gpt-6.1-sol).
Both use the OpenAI Responses API with low reasoning effort. Requests are stateless;
tool continuations replay the complete response items, including encrypted reasoning.
Support controls
The support tuning panel displays the effective model. An explicit supported OpenAI model override remains in effect, including an existing Astra override. Missing or blank overrides use Sol; legacy Claude overrides continue to fall back to the default. Changing the code default does not rewrite saved configuration or enable the agent. The default limit remains six model turns. Billing, unclear requests, escalation keywords, and malformed or exhausted runs still go to human review. Draft review, Location policy, confidence checks, human handoff, send reservations, and both send kill switches remain required as before.Other AI uses
Underwriting overview stays ongoogle/gemini-3.1-flash-lite. Card research and
prospecting drafts stay on openai/gpt-4o-mini. Underwriting eligibility is decided
by the deterministic underwriting engine, not these models.
Rollout and recovery
This default change requires the normal application and backend release. No saved production configuration or customer threads are modified by the code change. Synthetic tests validate request settings, tool replay, and escalation controls; they do not establish live model quality, account access, latency, or actual cost. Paid evaluations and customer-thread trials need separate authorization. To recover the prior defaults, revert the model constants inconvex/lib/supportAgent.ts and app/api/chat/route.ts to gpt-6-astra and release
through the normal approval gates. Explicit support overrides remain unchanged.
Compatibility was checked on October 9, 2026 against the official
model page and
migration guide.
Sol supports Responses tools and low, medium, high, xhigh, and max reasoning;
none and minimal are unsupported.