Model catalog

Four open-weight text models: a small fast rung, a mid workhorse, a large flagship, and a coding model.

Atlas serves a deliberately short ladder. Every model in it is kept warm, so there is no cold-start rung and no model that is cheap because it is rarely available.

This page is a summary. GET /v1/models is the authoritative catalog — it publishes each model's context length, capabilities, JSON Schema subset, and provisional price as data, and it reflects operator changes without a docs release. Read it from your own code rather than hard-coding the table below.

The ladder

Model	Context	Tool calling	Structured output	Price / MTok
`atlas-small-1`	32,768	Yes	Yes	$0.15 in · $0.45 out
`atlas-mid-1`	131,072	Yes	Yes	$0.50 in · $1.50 out
`atlas-large-1`	262,144	Yes	Yes	$2.00 in · $6.00 out
`atlas-code-1`	131,072	Yes	No	$0.40 in · $1.20 out

All prices are provisional and all beta usage is free. See Pricing.

atlas-code-1 does not support structured output. A request that sends response_format: {"type": "json_schema"} to it is refused with a 400 invalid_request_error carrying the code model_does_not_support_structured_output — refused before the request is admitted, so a capability mismatch never costs you tokens. Check the model's capabilities rather than assuming the ladder is uniform.

Choosing a rung

atlas-small-1

The fast, cheap rung. Classification, extraction, routing, short rewrites — high-volume work where latency and unit cost dominate and 32K of context is enough.

atlas-mid-1

The default. Start here unless you know you need something else: it carries the same capabilities as the flagship at a quarter of the price, with 128K of context.

atlas-large-1

The flagship, and the only 256K-context rung. Long documents, multi-step reasoning, and prompts where the mid model's answers are not good enough on your own evaluations.

atlas-code-1

Code generation, review, and transformation, at close to the small model's price. Supports tool calling; does not support structured output.

The honest advice is to benchmark the mid rung on your own prompts before reaching for the large one. Atlas does not claim a quality ranking that holds across workloads, and the catalog is short precisely so that trying two rungs is cheap.

Reading the catalog

GET /v1/models returns the OpenAI list envelope, with Atlas's published metadata added to each entry. The added keys are additive, so an OpenAI SDK parsing this ignores them.

curl https://api.inference.runatlas.com/v1/models \
  -H "Authorization: Bearer $ATLAS_API_KEY"

{
  "object": "list",
  "data": [
    {
      "id": "atlas-mid-1",
      "object": "model",
      "created": 1749513600,
      "owned_by": "atlas",
      "context_length": 131072,
      "max_request_bytes": 1048576,
      "capabilities": {
        "tool_calling": true,
        "structured_output": true,
        "json_schema_subset": "atlas-json-schema-subset-v1"
      },
      "pricing": {
        "input_per_mtok": "0.50",
        "output_per_mtok": "1.50",
        "currency": "USD",
        "provisional": true,
        "reasoning_tokens_billed_as": "output"
      },
      "lifecycle": {
        "status": "active",
        "superseded_by": null,
        "shutdown_at": null
      }
    }
  ]
}

context_lengthinteger

Total tokens across the prompt and the completion.

max_request_bytesinteger

The largest request body accepted for this model. A larger one is refused with 413 and the code request_too_large.

capabilities.tool_callingboolean

Whether the model accepts tools and tool_choice. See Tool calling.

capabilities.structured_outputboolean

Whether the model accepts response_format: json_schema. See Structured outputs.

capabilities.json_schema_subsetstring | null

The identifier of the JSON Schema subset this model's schemas are validated against, or null when the model has no structured-output support. Currently always atlas-json-schema-subset-v1.

pricing.reasoning_tokens_billed_asstring

Always "output". Where a model emits a reasoning prelude, those tokens are counted inside completion_tokens — never add the two figures together.

lifecycleobject

status is active or superseded; superseded_by names the successor identifier; shutdown_at is when a superseded identifier stops serving. See Model lifecycle.

Retrieving one model

curl https://api.inference.runatlas.com/v1/models/atlas-mid-1 \
  -H "Authorization: Bearer $ATLAS_API_KEY"

An identifier that was never published returns 404 with the code model_not_found. An identifier that has been retired — past its published shutdown date — also returns 404, but with the code model_retired, so you can tell "you typed it wrong" from "this used to work". See Model lifecycle.

Not in the catalog

Atlas serves text-generation models only. There are no embedding, audio, image-generation, or vision models, and no /v1/completions, /v1/embeddings, or fine-tuning endpoints. Requests carrying image content parts are not served by any catalog model.