close to Mythos 5 at finding vulnerabilities. It remains far behind at developing exploits. Its classifiers allow source-code vulnerability analysis while blocking categories such as penetration testing, binary scanning, and exploit generation.
That creates a less obvious operational consequence: a flagged request can automatically fall back to Opus 4.8 in Claude, Claude Code, and Claude Cowork. The API now offers optional automatic fallbacks as well. The requested model name, therefore, does not by itself guarantee which model performed the work.
Fallbacks can preserve availability, but they need to appear in telemetry. Evaluations, audits, and regulated environments should record the effective model, effort level, fallback, and version as evidence. Otherwise, two runs with the same input may use different capabilities and safeguards without the team noticing.
The architecture around the model matters more
The launch also adds mid-conversation tool changes without invalidating the context cache. In practice, an agent can gain or lose tool access as work progresses without paying again to process all prior context. Opus 5 is also available through Amazon Bedrock, Anthropic’s API, and the many surfaces supported by Copilot.
Broad availability makes adoption easier, but it also makes one fact harder to ignore: the model is only one component. Effort budgets, tool permissions, caching, fallbacks, observability, and acceptance criteria determine how the system actually behaves.
The most important development is not a scoreboard claiming that one model became smarter. It is intelligence moving from a fixed choice to a managed computing resource, allocated according to the value and risk of each step. Maturity will be less about always selecting the maximum and more about knowing where the maximum actually pays.