Extended context

The standard context window is 262,144 tokens for prompt and reply together. Extended context raises that to 1,048,576 tokens on models that support it, for work that genuinely needs everything in view at once: a whole codebase, a year of contracts, a long transcript set.

Turning it on

syntic-beta: extended-context-2026-12-01

{
  "model": "Amara S2",
  "context": "extended",
  "max_tokens": 16000,
  "messages": [{ "role": "user", "content": [{ "type": "file", "file": { "file_id": "file_01JD…" } }, { "type": "text", "text": "List every clause that limits liability, by contract." }] }]
}

Models that support it show context_window_extended: 1048576 in GET /v1/models. Others answer 400 unsupported_parameter.

Pricing

Proposed — not offered. Whether requests above 262,144 input tokens carry a separate long-context rate, and what it would be, is an owner-level commercial decision that has not been made. No rate is offered. The response field usage.context_tier (standard or extended) is part of the design so that any future rate can be reported per request.

Rate limits

Long requests draw on a separate long-context token-per-minute limit, so one large job does not starve your normal traffic. The limit is listed per model on Rate-limit tiers.

Getting good answers from very long inputs

• Put the documents first and the question last; models attend better to instructions at the end of a long prompt.
• Wrap each document in tags with its name: <document name="msa-2024.pdf">…</document>. Ask for answers that name the document.
• Ask the model to quote the relevant passage before it answers. Quotes are cheap to verify and expose guessing.
• Cache the large prefix with prompt caching if you will ask several questions about the same material.
• Consider whether retrieval would do: it is usually cheaper and faster when only a small part of the material matters to each question.

Errors