Gemini 4 Argon Is Half Price Until Nov 30

News · No. 36 · Tue Oct 20 Switching Week, Not Skills Week Google's top model is half price for six weeks, Meta joined the frontier price list, and OpenAI set a hard API end date for GPT-4o.

Two frontier models opened up last week and a familiar one got an end date. No new skills needed, just a decision about which model you run.

In This Email » Gemini 4 Argon at half price until Nov 30, and the three settings to change before you test it » GPT-4o and GPT-4.1 leave the API on Jan 14, plus Meta's Muse Wonder and DeepSeek's free search window » Steal This: a 20-minute model audit you can run Monday

The Big Thing: Google's Top Model Is Half Price

Gemini 4 Argon Is Open to Every Paid Developer

Google made Gemini 4 Argon broadly available on Oct 13: the paid Gemini API and AI Studio for developers, plus the Gemini app and AI Mode for Google AI Pro and Ultra subscribers.

The Price, Defined: models bill per token, a small chunk of text. Argon costs $2 per million input tokens and $10 per million output tokens through Nov 30, then $4 and $20 from Dec 1. It isn't on the Free tier, so you need a paid API key.

Three Migration Notes From the Docs: explicit context caching (storing a long prompt so repeat calls cost less) isn't supported. Temperature, top_p and top_k are ignored. Google recommends max_output_tokens of at least 32,768 and HTTP timeouts of at least 15 minutes. All three are in What's New in Gemini 4 Argon.

My Take: test your real workload now while it costs half, but budget at $4 and $20, because that's the price you'll live with.

Worth Your Time

GPT-4o and GPT-4.1 Leave the API on Jan 14

OpenAI's deprecations page lists gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14 for shutdown on Jan 14, 2027. The named replacement is gpt-6-luna; fine-tunes follow their base model's date.

The same day, per the Codex changelog, GPT-5.5 left Codex for ChatGPT sign-ins. API-key sessions aren't affected, and OpenAI points to gpt-6.1-sol.

Monday Use: find those strings now so the switch happens on your schedule, not as a 2am outage. The audit below takes 20 minutes.

Meta Is Now a Frontier API Option

Meta launched Muse Wonder (muse-wonder-1.0, in preview) on its Model API and in Muse Code on Oct 12. It costs $4 input, $0.20 cached input and $20 output per million tokens (pricing), with a 1M-token context window, meaning how much text it can read at once.

On Oct 15 Meta also released the Muse Spark Open weights, the downloadable model files, under a license with limited commercial terms. Self-hosting needs at least 700GB of GPU memory, so most people will use a host like Together or OpenRouter.

The Trap: Contributor mode, on by default in Muse Code subscriptions, lets Meta train on your prompts and code. Turn it off before pasting client code.

DeepSeek's Web Search Is Free in Claude Code and Codex Until Oct 31

On Oct 17 DeepSeek upgraded V4.1 Flash to V4.1-Flash-1017, with MIT-licensed weights on Hugging Face, and opened its Search & Fetch APIs at $0.75 per 1,000 calls per endpoint. Calls through server-side tools in DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users until billing starts Oct 31 at 12:00pm ET.

Monday Use: if your agents pay for web search, trial it free for two weeks. For scale, Cloudflare lists Exa at $7.00 per 1,000 requests.

Quick Hits

Cursor Pro Now Comes in a $30 Bundle. The Premium tier of xAI's X Subscription includes SuperGrok, Cursor Pro and X Premium+, which xAI says cost $90 a month combined. Your Claude and GPT usage in Cursor still bills separately at API list prices, so check the plan details before your renewal date.

Claude Code Review Is Out of Preview. As of Oct 15, Code Review is generally available for Claude Team and Enterprise. Anthropic says most reviews now finish in about 6-10 minutes, down from about 20. Admins pick Quick, Standard or Deep per repo. Set Quick on routine repos and a monthly spend limit first (docs).

Steal This

The 20-Minute Model Audit

1. In each repo, run grep -rnE "gpt-4o|gpt-4\.1" ., then check your automation tools and saved Codex settings by hand.
2. Log every hit in a sheet: where it lives, what it does, who notices if it breaks. That last column is your priority order.
3. Pull 20 real prompts from the job at the top of the list. Run them on gpt-6-luna, OpenAI's named replacement, and on Gemini 4 Argon while it's half price.
4. For Argon, set max_output_tokens to 32,768, raise your timeout to 15 minutes and delete temperature settings, or the test undersells it.
5. Score outputs side by side and price the winner at its Dec 1 rate. Switch your top job first, well before Jan 14.

Agents that search the web? Add DeepSeek's Search & Fetch to the test before Oct 31, while it's free.

Skip It

Microsoft's Maia 300 Chip, for Now Microsoft expects 40% more tokens per megawatt than leading GPU systems, but it arrives in 2027 for Microsoft's own services. Nobody can rent it Monday.

Step-by-step lessons for tools like these live in the free AI Academy.

Join the free AI Academy →

Free, no card. The lessons behind Friday's tools live here.

---

That's the Debrief.

See you Friday for the tools.

From Drew and The AI Debrief team

---

P.S. Run the audit and tell us how many GPT-4o strings turned up. Hit reply, we read every one.

P.P.S. New here, or skipped one? Every past edition is in the archive.