Claude Opus 5 transforms reasoning into compute budget: near-Fable intelligence at half price, but max effort is not for every task
Anthropic releases Claude Opus 5 with 5 effort levels, 1M context window, and 128k output tokens. Test-time compute analysis, task economics, and migration guide.
- 5 Levels
- compute effort levels
- 1M Tokens
- context window capacity
- $5 / $25
- input/output price per 1M tokens
Claude Opus 5 launched with a straightforward premise: significantly higher performance at the same pricing of $5 per million input tokens and $25 per million output tokens. Yet this release introduces a fundamental shift beyond leaderboard standings by making test-time reasoning compute an explicit product control parameter.
Developers configure Opus 5 across low, medium, high, xhigh, or max effort levels. Reasoning is enabled by default, and output token budgets cover both internal reasoning trajectories and final visible responses. At xhigh and max levels, thinking cannot be disabled. The model also spawns subagents faster, communicates trajectory progress frequently, and verifies intermediate outputs automatically (S3, S7).
Opus 5 within the Claude model hierarchy
Anthropic places Fable 5 at the top tier for long-running agentic tasks and deep research. Mythos 5 remains in limited availability for specialized security and biology workloads. Sonnet 5 serves as the scale-oriented model for standard coding and agent workflows. Opus 5 occupies the core enterprise tier: offering near-Fable capabilities at half the token cost of Fable (S4, S5).
| Claude Model | Input per 1M Tokens | Output per 1M Tokens | Positioning |
|---|---|---|---|
| Claude Sonnet 5 | $2.00 | $10.00 | Frontier intelligence at scale |
| Claude Opus 5 | $5.00 | $25.00 | Complex agentic coding & enterprise work |
| Claude Fable 5 | $10.00 | $50.00 | Highest available capability |
| Claude Mythos 5 | $10.00 | $50.00 | Limited availability specialized work |
Configurable effort ladders in production
Opus 5 refines test-time compute control into a structured parameter. Anthropic specifies high as the default effort level rather than max. Official developer guidelines recommend starting at high, stepping down to medium or low for lower-latency pipelines, and reserving max effort for highly complex agentic tasks (S3, S7).
- Reasoning enabled by default: Migration from Opus 4.8 automatically incorporates adaptive thinking tokens.
- Shared max_tokens ceiling: Output token limits enclose both internal reasoning steps and final answer content.
- Enforced reasoning at xhigh/max: Disabling reasoning at these levels returns an HTTP 400 validation error.
- Low/medium effort utility: Lower effort levels offer cost-efficient execution for well-defined pipelines.
Conclusion
The release of Claude Opus 5 underscores an evolving paradigm in enterprise AI deployment. Model selection now expands into configuring dynamic compute budgets tailored to task complexity. By providing adjustable effort parameters and adaptive reasoning, developers can optimize success-adjusted operational costs across production workflows.
