Claude Opus 5 launched on July 24, 2026 at $5 per million input tokens and $25 per million output tokens. That price makes routing discipline more important, not less.
Why flagship is not a routing policy
claude-opus-5 is Anthropic's everyday flagship for coding, agentic tasks, and knowledge work, and it is the default model on Claude Max. Those labels describe its intended breadth. They do not settle every production decision.
Anthropic still recommends the larger Fable 5 for the most advanced, longest-horizon autonomous agents. Opus 5 is therefore a new route, not a universal replacement. A bounded coding task and a long autonomous run should not inherit the same model choice merely because one model now carries the flagship label.
Read the benchmark as evidence, not a verdict
On Frontier-Bench v0.1, Opus 5 scores 43.3%, compared with 33.7% for Fable 5. Anthropic also reports state-of-the-art results on Frontier-Bench and GDPval-AA. That is a meaningful reason to test the model on agentic work.
It is not a production acceptance test. A public benchmark cannot represent your repositories, tools, data boundaries, reviewer standards, or recovery paths. Track model spend across every attempt, then divide it by the outputs your evaluation actually accepts. That cost per accepted outcome is more useful than price per token alone.
Build an evaluation before changing defaults
Start with a small, representative set of real tasks. Keep the prompts, tools, context, and acceptance rules stable across model tiers so the comparison remains interpretable.
- Segment the workload into bounded coding, tool-using agent tasks, knowledge work, and longest-horizon autonomy. Model selection should follow the task rather than the product label.
- Define acceptance before running the test. Use relevant checks such as passing tests, factual support, reviewer approval, and completion of the requested tool actions.
- Record the full outcome cost. Include input and output tokens, retries, review burden, and failure recovery instead of counting only the first model call.
- Exercise the guardrails. Test tool permissions, spending limits, stop conditions, escalation paths, and the agent's behavior when a tool or intermediate step fails.
Run Opus 5 against the model currently handling each segment. Include Fable 5 where the task genuinely requires extended autonomy, rather than forcing it into every comparison.
Let pricing shape the experiment
Standard Opus 5 costs $5 per million input tokens and $25 per million output tokens. That matches Opus 4.8 and is half the price of Fable 5.
For a team already using Opus 4.8, the token price does not create savings by itself; the evaluation must show a better accepted outcome rate, lower review burden, or another useful gain. For a Fable 5 route, the lower headline price is attractive, but only if quality and recovery costs remain acceptable for that workload.
The sensible unit of comparison is a completed, approved task. Cheap attempts that require repeated retries or heavy human repair can erase the apparent advantage.
Give Fast mode its own lane
Fast mode costs $10 per million input tokens and $50 per million output tokens and runs about 2.5x faster. Treat it as a separate operating tier, not a free speed toggle.
Measure whether lower latency changes the value of an interactive workflow enough to justify the higher token price. Batch work may favor standard mode; latency-sensitive work may justify Fast mode. Both conclusions should come from the same acceptance criteria used for the standard route.
Guard the agent, not just the model
A capable, cheaper model can make broader automation tempting. Keep the boundary explicit: grant only the tools required for the task, constrain what those tools may change, preserve an audit trail, and define when the agent must stop or escalate.
Review burden and failure recovery belong in the architecture. If a self-tooling agent can modify code or data, its model choice cannot compensate for missing permissions, rollback paths, or evaluation. Guardrails determine the blast radius when the output is wrong.
Use a policy you can revise
A reasonable starting hypothesis is to evaluate Opus 5 for everyday coding, agentic, and knowledge workloads, while retaining Fable 5 as a candidate for the most advanced, longest-horizon runs. Keep the current route wherever your evidence does not support a change.
Document the task class, selected tier, acceptance threshold, and fallback. Re-run the evaluation when the workflow, tools, or model changes. That turns a launch-day decision into an operating practice.
Which workload would you put in the first Opus 5 routing evaluation, and what exact result would your team count as accepted?
📖 Read the full guide → Claude Opus 5: What Anthropic's New Flagship Means for AI Agents
Top comments (0)