Claude Opus 5 API Pricing: Complete Cost Guide and Optimization Playbook
Claude Opus 5 API costs explained: $5/$25 pricing, fast mode economics, effort-based savings, batch API, and a practical playbook for cutting spend.
TL;DR
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens — unchanged from Opus 4.8 and half of Fable 5's input price. Fast mode doubles the price ($10/$50) for roughly 2.5× speed. The biggest levers for cutting spend: effort-based routing, prompt caching against the 1M context window, the Message Batches API for offline work, and letting the model's improved verification reduce retry loops.
The Price List
- Standard: $5/M input, $25/M output
- Fast mode: $10/M input, $50/M output, ~2.5× faster
- Context window: 1M tokens; max output 128K (300K via Message Batches API beta)
For context within the family: Sonnet 5 remains the high-volume budget tier, Fable 5 charges $10/M input for flagship capability, and Opus 5 now occupies the value-frontier position between them.
What a Month Actually Costs
Rough shapes for common workloads at standard pricing:
- Code assistant, 5 engineers, moderate use: typically low hundreds of dollars per month at medium effort — often less than the Opus 4.8 equivalent because fewer retries mean fewer total tokens
- Agentic pipeline, 10K tasks/month: dominated by output tokens; effort routing (low for triage, high for the hard tail) commonly cuts 30-50% versus flat-high
- Document analysis over large corpora: input-heavy, so the $5/M input rate and prompt caching do most of the work
The Optimization Playbook
1. Route by effort. Don't run everything at one setting. Low effort for mechanical steps, medium as default, high for the hard tail. This is the single largest lever. 2. Cache aggressively. With a 1M-token window it's tempting to resend enormous contexts every call. Prompt caching makes repeated system prompts, codebases, and document sets dramatically cheaper — structure prompts so the stable prefix stays stable. 3. Batch what can wait. The Message Batches API trades latency for discounts and unlocks 300K output. Nightly report generation, backfills, and eval runs belong there. 4. Cap outputs. Set max_tokens realistically per route. Output tokens cost 5× input; unbounded generation is the classic silent budget leak. 5. Count retries, not requests. Opus 5's iterative verification means it succeeds first-try more often than 4.8. When you compare model costs, compare cost per successful task — a cheaper model that needs three attempts is not cheaper. 6. Reserve fast mode for humans. Fast mode's 2× premium is worth it for interactive UX, wasted on background jobs.Cost vs the Alternatives
The strategic point of Opus 5's pricing: near-Fable performance at half the input rate, and on several published benchmarks (Frontier-Bench, GDPval-AA v2, OSWorld 2.0) it beats Fable 5 outright — meaning for many workloads the cheaper model is also the better one. Cost-per-task comparisons should also include the effort dial: low-effort Opus 5 encroaches on territory that used to belong to mid-tier models.
Conclusion
Opus 5's pricing headline is "unchanged," but its economics are not: effort routing, caching, batching, and higher first-try success compound into a meaningfully lower cost per completed task than any previous Opus. Budget on cost-per-outcome, not cost-per-token, and the model pays for itself quickly.