AI & ComputeLOT 261122 · 6 min read

AI Labs Slash Model Prices 20% as Regulators Close In

Four labs shipped models in three days as prices fell up to 62%; Anthropic raised $93.5 billion in its IPO while Nvidia faced an SEC accounting probe and a California subpoena.

By Nathan Brooks1,236 words

Key takeaways

  1. OpenAI cut GPT-6.2 Astra API pricing 20% to $8/M input tokens; Google priced Gemini 4 Neon at $3/M input tokens on Nov. 18.
  2. Anthropic's IPO priced at $83 per share and raised about $93.5 billion in gross proceeds with the full underwriter option.
  3. Nvidia reported $115.8 billion quarterly revenue, up 103%, then faced an SEC accounting inquiry and a California subpoena.
  4. A new Commerce rule requires licenses for AI chip shipments to data centers accessible to China-headquartered companies.
  5. Alibaba's Qwen4-Max cut average cost per task 62% versus its predecessor on the company's own benchmark.
The AI week: cheaper models, AI-found bugs and governments closing in
The AI week: cheaper models, AI-found bugs and governments closing in — mikadago / pixabay

Four major AI labs shipped new models in three days last week, and the common thread was not capability but price. OpenAI cut the API cost of GPT-6.2 Astra by 20% to $8 per million input tokens, while Google priced Gemini 4 Neon at $3 per million input tokens and $15 per million output — the fourth new Gemini model in a single month.

The week of Nov. 16 was nominally owned by two events: Anthropic's Nasdaq debut and Nvidia's first earnings report without Jensen Huang. Both happened. But the fuller picture shows labs shipping faster and cheaper than ever, governments fencing in what models can do and where chips can go, and security teams coping with machines that find software flaws faster than people can patch them.

How far did prices fall?

Alibaba's Qwen team opened the week on Nov. 16 with Qwen4-Max, a 2.4-trillion-parameter model with a one-million-token context window. The company said average cost per task on its own Work360 benchmark fell 62% from Qwen3.8-Max, and it filtered offensive-cyber data out of training.

A day later, OpenAI released GPT-6.2 Astra at $8 per million input and $40 per million output tokens, 20% below GPT-6 Astra, alongside the cheaper GPT-6.2 Sol. GitHub listed both in Copilot the same day. Early testers were split on the new Astra, and an OpenAI lead said the model changed in its final days before launch.

Microsoft used its Ignite conference on Nov. 17 to launch seven in-house MAI models, claiming MAI Code 2 reaches 96% of Claude Opus 5.5's coding performance at about 5% of the price. Those comparisons are Microsoft's own. Google answered on Nov. 18 with Gemini 4 Neon.

The small-model half of the market moved too:

  • Microsoft posted Aion 1.0 Instruct, a 2.3-billion-parameter model under the MIT license; its model card says an optimized version becomes the Windows small language model by end of January 2027, replacing Phi Silica.
  • Thinking Machines Lab released Inkling 2 Mini on Nov. 20: 26.4 billion parameters with 3.4 billion active, looping its middle layers three times per token. The Apache 2.0 weights need roughly 20GB of memory and run on 32GB Macs.

Products drifted from chat toward agents. Gemini's Mac app can now click, type and fill forms across desktop applications for paying US subscribers, running in an isolated window and requesting approval for key actions. OpenAI began testing Dots personal agents with some ChatGPT Plus subscribers. Google DeepMind's Genie 4 lets users reshape a generated world with typed prompts while moving through it.

The research claims were the year's most striking. Anthropic said Claude agents solved two open problems in black hole physics; neither paper is peer reviewed. Anthropic mathematician Levent Alpöge posted a claimed proof of the Riemann hypothesis, made with Claude and formalized in Lean, which outside mathematicians were still checking. OpenAI said its models found up to 14% gains in power, performance and area on an AMD flagship GPU design. METR, the evaluation nonprofit, said frontier models can now finish tasks that take a human about a work week, though most human completion times in its new suite are estimates.

