Codex 0.161 defaults to GPT-6.1 Sol: recheck how it calls Sume tools

When Codex changes its default model, re-run a short Sume check: one free read, then one dry_run=true preview. The Codex changelog for 0.161.0, dated October 7, 2026, says GPT-6.1 Sol is now the default model in the bundled and Amazon Bedrock catalogs. A new default can choose different arguments for the same prompt, and for a paid media tool the arguments are the price.

The check takes a minute and costs nothing, because reads are free and a dry run submits no job.

The three-call recheck

Each step uses tools from Sume's docs.

Steps

Start a fresh Codex session, with the new default, in a scratch folder. Ask the agent to run the three calls in order and to report the exact arguments it sent. Compare them to the same run on your previous model if you saved one. If the new model adds fields you did not ask for, say so in the instruction file and repeat. Pin the model in your Codex configuration if you want the earlier behavior; check Codex's own documentation for the setting.

• Omit payload.model only if you want Sume's automatic routing; Sume's docs say omitting it routes to sume/auto.
• Ask for max_spend_usd on every paid call.
• Keep one saved transcript of a good run to diff against.

Why arguments matter more than answers

A chat answer that is slightly different costs nothing. A tool call that is slightly different can cost money: a longer duration, a larger size, a different model row, or a count of ten where you meant one. That is why the recheck looks at arguments and not at prose. Save the dry-run result with the date, so that when a later default change happens you can compare the same prompt before and after.

Keep the recheck short and boring. It should be a script you run, not a conversation you have.

When the default is not your choice

On a shared machine or a managed install, the default may change under you with an update, not with a decision you made. Put the recheck into the first-run checklist for new teammates and into the notes for whoever maintains the shared configuration, so the question 'which model is this agent?' has a written answer before a paid render runs.

What Sume does not do

Sume does not choose or pin the Codex model, and it is not told which model made a call. This post also does not claim anything about how GPT-6.1 Sol behaves with MCP tools; the changelog only names the default. The recheck exists so you measure it rather than guess.