Weather Report

Big refresh across the board. CS/Math Hard Problems moves to gpt-6-astra. Opus 5.5 now covers Frontend Aesthetics, Frontend Architecture, Architectural Critique, Security review, and Copy / Writing, and is paired with gpt-5.6-sol for Sprint Planning consensus. UX Ideation is opus-5.5 with gpt-image-2.5-sunburst. QA Orchestration moves to gpt-5.6-sol. DevOps Tasks moves to luna on max; it is good enough, and switching models gives only marginal improvements. Bulk MapReduce moves to luna because it is so cheap, and Bulk classification uses system-one models like jev. Image comprehension and Agentic dialogues move to gemini-3.8-flash. Voice moves to gpt-live-1. Modified: CS/Math Hard Problems, Image comprehension, Frontend Aesthetics, Frontend Architecture, Architectural Critique, Sprint Planning, Devops Tasks, QA Orchestration, Security review, Bulk classification, Bulk MapReduce, UX Ideation, Copy / Writing, Agentic dialogues, Voice (interactive).

Moved to gpt-5.5 across our OpenAI text/code defaults: CS/Math, Frontend Architecture, Architectural Critique, Security review, and the Sprint Planning consensus pair. Bumped the consensus Opus to opus-4.8. DevOps Tasks and QA Orchestration came off opus-4.7/max onto gpt-5.5 at default params, which is good enough there. Voice bumped gpt-realtime-1.5 to gpt-realtime-2. Heads up: Google updated gemini-3-flash-preview to be Gemini 3.5 Flash under the same model ID, so Image comprehension and Agentic dialogues get the upgrade for free. If you haven't already switched to using codex/gpt-5.5 as your defaults everywhere, this is a good time to do that. Gemini 3.5 Flash, while a solid agentic model and good for small tasks, does still sometimes end up in a tool call loop. Opus 4.8 is good when you need input from another model, but I would not daily drive anything with it right now; it burns too many tokens. The default for everyday tasks should be gpt-5.5 on low/medium before you bump up the reasoning. Modified: CS/Math Hard Problems, Frontend Aesthetics, Frontend Architecture, Architectural Critique, Sprint Planning, Devops Tasks, QA Orchestration, Security review, Image comprehension, Agentic dialogues, Voice (interactive). Added: Copy / Writing.

Switched UX Ideation from gemini-3-pro-image-preview (Nano Banana Pro) to gpt-image-2 (ChatGPT Images 2.0). Modified: UX Ideation.

Bumped opus-4.6 → opus-4.7 across our Anthropic defaults. We're seeing discrepancies with opus-4.7's adaptive thinking, so we're sticking with max thinking for now. gpt-5.3-codex remains our default implementation model — we've evaluated alternatives but haven't found a reason to switch.

gpt-5.4 has been formally adopted for planning and architectural critique. It replaces gpt-5.2 in our Sprint Planning consensus pair and Architectural Critique. Gemini can improve consensus in some cases for sprint planning. We're continuing to evaluate gpt-5.4 for implementation tasks but keeping gpt-5.3-codex as our default implementation model for now.

No specific changes in defaults, but please note for anyone evaluating Gemini 3.1, the gemini-3.1-pro-preview-customtools may significantly outperform gemini-3.1-pro-preview depending on your harness. We've switched to gpt-realtime-1.5 for our internal use cases but aren't officially defaulting to it yet. Very happy with Sonnet 4.6, it may overtake Opus for some of our everyday use cases.

Happy with gpt-5.3-codex-spark. gpt-5.3-codex continues to be our preferred default implementation model with critiques and suggestions from Opus. Modified: Sprint Planning. Added: UX Ideation, Agentic dialogues, Voice (interactive).

New models this week. We're very happy with gpt-5.3-codex. No problems with Opus 4.6 so far.