The roster.
One base URL; the model field picks the route. Each model's
page carries three things kept deliberately separate: its maker's own
card, the benchmarks its maker publishes, and the numbers we measured on
the live endpoint.
Generation
2 models · chat, tools, vision-
Ornith-1.5 35B-A3Bflagshipornith-ai/ornith-1.5-35b-a3bAgentic coding at ~3B active parameters ; its maker cites SWE-bench Verified, Terminal-Bench 2.1
- first token
- 90 ms
- output, up to
- 330 tok/s
- in / out per 1M
- $0.25 / $1.20
-
Qwen3.8 27Bqwen/qwen3.8-27bVision, video and a reasoning-effort dial ; its maker cites LiveCodeBench v6, OSWorld-Verified
- first token
- 150 ms
- output, up to
- 250 tok/s
- in / out per 1M
- $0.30 / $2.03
Rail figures are ours, measured on the live endpoint: first token is the median to the first content token with reasoning off, cold cache, from the serving region; output speed is a single-stream ceiling. Dates, builds and full conditions are on each model’s page.
Retrieval
2 models · embeddings & rerank · batch-class- Qwen3-Embedding-8B
qwen/qwen3-embedding-8bSpecifically designed for text embedding and ranking tasks
- surface
/v1/embeddings- context
- 33K
- input per 1M
- $0.01
- Qwen3-Reranker-8B
qwen/qwen3-reranker-8bSpecifically designed for text embedding and ranking tasks
- surface
/v1/rerank- context
- 33K
- input per 1M
- $0.03
No speed cell on purpose: retrieval is batch-class: it
rides the generation models’ serving valleys, sheds with a retryable
429 under load, and that subordinate scheduling is the price
basis. Billing is input tokens only.
Whose numbers are whose
the rule every page here followsThe maker’s numbers
Benchmark scores on a model’s page are its maker’s own published figures, quoted with sources and retrieval dates, not our measurements, and not comparable between makers, who each ran their own harness.
Our numbers, and the gates behind them
First-token and output-speed figures are ours, measured on the live endpoint: single stream, medians, cold cache, conditions and build pins on each model’s page. Every build passes correctness gates before it takes traffic: checkpoint tensors reconciled against the vendor’s index, speculative output verified byte-identical to plain decode, and a live gate on all three wire formats plus a tool-call round trip. These are observations, not guarantees; run your own and use those instead.