A direct API cost comparison: headline rates, the per-call math on a typical workload, and which one to pick beyond price.
On a standard 1,500-input / 500-output call, Claude Sonnet 4.6 is about 47% cheaper — $0.0120 versus $0.0225 per call. At 10,000 calls a month that is roughly $120.00 vs $225.00.
Output tokens cost several times more than input on both models, so the more your workload leans toward long generated answers, the more the output rate dominates the bill.
| Metric | Claude Sonnet 4.6 | GPT-5.5 |
|---|---|---|
| Vendor | Anthropic | OpenAI |
| Input price / M tokens | $3.00 | $5.00 |
| Output price / M tokens | $15.00 | $30.00 |
| Context window | 1M | 1.05M |
| Cost per typical call | $0.0120 | $0.0225 |
| Cost per 10,000 calls | $120.00 | $225.00 |
Beyond the per-token math, tokenizer efficiency on non-English text and your own quality evals can shift the real cost. On context, GPT-5.5 offers the larger window (1.05M), which matters for long documents and agent histories. Price your actual prompt in the language you serve with the calculator before committing.
Both providers support prompt caching (cached input bills at roughly 10% of the standard rate) and batch processing (about 50% off for 24-hour-tolerant jobs). If one model lets you cache a larger static prefix in your setup, it can become cheaper in practice even when its headline rate is higher.
Claude Sonnet 4.6 wins on raw price for a typical workload. Pick GPT-5.5 only when its quality, context, or tokenizer advantages on your specific task outweigh the difference.
Claude Sonnet 4.6 — about 47% cheaper on a typical 1,500-input / 500-output call ($0.0120 vs $0.0225).
Claude Sonnet 4.6: $3.00 input / $15.00 output. GPT-5.5: $5.00 input / $30.00 output.
The one with the lower output rate, since generation is output-bound: Claude Sonnet 4.6 ($15.00/M output).