OpenAI

OpenAI's gpt-5.5 pairs top-tier reasoning with a 400k context window at a mid-tier price — which makes it both a strong worker and the catalog's best cross-family judge over a Claude or GLM worker. Stella speaks the OpenAI Responses dialect (not the legacy chat-completions shape) against https://api.openai.com/v1.

Setup

export OPENAI_API_KEY="sk-..."

OpenAI sits third in auto-detection, so the key alone is enough. Prefer a file? Store the key in ~/.config/stella/credentials.toml, or rename the env var Stella reads with providers.openai.api_key_env — see overriding the key source.

Selecting models

The provider id is openai; the default model is gpt-5.5.

stella --model openai/gpt-5.5 run "add integration tests for the exporter"
export STELLA_MODEL=openai/gpt-5.5

Custom base URL

Route the built-in provider through an Azure OpenAI endpoint or an OpenAI-compatible proxy by overriding the base URL:

{
  "providers": {
    "openai": {
      "base_url": "https://my-resource.openai.azure.com/openai/v1"
    }
  }
}

One constraint: the built-in openai provider's dialect is fixed to OpenAI Responses, so the endpoint must speak it (Azure's /openai/v1 surface does). For a gateway that only speaks legacy chat completions, define a custom provider id with the default openai-compatible dialect instead. --base-url (env form STELLA_BASE_URL) does the same for one run and outranks the settings value; in a project's .stella/settings.json, base_url is ignored unless STELLA_TRUST_PROJECT=1.

Reasoning-model parameter rules

The Responses API has a hard rule for reasoning models, and Stella honors it: for the gpt-5 and o-series families, temperature and top_p are omitted from the request — the API rejects them outright, and Stella drops them rather than letting your run die on a 400. What you control instead:

service_tier: "flex" trades latency for a cheaper rate on tolerant workloads; "priority" does the reverse. Set any of these per agent in agent_engine_config.

In the pipeline

gpt-5.5 is the judge of choice when your worker is a Claude or GLM model — a different family judging at high effort resists same-model bias. It appears as the judge in the max-quality profile and as an alternate judge in the balanced profile; its 400k window and mid price also make it a first-rate worker in its own right. Wire it per agent with agent_engine_config.

Catalog pricing

List prices from Stella's seed catalog — verify with OpenAI; your local catalog re-syncs from models.dev via stella models refresh. The cached-input rate is what most worker input costs mid-run (see the Model guide).