Models and pricing

Struck-through figures are the model vendor’s own published rate; the figure beside them is what we charge. This table is generated from the same catalogue the gateway routes on, so it cannot drift from what you are actually charged.

The figures in the table are the peak rates. DeepSeek V4.1 Flash has a second, cheaper rate outside the peak window, shown under each one: peak is Mon-Fri 09:00-12:00 and 14:00-18:00 (UTC+8) and every other hour is off-peak. Which rate a request pays is decided by when it arrives, not when it finishes, so a long generation that crosses the boundary is billed entirely at the rate in force when you sent it. Both rates are in /v1/models as pricing and pricing_offpeak, with the window itself as pricing_window, so a client can price a request before making it.

• DeepSeek V4.1 Flash — Fast multimodal model. Reads text, images and video; cheapest per token we sell.
• DeepSeek V4 Pro — Frontier reasoning model. Best for agents, long-context analysis and code generation.
• DeepSeek V4 Flash — High-efficiency workhorse. Classification, extraction, chat and high-volume pipelines.
• GLM 5.3 Flash — Fast multimodal model with enforced structured output. Reads text, images and video.
• GLM 5.2 — Flagship model for long-horizon agent tasks.

Fresh prompt tokens at the input rate, cached prompt tokens at the cache-read rate, and generated tokens at the output rate. Reasoning tokens are generated tokens and are counted once, inside completion_tokens.

DeepSeek V4.1 Flash, DeepSeek V4 Pro, DeepSeek V4 Flash, GLM 5.3 Flash, GLM 5.2 think before answering. Those thinking tokens are generated tokens: they are counted inside completion_tokens, billed at the output rate, and reported separately as reasoning_tokens so you can see the split.

If you want short answers without paying for deliberation, pick a model whose Reasoning column below is a dash.

Ids are vendor/model, and they are the same strings OpenRouter publishes for the same models — so if you are moving from OpenRouter, the model names in your code do not change either.

"model": "deepseek/deepseek-v4-flash"
"model": "z-ai/glm-5.3-flash"

The un-namespaced ids we published earlier — deepseek-v4-flash, glm-5-3-flash — keep working, and will. So does the namespaced id with the vendor left off, and any capitalisation. GET /v1/models reports the namespaced id as id and the older one as slug; usage rows, key scopes and the URLs on this site are keyed by the slug.

Every model in the catalogue reasons before it answers unless you tell it not to, and the reasoning is billed at the output rate. That is the single largest thing a rate card cannot tell you, because it is a difference in how many tokens a task takes rather than what each one costs.

A customer measured it on a JSON extraction: the same six documents cost 2.7 times more on the model with the lower per-token price, because it spent most of its output budget thinking. One document in six never finished — the whole max_tokens budget went on reasoning and the answer came back empty with finish_reason: "length". On a short, mechanical task like extraction, the thinking buys very little.

"reasoning": {"enabled": false}

On the same extraction that one line took a request from 1,047 reasoning tokens to zero — about thirty times cheaper — with identical output. Reach for it whenever the task is mechanical: extraction, classification, reformatting, anything where you already know the shape of the answer. Leave thinking on for the work it is for.

Median reasoning tokens on a five-person seating puzzle at temperature 0, three runs per value, sending exactly what the gateway sends. Measured 2026-10-03. The prompt matters more than anything else here: a model given a question it can answer without thinking produces almost no reasoning whatever you ask for, which looks exactly like obeying you.

Reasoning tokens are counted in completion_tokens and broken out in completion_tokens_details.reasoning_tokens, so you can see what a task actually spent on thinking before deciding whether to switch it off.

JSON mode is response_format: {"type": "json_object"} — valid JSON, shape up to the model. JSON schema is type: "json_schema", where the shape is yours and the model is held to it. A model without the second tick may accept a schema and ignore it, so the column is measured rather than taken from a datasheet — see structured outputs.

Every model is served by at least one supplier, and where there is more than one we pick per request on cost and live latency, failing over before the first byte reaches you. A model that no supplier can currently serve returns 503 rather than a silent substitution — we will never answer with a different model from the one you asked for.