GPT-6 Astra and Gemini 3.8 Flash - Changelog

GPT-6 Astra (gpt-6-astra, rolled out from September 3) is the new recommended model on the Codex harness — OpenAI's most capable model for the hardest reasoning, coding, computer-use, and research work.
1.05M token context, 128K output tokens, image input, reasoning effort low → max (the GPT-5.6 "none" tier is gone), defaulting to high.
$10 / $50 per million tokens ($1 cache hits, $12.50 cache writes), 2.5x GPT-5.6 Sol. Long requests over 272K input tokens bill at $20 / $75.
GPT-5.6 Sol, Terra, and Luna stay selectable; Sol keeps its $4 / $20 promo through November 21. Saved chats and agents keep their model — nothing moves to the pricier tier automatically.
Requires Codex CLI 0.153.4 or newer, which now ships with Teamday's runner.

Gemini 3.8 Flash (GA September 2) is the recommended default on the Google Antigravity harness in Medium, High, and Low thinking levels — Google's most intelligent Flash for long-horizon software engineering and agents, at the same $0.75 / $3.75 promo price as 3.7 Flash.
Gemini 3.7 Flash stays available for compute-efficient work. Gemini 3.5 Flash has left Antigravity's selector; saved selections map to the same Gemini 3.8 Flash level.

GLM-5.3 Flash — Z.ai's 320B / 18B-active multimodal model on the Pi harness via OpenRouter (1M context; $0.075 / $0.25 per million tokens on a 50% launch promo).
Qwen3.8-Flash — Alibaba's efficient multimodal model on the Qwen Code harness for orgs with a Model Studio API key (1M context, image and video input; $0.15 / $0.47 per million tokens; not part of the Coding Plan).

GPT-5.6 cache writes — OpenAI bills cache writes at 1.25x uncached input (Sol $5, Terra $2.50, Luna $0.25 per million tokens). Usage estimates for GPT-5.6 now include them; the earlier "no cache-write fee" note was wrong.