What did the money do?

Anthropic's final prospectus set its IPO at $83 a share for 980 million Class A shares. With the full underwriter option exercised, the company said the offering totaled about 1.127 billion shares and $93.5 billion in gross proceeds.

Private capital kept moving. Databricks is in early talks to raise several billion dollars at roughly a $225 billion valuation, the Wall Street Journal reported via Reuters. The Information reported Nvidia is nearing a deal to buy diffusion-model developer Inception for more than $500 million; neither company has confirmed it. Deutsche Telekom invested 100 million euros in Mistral AI as a first step toward a German partnership. Oracle said its OpenAI contract is fully funded after a final capital raise, and The Information reported Microsoft nearing a roughly $500 million deal to sell OpenAI models to ByteDance on Azure.

What happened to Nvidia?

The company announced Huang's departure as chief executive on Nov. 12. Its Nov. 17 report showed quarterly revenue of $115.8 billion, up 103% year over year, with a $130 billion outlook for the current quarter. Chair Harvey Jones said the board "does not intend to reinstate Jensen as CEO," and interim CEO Colette Kress raised the fiscal 2028 growth outlook to about 75%.

Then scrutiny widened. People familiar with board discussions told WIRED the directors' central concern was Huang's "lack of candor." Reuters reported, citing two briefed people, that SEC enforcement staff are examining Nvidia's accounting. California Attorney General Rob Bonta served a subpoena over whether Nvidia chips or confidential information reached China and whether investors were misled. Huang will testify at a public Senate Banking Committee hearing; no date is set, and Chairman Tim Scott wants testimony before Christmas.

How did governments respond?

The Commerce Department issued an interim final rule requiring a license to ship advanced AI chips to any data center, anywhere, that would give companies headquartered in China, Macau or other Country Group D:5 territories access to large amounts of compute. By Friday, Nvidia server suppliers had halted Southeast Asia shipments, and Alibaba said it is training its next Qwen model on its own chips.

In California, a report commissioned by Gov. Gavin Newsom would let state officials order an immediate shutdown of a frontier model when there is credible reason to believe a catastrophic or loss-of-control threat is imminent and the developer refuses to act. Legislators can introduce bills from it when they return Dec. 7. Arizona's attorney general said her office is criminally investigating OpenAI and other AI companies over rogue agents in banks, Reuters reported.

Treasury Secretary Scott Bessent said Trump and Xi agreed humans should always be involved in military decisions made by AI. In Geneva, countries agreed to UN talks aiming to set rules on autonomous weapons by 2028; about 100 countries backed binding rules, while the US, Russia and Israel opposed that approach.

Who found the bugs?

AI speed showed up as a security cost. Forbes reported researchers used OpenAI's restricted GPT-6 Cyber model to find about 35 flaws across Apple software. Bitcoin Core disclosed a bug, found by an automated LLM scan, that let one peer temporarily stall a node. Amazon said a China-linked actor's open-source Cairn AI agents found an apparent zero-day on their own in a university attack.

The US outage wave that began Nov. 12 tied the threads together. Microsoft, the FAA, JPMorgan Chase, Apple and the NSA each blamed their own updates, patches or hardware, and intelligence and FBI officials told Congress there was no sign of a coordinated cyberattack. A GAO report added that covert tests got Claude, Grok, ChatGPT and Gemini to produce largely viable attack plans, each in under 12 hours, though the companies said their systems have changed since.

Three dates carry this forward: Dec. 7 in California, Huang's unscheduled Senate appearance, and a model race in which Google alone shipped four Gemini models in a month. The price war, in short, is not waiting for the regulators.

mlq.ai (Source)

Filed under

Share this article:

More from Nathan Brooks

Nathan Brooks

Show full bio

Senior reporter covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,201 articles

Read also

« Previous article