[BOT ISSUE] Missing AI21 Labs as a direct API provider family (Jamba models)

[github-actions]

Summary

AI21 Labs operates a direct API with its proprietary Jamba model family, but is not represented as a provider family in the Braintrust proxy. The ModelEndpointType enum in packages/proxy/schema/models.ts does not include an ai21 entry, and no Jamba models appear anywhere in model_list.json.

AI21 API details

• Base URL: https://api.ai21.com/studio/v1/chat/completions
• Format: OpenAI-like chat completions endpoint pattern (/v1/chat/completions)
• Available models:
  • jamba-mini — lightweight Jamba model
  • jamba-large — full-size Jamba model
• Architecture: Jamba is a proprietary hybrid SSM/Transformer architecture unique to AI21 Labs — not a fine-tune or variant of any open-source model.

Verified pricing

From the official AI21 pricing page:

Model	Input $/MTok	Output $/MTok
jamba-mini	$0.20	$0.40
jamba-large	$2.00	$8.00

Official sources

• API reference: https://docs.ai21.com/reference/jamba-1-6-api-ref
• Pricing: https://www.ai21.com/pricing
• Model availability across platforms: https://docs.ai21.com/docs/model-availability-across-platforms

Local files inspected

• packages/proxy/schema/models.ts — ModelEndpointType array (line 20-41) does not include ai21
• packages/proxy/schema/model_list.json — grep for ai21 and jamba returns no matches
• packages/proxy/scripts/sync_models.ts — no AI21 provider mapping exists

Why high confidence

AI21 Labs has a publicly documented direct API with published per-token pricing and a /chat/completions endpoint that follows OpenAI conventions. The Jamba architecture is genuinely proprietary (not derived from open-source weights). Jamba models are also available on AWS Bedrock (Jamba Large 1.5, Jamba Mini 1.5), but even those Bedrock variants are absent from the catalog. The proxy already supports other OpenAI-compatible providers with similar patterns (Together, Fireworks, Groq, Cerebras, Perplexity).

Fields not verified

• Context window sizes (max_input_tokens, max_output_tokens) — not published on the pricing page or model availability page. The downstream fix job should check model detail pages or the API response for these.
• Explicit OpenAI SDK compatibility — the endpoint pattern (/v1/chat/completions) strongly suggests compatibility, but the docs don't explicitly state "OpenAI-compatible." The downstream fix job should verify.

Relationship to existing issues

• Issue BT-4825 tracks missing DeepSeek as a provider family — separate provider.
• Issue BT-4841 tracks missing Cohere as a provider family — separate provider.

{
  "kind": "missing_model",
  "provider": "ai21",
  "models": ["jamba-mini", "jamba-large"],
  "status": "active",
  "model_specs": {
    "jamba-mini": {
      "format": "openai",
      "flavor": "chat",
      "input_cost_per_mil_tokens": 0.2,
      "output_cost_per_mil_tokens": 0.4,
      "available_providers": ["ai21"]
    },
    "jamba-large": {
      "format": "openai",
      "flavor": "chat",
      "input_cost_per_mil_tokens": 2.0,
      "output_cost_per_mil_tokens": 8.0,
      "available_providers": ["ai21"]
    }
  },
  "source_urls": [
    "https://www.ai21.com/pricing",
    "https://docs.ai21.com/reference/jamba-1-6-api-ref"
  ]
}

[github-actions]

Autofix skipped because the issue body/title did not contain parseable bot metadata for a model catalog change.