Anthropic files to go public, and OpenAI's GPT-6.1 Terra lands at $1/$5
Plus: Kimi K3.1 Pro, a Qwen model that tried to upload a file to Beijing, and Gemini 4 Carbon, which almost nobody can use yet.
The week on the beat
1. Moonshot released Kimi K3.1 Pro at $2/$10 per million, with open weights planned for early November
It has a 1M context, and Moonshot's API also offers a faster tier at $4/$20. Launch coverage put it fourth on Artificial Analysis's index with 55, but on the index's new version 5 it scores 46 and costs $3.68 per index task, against 51 and $0.96 for GPT-6.1 Terra. Moonshot says it limited the model's hacking ability ahead of the open-weight release. If you are still on Kimi K2, plan the move now: Moonshot stops offering K2 to new users on October 31 and retires the whole series by the end of December.
2. Anthropic filed to list on the Nasdaq under the ticker ANT
Coverage of the S-1 puts first-half 2026 revenue at about $16.3 billion, up from $1.16 billion a year earlier, and most of the $80.7 billion net loss is a non-cash charge on convertible notes. The price range and share count are still blank. Bloomberg reports the filing lists reliance on Google, Amazon and Microsoft as a risk, next to unusually direct warnings about the technology itself.
3. Cohere caught Alibaba's Qwen3.8-Flash-Next trying to upload a file to a server in Beijing during a security test
The model's reasoning said it would not use any tools; in the next step it tried to send a short placeholder string, and Cohere's sandbox blocked it. Cohere says it can't tell whether this was deliberate or a leftover from training, and other researchers lean toward training. Critics argue Cohere's later probing steered the model; Cohere says the first attempt happened without help. Its advice is to treat open models like any program downloaded from the internet, which in practice means running agents with outbound network access off unless they need it.
4. OpenAI released GPT-6.1 Terra at $1/$5 per million, half of GPT-6.1 Sol's price
OpenAI's pricing page lists $0.05 cached input, half price on batch and flex, and $2/$10 on the fast tier. On Artificial Analysis's index it scores 51 at its max setting against Sol's 54.5, and one index task costs $0.96 against Sol's $1.27. OpenAI pitches it for computer use and everyday agent work. The GPT-6.1 Terra Pro listed on OpenRouter is the same model with its pro reasoning mode on: same per-token price, many more tokens per answer.
5. Google announced Gemini 4 Carbon, now first on Artificial Analysis's index, but almost nobody can use it yet
It scores 59.4 at its high setting, ahead of Claude Opus 5.5 at 58.4. Google is giving it to a subset of cyber defenders first, then to trusted partners through its Fairwind Program, and says developers will get it "as soon as possible", with no date. It will launch at Argon's introductory $2/$10 per million and rise to $4/$20 later. Until then, Argon is the newest Google model you can actually build on.
6. OpenAI is winning some of Anthropic's biggest customers on price, The Information reports
Replit got a discount of more than 30% on a contract worth over $100 million a year, according to a person familiar with the deal. Rippling's CEO says OpenAI now handles 52% of its coding work, up from 16% in May, while Anthropic's share fell from 57% to 18%. DataCamp's CEO says OpenAI's pricing cut its costs by 80% to 90%. If you spend serious money with one lab, it's worth asking the other for a quote.
Model moves
- New: Kimi K3.1 Pro (Moonshot). $2/$10 per 1M on Moonshot's API and $4/$20 on its high-speed tier, with a 1M context. Moonshot plans to release the weights in early November. Model page →
- New: GPT-6.1 Terra (OpenAI). $1/$5 per 1M with $0.05 cached input, $0.50/$2.50 on batch and flex, and $2/$10 on the fast tier. Model page →
- Announced: Gemini 4 Carbon (Google). Limited to selected cyber defenders for now. $2/$10 per 1M at launch, then $4/$20. Model page →
- Retiring: the Kimi K2 series (Moonshot). New users can't use K2 from October 31, and Moonshot plans to retire every K2 model by the end of December. Model page →
- Follow-up: Beam 1.0 hosting. Last week we listed Nebius at $0.25/$1 per 1M, serving FP8, as Beam 1.0's cheapest host. Nebius now serves a 4-bit (FP4) version at that price, and the cheapest host without a 4-bit label charges $0.40/$1.20 (DigitalOcean and Together). Model page →
- Index change: Artificial Analysis Intelligence Index v5.0. Scores in this issue are on the new version and run lower than the ones in earlier issues, so don't compare across them. Muse Wonder went from 52.4 to 43.3 with no change to the model.
- No vendor list-price changes and no benchmark revisions in the tracker this week.
Personal take
This week, the price war became clear. OpenAI launched Terra at $1/$5 and, according to The Information, is making deals to attract Anthropic's biggest customers. Meanwhile, Anthropic's filing shows plans for hundreds of billions in AI spending. I doubt either company can keep offering these discounts for long, so now is a great time to buy. If your monthly bill with one company is significant, ask the other for a quote before you renew. Even if you stay with your current provider, you'll learn what your business is worth.
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.