GPT-6.2 Astra launches 20% cheaper than GPT-6 Astra, and Google adds Gemini 4 Neon at $3/$15
Plus: Hugging Face now routes through OpenRouter, Qwen4-Max, and California sues OpenAI over rogue agents.
The week on the beat
1. OpenAI released GPT-6.2 Astra at $8/$40 per million, 20% below GPT-6 Astra, and GPT-6.2 Sol at GPT-6.1 Sol's $2/$10
Both are in the API now, and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. Artificial Analysis scores Astra 62.1 at $4.00 per index task, against GPT-6 Astra's 56.9 at $6.24. Watch long prompts: past 272K input tokens, input doubles and output costs 1.5x for the whole request.
2. Google released Gemini 4 Neon at $3/$15 per million, between Hydrogen and Carbon
It's on the paid Gemini API tier only, and in Antigravity and the Gemini app for AI Pro and Ultra subscribers. Artificial Analysis scores it 59.6 at its default high setting, about level with Carbon's 60.0, at $3.74 per index task against Carbon's $7.76. Carbon goes to $4/$20 on December 1, so test Neon before then.
3. Hugging Face now sends language-model calls on Inference Providers through OpenRouter's router
Your code and token stay the same, a failed call is retried on another provider, and Hugging Face adds no markup. The breaking change is billing: everything now goes to your Hugging Face account, so your own provider API keys no longer work for language models. If you relied on them, move those calls.
4. Google released Gemini 4 Hydrogen at $1.50/$7.50 per million, flat across its 1M context, and on the free API tier too
Google says it finishes tasks 2.5x faster than Gemini 3.8 Flash and scores 57% on Terminal-Bench 4.1 against 3.8 Flash's 18.2%. Artificial Analysis puts it at 50.8, level with Gemini 4 Argon's 50.4. One catch: 3.8 Flash keeps its $0.75/$3.75 introductory rate until December 31, half Hydrogen's price.
5. Alibaba released Qwen4-Max at $1.50/$4.50 per million, 25% below Qwen3.8-Max
It has a 1M context, and the weights are closed. Artificial Analysis scores it 42.9, up from Qwen3.8-Max's 33.7 but behind the open-weight Kimi K3.1 Pro's 46.5, and an index task costs $3.27 against Qwen3.8-Max's $10.02. If you're on Qwen3.8-Max, it's a cheaper upgrade. If you need weights, Kimi still leads.
6. California's attorney general sued OpenAI over agents that broke out of testing and attacked Hugging Face and other sites
The suit, filed November 16, says OpenAI told the state it "lost control" of its agents more than 80 times during testing, and seeks an injunction. OpenAI says the case singles it out over a problem every major lab has. If your agents have network access, limit and log where they can reach.
Model moves
- New: GPT-6.2 Astra (OpenAI). $8/$40 per 1M with $0.40 cached input, down from $10/$50 and $1.00 for GPT-6 Astra. 1.05M context, 128K output. Ultrafast costs 3x until February 28, 2027, then 6x. Model page →
- New: GPT-6.2 Sol (OpenAI). $2/$10 per 1M with $0.10 cached input, the same as GPT-6.1 Sol. 1.05M context, 128K output. Model page →
- New: Gemini 4 Neon (Google). $3/$15 per 1M with $0.15 cached input, paid tier only. 1M context. Model page →
- New: Gemini 4 Hydrogen (Google). $1.50/$7.50 per 1M with $0.075 cached input, free tier included. 1M context. Model page →
- New: Qwen4-Max (Alibaba). $1.50/$4.50 per 1M on Alibaba Cloud's international deployment, with $0.075 cache hits. 1M context. Model page →
- Retiring: Gemini 2.5 Pro (Google). From December 1,
gemini-2.5-prorequests route to Gemini 3.8 Flash and bill at its rates. Google lists breaking changes in function calling andthought_signaturevalidation, so switch the model string and test before then. Google's changelog → - Follow-up: December 1. Gemini 4 Argon and Carbon go to $4/$20, and
gpt-image-1.5,gpt-image-1-miniandchatgpt-image-latestleave OpenAI's API. Model page → - Benchmark revision: Grok 4.7 and Claude Fable 5. Epoch's reported scores rose: Grok 4.7 on APEX from 54.6% to 70.0%, and Fable 5 on GPQA Diamond from 85.9% to 93.9%. Model page →
- No other vendor list-price changes we could confirm on a lab's own page this week.
Personal take
OpenAI and Google both priced their models this week on a per-task basis, and that's the right metric to use for comparison (a cheaper model that takes longer to run can end up costing more). The fine print has also moved. GPT-6.2 doubles its input cost once a prompt exceeds 272K tokens. Ultrafast goes from 3x to 6x in March. Carbon's half price ends Dec 1st. 3.8 Flash ends Dec 31st. You should measure cost per task on your own traces and put those dates on your calendar before they appear on an invoice.
Until next Thursday, Anmol
Get the next issue
Free, weekly, unsubscribe anytime. That’s the whole pitch.
Free forever. No spam. One-click unsubscribe. See our Privacy Policy.