Skip to main content
Dari defaults to excluding model/provider candidates when the estimated full input exceeds their standard-price context limit. This is a pricing boundary, not a universal context cap: models with no long-context surcharge retain their full window. Cached input counts toward this boundary. If no compatible candidate remains, the request fails with context_length_exceeded; Dari does not silently truncate the conversation. Compact or shorten the conversation before retrying. The model catalog API exposes standard_context_limit alongside context_window. A null standard limit means the offering has no configured pricing boundary. Codex receives a context window capped at that boundary. Input accounting uses estimates, including prior provider-reported usage when available. Estimates are not exact provider token counts and cannot guarantee a request close to the boundary will avoid a surcharge.

Allow Long Context

Enable Allow Long Context in a router’s Advanced Settings to opt in to the provider’s full context window. This setting defaults to false for existing and new routers. Long-context rates can apply to the entire request, including cached input and output, not only the tokens above the threshold. Set allow_long_context: true in the create/update router API or a router YAML manifest. From the CLI:
Omitting the field on update preserves the current setting. The model catalog exposes pricing tiers with input_tokens_above and the applicable per-million rates. Routing estimates reevaluate tiers as projected conversations grow; provider usage accounting selects the tier using the full reported input count.