Claude Opus 4.8 API pricing and benchmark context

Pricing breakdown

How to interpret the coding benchmark

Claude Opus 4.8 is reported at 88.6% on SWE-bench Verified. DeepSeek V4-Pro has a separate provider-published 80.6% result, while the checked Google material does not publish a directly comparable SWE-bench Verified figure for Gemini 3.5 Flash and the GPT-5.5 value formerly shown by benchr was unsupported. Those unsupported comparisons have been removed. Even properly sourced point gaps do not establish fewer retries, lower cost, or a higher success rate in your repository.

The same scorecard lists GPQA Diamond at 93.6%, HumanEval at 96%, and MMLU at 93.5%. Treat these as task-specific signals and confirm prompt format, tool use, correctness, and cost on a representative private evaluation set.

The $25 vs $30 output-price comparison

In this dated rate snapshot, Claude Opus 4.8 and GPT-5.5 both list $5/1M input, while Opus output is $25/1M vs $30/1M — 17% lower per output token. Opus has a lower token-only cost only when the compared tasks use the same input and output volumes; different success rates, retries, or response lengths can reverse the per-successful-task result.

Example: with identical 5K input and 20K output token counts, Opus 4.8 totals $0.525 and GPT-5.5 totals $0.625. At 100K such tasks, the arithmetic difference is $10,000. Their separately published benchmark figures do not prove a quality advantage for this workload.

Prompt caching: the 90% discount

Cached input tokens cost $0.50/1M — 90% off the standard $5/1M rate. The cache persists for up to 5 minutes. For agents with large, stable system prompts, caching can materially change cost. A 100K-token system prompt sent 100 times per hour costs $50 without caching; at a 90% cache-hit assumption, the modeled input cost is $9.50. Actual savings depend on eligible tokens, cache writes, expiry, and the hit rate you observe.

Fast Mode: when latency is the constraint

Fast Mode doubles the listed price to $10/1M input and $50/1M output in this snapshot. Anthropic does not publish one universal first-token or tokens-per-second guarantee for every region and request shape, and benchr has no controlled cross-region measurement. The undocumented latency values previously shown in comparisons have been removed. Measure both modes from your deployment region before making a customer-facing routing decision.

When Sonnet is enough

Claude Opus 4.8 can cost materially more than a Sonnet-tier model depending on token mix and caching. A defensible workflow is to evaluate the cheaper eligible model first, identify failure cases, and test whether routing those cases to Opus improves the outcome enough to cover the added cost. The appropriate routing share comes from your own evaluation; benchr does not claim a universal production split.

Use-case fit

Candidate for: difficult automated coding tasks, multi-step agentic pipelines where failures are expensive, and architecture-level reasoning—provided a representative evaluation shows an advantage at acceptable cost.

Compare alternatives if: you need low-cost volume or low latency, or your task does not benefit from the reported benchmark delta. Test current Sonnet, GPT, Gemini, and other eligible options rather than assuming a universal default.

Decision checklist

Before routing production traffic to Opus 4.8, run the same task on Claude Sonnet 4.6 and measure accepted outputs, review effort, retries, latency, and token cost. If Sonnet passes your threshold, it is the lower-priced candidate. Against GPT-5.5, use an output-weighted comparison rather than input price alone.

Check whether your workload benefits from caching. Large, repeated system prompts make Opus 4.8 far cheaper than the headline rate suggests. A 50K-token system prompt with 90% cache hit rate drops effective input cost to approximately $0.95 per million.