Day 0: Claude Opus 5 Is Now Live on RouterHub
RouterHub Team · Updated 2026-07-25
Claude Opus 5 is now available on RouterHub under the model ID anthropic/claude-opus-5. RouterHub published the model during Anthropic’s launch window.
Anthropic launched Claude Opus 5 on July 24, 2026, positioning it as its most capable Opus model to date and reporting improvements across agentic coding, professional knowledge work, computer use, and end-to-end business workflows. Teams can evaluate the model through the same RouterHub access layer they use for other model providers.
Claude Opus 5 on RouterHub at a glance
| Item | Detail |
|---|---|
| Provider | Anthropic |
| RouterHub model ID | anthropic/claude-opus-5 |
| RouterHub availability | Published |
| Provider launch date | July 24, 2026 |
What does Anthropic say Claude Opus 5 is built for?
Anthropic positions Claude Opus 5 for complex, multi-step work, with particular emphasis on agentic coding, professional knowledge work, computer use, and end-to-end business workflows.
Long-horizon coding and engineering
Opus 5 is built for tasks that span planning, implementation, testing, and revision. Anthropic reports examples in which the model traced bugs to their root cause, caught edge cases that a surface-level patch missed, and built its own test harness when a live validation source was unavailable.
That makes repository-level changes, difficult debugging, architecture review, code review, and multi-step engineering tasks useful starting points for evaluation.
Professional knowledge work
The model also targets analytical work that combines numerical reasoning, tables, documents, and domain judgment. Anthropic’s published results and early-access reports highlight financial research, legal review, scientific analysis, and other workflows where the model must remain coherent across many steps rather than produce a quick summary.
Computer use and business workflows
Anthropic reports stronger results on computer-use and end-to-end workflow evaluations. For teams, that makes Opus 5 worth testing for research agents, operations automation, and assistants that act across multiple systems.
Verification and careful iteration
Anthropic also reports that Opus 5 more often verifies its work and iterates before declaring a task complete. Teams should test whether that behavior improves accuracy and reduces review effort on their own workflows.
At the model level, Anthropic lists a 1-million-token context window, up to 128,000 output tokens per request, and adaptive thinking. Platform-specific limits and parameter support can vary, so teams should confirm the current RouterHub route configuration before production rollout.
What do Anthropic’s published benchmarks show?
The following image is Anthropic’s official benchmark overview. These are provider-published results, not independent RouterHub measurements, and each benchmark has its own harness, tools, effort setting, scoring method, and comparison set.
Several results stand out:
- Agentic terminal coding: Anthropic reports 43.3% on Frontier-Bench v0.1 for Opus 5, compared with 21.1% for Opus 4.8 in the overview table.
- Knowledge work: Opus 5 records 1861 on GDPval-AA v2, ahead of the other models shown in the table.
- Novel problem-solving: Opus 5 reaches 30.2% on ARC-AGI-3.
- Computer use: Opus 5 reaches 70.6% on OSWorld 2.0.
- Business workflows: Opus 5 reaches a 26.0% pass rate on AutomationBench.
The chart also shows why no launch benchmark should be treated as a universal ranking: results vary by task, harness, effort setting, and model configuration. Opus 5 is a strong candidate for several demanding workloads, but teams should validate it against their own production tasks.
Performance per task matters more than token price alone
Anthropic positions Opus 5 as approaching Fable 5’s capability across many tasks at half Fable 5’s standard per-token price. Anthropic lists Opus 5 at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8. RouterHub displayed the same input and output rates as of July 25, 2026.
The launch article’s effort curves are particularly relevant for production teams. They show that model quality and cost change with the selected effort level. On CursorBench 3.2, Anthropic reports that Opus 5 at maximum effort comes within 0.5% of Fable 5’s peak score at roughly half the cost per task.
This is a more useful way to evaluate a model than comparing token prices in isolation. A higher-quality run can be cheaper overall if it needs fewer retries, fewer tool calls, less human correction, or fewer model escalations. The reverse can also be true when a workload does not benefit from additional reasoning.
Where does Claude Opus 5 fit in production?
Start with workloads where stronger judgment and sustained execution can create a measurable difference:
- A repository-level engineering task that requires investigation, a multi-file change, tests, and a reviewable explanation.
- A complex analytical deliverable that combines several documents, tables, calculations, and explicit assumptions.
- A tool-driven workflow that must complete a sequence of actions and recover when an intermediate step fails.
- A computer-use task with clear completion criteria and observable failure modes.
- A long-running task where the model must verify its own work before handoff.
Run each task more than once and measure completed-task rate, end-to-end latency, total cost, output variance, refusal behavior, and the amount of human review required. Benchmarks help narrow the shortlist; representative workloads determine whether a model belongs in production.
Why evaluate Claude Opus 5 through RouterHub?
New-model adoption often adds operational work beyond the model itself: a separate provider account, integration path, billing flow, and approval process. RouterHub provides a unified access and routing layer across model providers, allowing teams to add Opus 5 to an existing evaluation workflow without introducing a separate provider integration.
The production question is not whether Opus 5 leads every launch-day benchmark. It is whether the model improves completed-task quality, cost, and review effort on the workloads your team actually runs. Start with a controlled evaluation and measure those outcomes end to end.