benchrHead-to-head: how this rate card stacks against OpenAI's daily driver is covered in Sonnet 4.6 vs GPT-5, and against the open-weight challenger in DeepSeek V4-Pro vs Sonnet 4.6.
Pricing breakdown
| Tier | Rate / 1M tokens |
|---|---|
| Standard input | $3.00 |
| Standard output | $15.00 |
| Cached input | $0.30 |
| Context window | 1,000,000 tokens |
| Max output | 64,000 tokens |
How to test Sonnet against Opus
Anthropic's published scorecard reports 79.6% SWE-bench for Sonnet 4.6 and 88.6% for Opus 4.8. Those provider-reported scores describe a defined benchmark setup; the nine-point difference does not predict the acceptance-rate gap for chat, summarization, code review, or your repository.
Sonnet is a sensible lower-cost candidate. Run the same versioned prompts and tasks on both tiers, blind-review the outputs, and escalate only the task classes where Opus produces a measured improvement that justifies its higher token cost.
The 64K max output ceiling
Claude Sonnet 4.6 supports 64,000 output tokens per response. Claude Opus 4.8 supports up to 128,000 output tokens; both models list a 1M-token context window. A full 64K output at Sonnet pricing ($15/1M) costs $0.96, while the same token count at Opus 4.8 pricing ($25/1M) costs $1.60. Opus does not require splitting at 64K.
Caching and effective rates
Claude Sonnet 4.6 cached input costs $0.30/1M — 90% off the $3 standard rate. For a typical agent with a 50K-token system prompt running 1,000 calls per day, caching saves approximately $135/day versus uncached at $3 per million. With high cache hit rates, the effective input cost drops to roughly $0.57 per million — competitive with Haiku uncached pricing for input-heavy workloads.
Cost scenarios
At 20M input + 5M output per month: $60 + $75 = $135/month. Opus 4.8 at the same volume: $100 + $125 = $225/month — $90 more for the Opus premium. With 90% cache hit rate on Sonnet 4.6: approximately $11.40 + $75 = $86.40/month. For a startup processing customer queries, this is a practical monthly budget difference that compounds at scale.
Use-case fit
Candidate for: Production chat and assistant applications, code review, moderate-complexity generation, long-document workflows, and deployments where its lower listed rate may matter.
Compare alternatives if: The workload has costly failure modes, strict latency targets, or simple high-volume tasks. Test Opus, Haiku, and other eligible models on the same acceptance criteria rather than inferring fit from tier names.
Decision checklist
Before using Opus 4.8 instead: run a 50-sample eval of your hardest task category on both Sonnet 4.6 and Opus 4.8. If you can't detect a quality difference, stay on Sonnet — the 40% cost reduction compounds at production volume.
Before using Haiku 4.5 instead, compare it with Sonnet on the same tasks. Anthropic's published SWE-bench figures differ by about six points, but that benchmark gap does not establish the quality difference for classification, extraction, or your analysis workflow.