On this page
Models
GET /v1/models lists the models a key may use, with the metadata a client needs to configure itself. The list is per key: a key limited to some models lists only those.
The models#
| Model id | Context | Max output | Vision | Tools | Reasoning efforts | Input modalities |
|---|---|---|---|---|---|---|
glm-5.3 | 1,048,576 | 131,072 | no | yes | low, high, max | text |
deepseek-v4.1-flash | 1,048,576 | 393,216 | yes | yes | low, high, max | text, image |
deepseek-dgx | 1,048,576 | 393,216 | yes | yes | low, high, max | text, image |
glm-5.3-flash | 1,048,576 | 131,072 | yes | yes | no effort values published | text, image, video, file |
deepseek-v4.1-flash-abliterated | 1,048,576 | 131,072 | yes | yes | low, high, max | text, image |
glm-5.3-abliterated | 1,048,576 | 131,072 | no | yes | low, high, max | text |
glm-5.3-flash-abliterated | 1,048,576 | 131,072 | no | yes | low, high, max | text |
An embedding model is listed with output_modalities: ["embedding"] and its dimensions; Embeddings is its endpoint. GET /v1/models is the current list for any key.
A prompt plus the output it may ask for must fit the context window. A request over it is refused with the context length error, context_length_exceeded in the OpenAI format, so an agent that compacts on it keeps doing so.
deepseek-dgx is DeepSeek V4.1 Flash served on our own DGX Spark, and it is the only model the Sparks serve.
One id per model#
A model has one id: what GET /v1/models lists and what a request sends. A request may add the :zdr suffix to ask for zero-retention serving only (glm-5.3:zdr): the suffix is stripped before routing, and the request answers 503 zdr_unavailable rather than being served by a transient rented-GPU route. A key carries the same restriction as a switch, "zero-retention routes only", on the key's settings page: the simpler choice for a client that must never leave the zero-retention tier. A zero-retention key's GET /v1/models lists only the models a zero-retention route serves, so an agent does not learn an id it cannot use. Zero retention explains the two serving tiers.
Abliterated variants#
An id that ends in -abliterated is the model as released with the refusal behaviour the provider trained into it removed. A variant has no retention rule of its own: it is served under the same retention rules and the same acceptable use policy as every other model, and it is priced like any other model. The models page marks the variant and says what the change does, and zero retention explains which routes are contractually zero-retention and which are the transient rented-GPU tier.
Pricing#
Prices are in EUR per 1M tokens, in three buckets: input, cached input and output. A cached input token is charged the cached price, at or below the input rate (equal where the model's provider gives no cache discount), and a cache write is free because it is included in the input price. Reasoning tokens are output tokens, never a fourth price.
Prices are published and fixed per model. The price of a bucket does not depend on how a request is served, so a request is charged the price shown when it ran. The published price is the promise. This page prints no figure because the list changes; GET /v1/models returns the current prices for the key that asks, and the dashboard shows the same numbers.
Ask for the list#
The fields#
| Field | Meaning |
|---|---|
id | The string to send as model |
object | Always model |
created | When the model was added, Unix seconds |
owned_by | Always zdr.sh |
name | The display name, for example GLM 5.3 |
context_window, context_length | The context window under both names |
max_output_tokens | The largest output the model accepts |
input_modalities | text, image, video, file |
output_modalities | text on a chat model, embedding on an embedding model |
embedding_dimensions | The vector size on an embedding model, 0 on a chat model |
reasoning | Whether the model can think before answering |
supported_reasoning_efforts | The values reasoning_effort may take, [] when the model has no effort control |
supports_tools | Whether the model can call tools |
input_eur_per_1m, cached_eur_per_1m, output_eur_per_1m | The price of each bucket, in EUR per million tokens |
Anthropic shape#
Send an anthropic-version header and the answer is the Anthropic list shape. The ids are the same; the metadata is the minimum the shape carries.
A claude- id is never listed, even when the key maps one. Messages has the mapping.
Discovery in practice#
- opencode does not call this endpoint for a custom provider: list the models in
opencode.json - Hermes Agent probes it, so a model added to the list shows up without a config change
- Claude Code keeps only ids containing
claudeoranthropic, so add zdr.sh ids by hand as in the Claude Code page