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. Deciding between this and running it yourself? Self-hosting, a rented GPU, or hosted, costed.
Generation
2 models · chat, tools, vision
GLM-5.3-Flashzai/glm-5.3-flashStrong, cheap and the fastest here: the pick for anything that finishes in a few calls, with the full 1M window when the prompt is big.
Approaching Claude Opus 4.8 on coding and agentic benchmarks, at one-tenth the price of GLM-5.2; its maker cites Terminal Bench 2.1, DeepSWE v1.1, GDPVal-AA v2
- output, up to
- 179 tok/s
- in / out per 1M
- $0.15 / $0.50
DeepSeek-V4.1-Flashdeepseek-ai/deepseek-v4.1-flashSmart, with reasoning on by default, and the cheapest here for long iteration: once a loop replays its history, the cache carries most of the bill at the cached rate.
A multimodal Mixture-of-Experts model with 552B backbone parameters and support for contexts of up to one million tokens; its maker cites Terminal-Bench 2.1, DeepSWE v1.1, Agent's Last Exam
- first token
- 220 ms
- output, up to
- 101 tok/s
- in / out per 1M
- $0.30 / $1.20
Rail figures are ours, measured on the live endpoint: first token is the published figure at the fast edge, pinned to its build, never a median; where the model has a reasoning-off mode it is measured with reasoning off, otherwise it is the first streamed token with the default thinking on (each model page carries the current measurement record). GLM-5.3-Flash publishes no first-token figure. Cold cache, and the RTT between you and the edge is not part of the number; output speed is a single-stream ceiling. Dates, builds and full conditions are on each model’s page.
Speech to text
2 models · streaming & whole file · billed by the audio hour
Whisper Large v3 Hebrewivrit-ai/whisper-large-v3-heHebrew transcription on the ivrit-ai Whisper Hebrew weights, streaming or whole file, priced by the audio hour rather than by the token.
- endpoints
/v1/realtimestreaming/v1/audio/transcriptionswhole file- streaming
- $0.18 / audio hour
- whole file
- $0.10 / audio hour
Qwen3 ASR Englishqwen/qwen3-asr-enEnglish transcription on a small, fast model, streaming or whole file, priced by the audio hour rather than by the token.
- endpoints
/v1/realtimestreaming/v1/audio/transcriptionswhole file- streaming
- $0.15 / audio hour
- whole file
- $0.04 / audio hour
No speed cell on purpose, and not because one is pending: what we measure on these models is how many real-time streams a card carries, which is a capacity figure we keep internal, not a customer claim. Billing is audio duration, silence included, with no minimum per request: an hour of audio costs the hourly rate whether it arrives as one file or as a live stream. Request shapes are in the API documentation.
Retrieval
2 models · embeddings & rerank · dedicated capacity- Qwen3-Embedding-8B
qwen/qwen3-embedding-8bPick it for memory: turn your documents into vectors once and retrieve them forever, at the lowest rate on the roster.
Specifically designed for text embedding and ranking tasks
- surface
/v1/embeddings- context
- 16K
- input per 1M
- $0.02
- Qwen3-Reranker-8B
qwen/qwen3-reranker-8bPick it for precision: score your top retrieval hits so the model answers from the right ones, not the nearest ones.
Specifically designed for text embedding and ranking tasks
- surface
/v1/rerank- context
- 16K
- input per 1M
- $0.05
No speed cell on purpose. Since 2026-09-02 retrieval hasdedicated capacity: it is admitted up to the advertised concurrency without yielding to the generation lanes, and only beyond that concurrency does the lane cap shed with a retryable 429. That posture 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, published reference figures (never medians), cold cache, conditions and build pins on each model’s page. Every build passes correctness gates before it takes traffic: what we serve is reconciled against the vendor’s published model, and a live gate covers all three wire formats plus a tool-call round trip. These are observations, not guarantees; run your own and use those instead.