graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs
0292. Sol's subscription context default is 500k
Status: accepted 2026-10-18.
Context
The blanket subscription input cap predates GPT-6.1 Sol. It reduced even a 500,000-token account catalog row to 270,000, causing early compaction. The upstream announcement raises this model's Codex and Work default from 272k to 500k. The larger advertised maximum is a separate value, not a reason to select it automatically.
Decision
Use 500,000 for the exact gpt-6.1-sol model on codex and chatgpt-new,
in both offline rows and the subscription clamp. Continue honoring smaller
provider catalog windows. Keep other subscription models at the existing
conservative cap. Leave direct API and gateway windows unchanged.
The shared provider context remains authoritative for the engine, children, REPL and ACP context reporting. Keep the compaction percentage unchanged: its default 80% now starts compaction at 400,000 for this subscription model.
Consequences
Long sessions retain more history without opting into the maximum window. A stale smaller account catalog stays conservative until refreshed. This change does not increase account quotas or promise that larger sessions use the same amount of subscription allowance. Revisit the cap only with new route-specific evidence, not a name-only API maximum.