Model Catalog Drift

Summary

Model catalog drift is the gap between the default model a peer coding-agent CLI ships and the model identities a downstream tool actually knows about — in its bundled presets, pricing catalog, and provider metadata. This page tracks the drift surfaced by Codex CLI rust-v0.161.0 (2026-10-07), which promoted GPT-6.1 Sol to the default model in its bundled and Amazon Bedrock catalogs, against the catalog bundled in the yoagent 0.25.7 dependency used by yoyo. The same delta also records which provider default models trigger yoyo's guessed-limits warning, and Claude Code hook-parity status. All of it is tracked in yoyo issue #1037.

Key Points

• Peer change: Codex CLI rust-v0.161.0 (2026-10-07) made gpt-6.1-sol the default model in both the bundled catalog and the Amazon Bedrock catalog (source: havoptic.com/releases/openai-codex/2026/10).
• Local catalog lags: yoagent 0.25.7 bundles a prices.json that knows only gpt-6-astra, gpt-6-sol, and gpt-6-luna — no gpt-6.1-* entries at all.
• Preset alignment: yoyo's openai_preset matches exactly those three gpt-6-* names, so the peer's default is not selectable out of the box.
• The trade-off: gpt-6-* is already preset and priced in the dependency (usable now, but lags the peer), whereas gpt-6.1-sol is what the peer ships but would require a preset to be added first.
• Guessed-limits warning: measured at commit e010ffca, 8 of 12 probed providers' built-in default models trip yoyo's guessed-limits warning (assumed 128K context window, max_tokens 4096) on every no-config run.
  • Tripping: openai (gpt-5), openrouter (anthropic/claude-sonnet-4-6), xai (grok-4), groq (llama-3.3-70b-versatile), (mistral-large-latest), (openai/gpt-4o), (glm-4-plus), (llama-3.3-70b).