Compare AI models
Pick up to three models. Price, limits and live performance, side by side.
| Metric | o1 | Claude Sonnet 5 | Claude Opus 5.5 |
|---|---|---|---|
| Input / 1M | $15.00 | $3.00 | $4.00 |
| Output / 1M | $60.00 | $15.00 | $20.00 |
| Input / 1M (credits) | 335,821 | 67,165 | 89,553 |
| Output / 1M (credits) | 1,343,284 | 335,821 | 447,762 |
| Context window | 200,000 | 1,000,000 | 1,000,000 |
| Longest answer | 32,000 | 32,000 | 32,000 |
| Latency (avg) | — | — | — |
| Latency p50 | — | — | — |
| First token p50 | — | — | — |
| Throughput | — | — | — |
| Uptime | — | — | — |
| Tokens (7d) | — | — | — |
| Reads images | Yes | Yes | Yes |
| Reads PDFs natively | No | Yes | Yes |
How these numbers are measured
Every request the chat and the API send to a model is counted the moment it finishes: tokens in and out, total latency (from the request to the last byte), time to first token (streamed answers) and throughput (output tokens per second, counted after the first token, only on answers with at least 20 tokens). Medians are estimated from a latency histogram, so they are close but not exact.
Uptime is 1 minus the share of requests the provider failed (timeouts, rate limits, server errors and network failures). A request rejected for being malformed does not count against the provider. Uptime is shown only after 50 requests in the period, and latency only after 10 answers, so a model with three lucky requests never tops the list.
Days are in UTC. Today counts from midnight UTC; this week is the last 7 days including today. Synthetic checks, when enabled, count for uptime and speed but never for popularity.