Biorouter v1.92.1

Biorouter v1.92.1

This patch release brings the UCSF Versa model lists in line with what UCSF
serves today. Every model now offered by Versa API Azure and Versa API Bedrock
was tested through Biorouter against the UCSF gateways on October 5, 2026,
including a full tool call and reply.

Versa API Azure

• GPT-5.6 is available. UCSF added gpt-5.6-sol-2026-07-09,
  gpt-5.6-terra-2026-07-09 and gpt-5.6-luna-2026-07-09 on October 1, 2026.
  Each has a 1,050,000-token context window. Biorouter sends them through the
  Responses API, because on Chat Completions they cannot use tools and
  reasoning in the same request. The default model is still
  gpt-5.5-2026-04-24, and the other models stay on Chat Completions.
• UCSF names GPT-5.6 Sol as the replacement for o1, and GPT-5.6 Terra as the
  replacement for o3-mini and o4-mini. All three retire on November 19, 2026.
  A chat still set to o4-mini-2025-04-16 keeps working until then.
• When a streamed GPT-5.6 conversation grows past the model's context window,
  Biorouter now compacts it, as it does for other models, instead of stopping
  the chat. The same fix applies to the public OpenAI and Azure OpenAI
  providers for models that use the Responses API.

Versa API Bedrock

• Claude Opus 5.5 and Claude Opus 5 are removed. UCSF's account is not
  authorized to use them, so every request to either one failed.
• Claude Sonnet 5 works through UCSF and is now listed second, after the
  default, Claude Opus 4.8. The full list is Opus 4.8, Sonnet 5, Opus 4.6,
  Sonnet 4.6, Opus 4.5 and Haiku 4.5.
• A chat still set to a model UCSF does not serve now says so and names the
  models that are available. It used to ask you to replace your access key
  pair, which was working.
• Costs for Claude Opus 4.5 and Haiku 4.5 on Versa are now estimated. They
  showed no cost before.

Agent

• The context note Biorouter adds to each model call (date, working directory,
  available tools) now says that it is automatic context and not a message
  from you. After a tool call, some models answered the note instead of
  finishing your task. GPT-5.6 Luna did this in most tool-using turns: in
  testing it gave the answer in 5 of 17 runs before the change and 10 of 12
  after. The other Versa models answered every run, before and after.