benchrPricing breakdown
| Tier | Rate / 1M tokens |
|---|---|
| Standard input | $5.00 |
| Standard output | $25.00 |
| Cached input | $0.50 |
| Fast Mode input | $10.00 |
| Fast Mode output | $50.00 |
| Batch (50% off) | $2.50 / $12.50 |
| Context window | 1,000,000 tokens |
| Max output | 128,000 tokens |
The coding benchmark gap
Claude Opus 4.8 scores 88.6% on SWE-bench Verified — 4.6 points clear of GPT-5.5 (84.0%), 8 points clear of Gemini 3.5 Flash and DeepSeek V4-Pro (both 80.6%). For automated software engineering agents — bug-fix pipelines, PR generation, test writing, code review — this gap translates into real production outcomes: fewer failed attempts, lower retry costs, higher task success rates per dollar.
The supporting benchmarks reinforce this: GPQA Diamond at 93.6%, HumanEval at 96%, MMLU at 93.5%. These aren't narrow wins. Opus 4.8 is the strongest general coding and reasoning model available via commercial API.
The $25 vs $30 output price matters more than you think
Both Claude Opus 4.8 and GPT-5.5 charge $5/1M on input. But Opus 4.8's output is $25/1M vs GPT-5.5's $30/1M — 17% cheaper per output token. For output-heavy workloads where the model generates long code, detailed analyses, or extended documents, the per-task cost of Opus 4.8 is lower than GPT-5.5 despite identical input pricing.
Example: a task with 5K input tokens and 20K output tokens. Opus 4.8: $0.025 + $0.50 = $0.525. GPT-5.5: $0.025 + $0.60 = $0.625. At 100K tasks/month, Opus 4.8 saves $10,000 — with a higher benchmark score.
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 changes the cost equation significantly. A 100K-token system prompt sent 100 times per hour: without caching, $50/hour input cost; with 90% cache hit, $5 + $45 cached = $0.05 + $4.50 = $4.55/hour. For persistent long-context agents, caching is not optional — it's the difference between Opus 4.8 being expensive and affordable.
Fast Mode: when latency is the constraint
Fast Mode doubles the price to $10/1M input, $50/1M output, but reduces first-token latency. Standard Opus 4.8 has a 700ms first-token time and 68 tokens/second throughput — slower than GPT-5.5 (520ms, 82 tok/s) and much slower than Gemini 3.5 Flash (195ms, 289 tok/s). Use Fast Mode when you're running synchronous, customer-facing tasks where latency directly affects the experience. For async agents and batch work, standard mode is almost always the right call.
When Sonnet is enough
Claude Opus 4.8 costs roughly 5–12× more than Claude Sonnet 4.6 depending on caching. If your task passes with Sonnet, the upgrade to Opus is not justified. The correct workflow: evaluate on Sonnet first, identify the failure cases, and route only those cases to Opus. Most production teams find that 80–90% of their volume stays on Sonnet; Opus handles the tail that requires the highest reasoning ceiling.
Use-case fit
Best for: Highest-stakes automated coding tasks; complex multi-step agentic pipelines where failure is expensive; architecture-level reasoning; production SWE-bench-level work where the 4–8 point benchmark gap over competing models translates into fewer retries.
Skip if: You need cheap volume — Claude Sonnet 4.6 is the right default; you need the lowest latency — use Fast Mode or Gemini 3.5 Flash; your specific task type doesn't benefit from the benchmark delta, which is worth measuring before committing to Opus pricing.
Decision checklist
Before routing production traffic to Opus 4.8: run the same task on Claude Sonnet 4.6 and measure the quality gap. If Sonnet passes your eval threshold, there's no economic case for Opus. If you're comparing against GPT-5.5: note that Opus 4.8 output is cheaper ($25 vs $30/1M) — run an output-weighted cost comparison, not just an input-price comparison.
Check whether your workload benefits from caching. Large, repeated system prompts make Opus 4.8 significantly more affordable 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.