Codeep 3.2.0 — GPT-6 Astra and Gemini 3.8 Flash

[VladoIvankovic]

Two frontier models shipped within a day of each other last week. Both are in Codeep now, on the CLI and the Mac app, on the same version number.

CLI 3.2.0 — npm install -g codeep
macOS 3.2.0 — https://codeep.dev/download/mac, or let the app update itself

---

GPT-6 Astra

OpenAI's frontier model, released 3 September. 1M context, 128K output, and the thinking control works on it.

It is not the default, and that is deliberate. It costs $10/$50 per million tokens against GPT-5.6 Sol's $5/$30 — twice the price — and OpenAI is still rolling it out organization by organization, so a key that works today may not have access yet. Sol stays the default. Astra is one /model away in the CLI, or Settings → Model on the Mac.

If you switch, know what you are switching to: a long session that cost you $2 will cost $4.

Gemini 3.8 Flash

Google's newest Flash, released 2 September, at the same price as 3.7 Flash — $0.75/$3.75, the promotional rate that runs to the end of the year. 1M context, 64K output, tuned for long-horizon agentic work.

If you are already on 3.7 Flash, there is no cost reason not to move.

---

Two things that would have failed quietly

Worth saying out loud, because neither would have produced an error you could act on:

The thinking control was gated on a gpt-5 prefix. GPT-6 is a reasoning model, so it would have arrived in the picker with /thinking hidden and nothing anywhere to explain why. Both prefixes now.

Google removed temperature, top_p and top_k in the 3.7 generation, and 3.8 is built on 3.7 per Google's own model card. Without recording that, every single call would have returned a 400.

On the cost figures

Codeep estimates what a run cost, so a wrong number is worse than no number. The Gemini rows carry the promotional rate that runs to 31 December 2026 — not the rise scheduled for 1 January 2027.

That is a lesson rather than a preference. Sonnet 5's scheduled rise was entered ahead of time, Anthropic then cancelled it, and Codeep over-reported every Sonnet run by 50% for weeks. No test can catch that: a wrong price and a right one are both just numbers. So prices go in when they are being charged, and the date goes in a comment.

One more thing

Existing model ids are never removed when new ones arrive. Dropping an id does not leave your pinned setting on the older model — it drops the lookup and lands your session on whatever that provider's default happens to be. Added beside, always.

---

Full notes: CLI · macOS

If Astra or 3.8 Flash behaves oddly for you, say so in Q&A — both are days old and neither has many miles on it yet.