Every cost on our model pages comes from 79,301 real coding-agent requests from 75 users on the Standard Compute gateway (Sep 13 to Oct 12, 2026), each priced at the vendor's published list rates for its own input, cached, and output tokens; nothing is estimated from a price card alone.
Run current Claude, GPT, Gemini and open models in your agent from $19/month
See plans| Window | Sep 13 to Oct 12, 2026 (30 days, UTC) |
| Requests | 79,301 successful requests (HTTP 200) |
| Users / sessions | 75 users, 1,177 sessions |
| Clients | OpenCode, Kilo Code, Roo Code and GitHub Copilot Chat, as identified by the client in each request |
| Median input | 99,534 tokens (p25 59K, p90 280K, p99 533K) |
| Cache reads | 87% of input tokens; cache writes 0.9% |
| Output | 698 tokens per request on average (median 214) |
| Prompt size | 50% over 100K, 21% at 200K or more, 11% over 272K |
| Sessions | median 16 requests, mean 67.4 |
OpenCode alone had 68 users and 64,668 requests, so the OpenCode guide uses its own figures. Every other agent had fewer than 50 users and is folded into the pooled sample; a separate figure needs at least 50 users and 500 requests.
Each logged request keeps its own input, cache-read, cache-write and output token counts. For every model we price it as fresh input × input rate + cache reads × cache-read rate + cache writes × cache-write rate (the input rate where a vendor charges none) + output × output rate. If the request's input passes the model's long-context threshold, the higher rates apply to the whole request, as the vendors bill it. Requests larger than a model's context window are left out of that model's figures, and its page says how many.
From the priced requests we report: the average request; 100 requests; the median session (requests from one key with no pause longer than 10 minutes); and a month of regular use, the median 30-day total of the 20 users active on 5 or more days (1,478 requests). The three prompt-size rows are under 50K, 50K to 150K and over 150K input tokens. Vs pages compare both models on the requests that fit in both context windows.
List prices only, never promotions or our own plan prices. Where a vendor has a temporary price (Gemini 3.8 Flash until December 31, 2026) or time-of-day pricing (DeepSeek peak hours), we use the standard or peak rate and say so on the model's page. Devstral 2 uses Mistral's endpoint price on OpenRouter because Mistral's pricing page does not list it.
The gateway logs token counts and the client name for each request, never prompt text. We publish aggregates only: no per-user figures, and no agent cell under 50 users. The figures are regenerated with scripts/analytics/model-task-costs.mjs, which writes the JSON every page reads. See the result ranked in the cheapest models per coding-agent request.