benchr
Side-by-side specs
| Dimension | GPT-5 | Claude Opus 4.8 |
|---|---|---|
| Provider | OpenAI | Anthropic |
| Released | Aug 7, 2025 | May 28, 2026 |
| Input / 1M | $1.25 | $5.00 |
| Output / 1M | $10.00 | $25.00 |
| Cached input / 1M | $0.125 (90% off) | $0.50 (90% off) |
| Context window | 400,000 tokens | 1,000,000 tokens |
| Max output | 128,000 tokens | 128,000 tokens |
| SWE-bench Verified | 74.9% | 88.6% |
| GPQA Diamond | not published | 93.6% |
| Fast mode | No | Yes — $10/$50, ~2.5× speed |
Cost for an example workload
At 10 million input tokens per month — a moderate agentic workload — you're spending $12.50 on GPT-5 or $50 on Opus 4.8. That $37.50/month difference is $450/year. At production scale of 500M input tokens monthly, it's $7,500 vs $30,000.
Both models discount cached input by 90%: GPT-5 charges $0.125/1M and Opus 4.8 charges $0.50/1M. Once cached, a reusable 40,000-token prefix costs $0.005 per GPT-5 call versus $0.02 per Opus call. Caching preserves GPT-5's 4× input-price advantage for large, static prompts. Model your actual cache-hit rate and output mix before budgeting.
How to interpret the coding results
SWE-bench Verified presents real GitHub issues from open-source Python repositories. The model must read the issue, understand the code, write a fix, and pass the existing tests — no scaffolding, no hints. At 88.6%, Opus resolves roughly 88 of 100 issues without human intervention. GPT-5 resolves about 75.
That gap matters most when humans are out of the loop. If you're running a CI bot that autonomously applies model-generated patches, those 13 additional failures per 100 issues translate directly to broken builds or manual cleanup. At 200 automated PRs per day, that's 26 more failures requiring engineer attention. At $30 per engineer-hour, the math moves fast.
Outside fully automated code repair, the gap gets much smaller. GPT-5 handles functions from a spec, docstrings, and unfamiliar-code explanations well enough for most teams. The Opus premium pays for more autonomous work, not routine assistance.
Context window: when 400K isn't enough
For most production workloads, GPT-5's 400K-token context is sufficient. Where it bites: entire codebases, long legal contracts, multi-session agent memory, or document analysis jobs that want to fit everything in at once. Opus 4.8's 1M-token context handles all of these without truncation. If your use case regularly approaches 300K tokens, this is a functional constraint, not a spec comparison.