benchrPricing breakdown
| Tier | Rate / 1M tokens |
|---|---|
| Standard input | $5.00 |
| Standard output | $25.00 |
| Cached input | $0.50 |
| Context window | 1,000,000 tokens |
| Max output | 32,000 tokens |
Opus 4.7 vs Opus 4.8: the 1-point question
Claude Opus 4.8 improves SWE-bench Verified from 87.6% to 88.6% — one percentage point. Both versions are priced identically at $5/1M input and $25/1M output. In practice, a 1-point SWE-bench difference means Opus 4.8 resolves roughly 1 additional issue per 100 that would have failed on Opus 4.7. For most production workloads, this difference is invisible unless you're running large coding pipelines at scale with measurable pass/fail rates.
If you already have Opus 4.7 running in production and your quality metrics are acceptable, the cost of migration (prompt re-testing, regression checks, system prompt tuning) likely outweighs the benchmark improvement. For greenfield projects, start with Opus 4.8 — there is no reason to use the older version.
Opus vs Sonnet: the real cost decision
Claude Sonnet 4.6 costs $3/1M input and $15/1M output — 40% cheaper on input, 40% cheaper on output. Sonnet 4.6 scores 79.6% on SWE-bench Verified, an 8-point gap below Opus 4.7. For complex multi-file refactors, architecture reviews, or tasks where output quality has a direct dollar value, that 8-point gap is meaningful. For general-purpose chat, summarization, and classification, Sonnet 4.6 handles the workload without paying the Opus premium.
A practical test: if you cannot tell the difference between Opus and Sonnet on your specific tasks in a blind eval, you should be on Sonnet. The 8-point SWE-bench gap shows up most clearly in multi-step code generation where errors compound across steps.
Caching and the effective rate
Cached input for Claude Opus 4.7 costs $0.50/1M — 90% off standard. For agents with a large fixed system prompt (say 80K tokens of tool definitions and instructions), a 90% cache hit rate brings the effective input cost to roughly $0.95 per million tokens — comparable to uncached Sonnet 4.6 at $3 per million. Output doesn't cache, so output-heavy workloads still pay the full Opus premium on the output side.
Cost scenarios
At medium volume — 5M input tokens and 1.5M output per month — Opus 4.7 costs $25 + $37.50 = $62.50/month. Sonnet 4.6 at the same volume: $15 + $22.50 = $37.50/month. The Opus premium is $25/month at this scale. At 90% cache hit on input: approximately $4.75 + $37.50 = $42.25/month — the cached Opus rate is within 13% of uncached Sonnet, which narrows the decision considerably for systems with stable, cacheable context.
Use-case fit
Best for: Complex multi-file code generation; architecture review; long-document analysis requiring sustained reasoning over 100K+ tokens; production agent loops where output quality directly maps to business value and the Opus premium is justified by error-reduction ROI.
Skip if: You're starting a new project — use Opus 4.8 instead. Also skip if Sonnet 4.6 handles your eval benchmarks — the 40% cost difference compounds quickly at production scale.
Decision checklist
Run your actual task sample on both Claude Sonnet 4.6 and Claude Opus 4.7 in a blind evaluation before committing to Opus pricing. If you can't detect the quality difference in your specific workload, Sonnet is the right economic choice.
For Opus 4.7 vs 4.8 specifically: measure whether your pipeline has task types where a 1-point SWE-bench improvement is detectable at your scale. If pass rates are already above 90%, the marginal 1% improvement from Opus 4.8 is noise.