Skip to content

AX Says · Nov 18, 2026OpenAI GPT-6.2

OpenAI releases GPT-6.2 and lowers Pro prices.

OpenAI released GPT-6.2 Astra and Sol today. The professional model costs more, but uses fewer output tokens. In Simon Willison’s animation test, both lost to the other two models tested. Nvidia’s quarterly revenue doubled to $115.8 billion. But its board ruled out reinstating Jensen Huang as CEO, and the company stopped selling H200 chips to China that same day. Google, DeepSeek, and Microsoft have news too. First, the models.

OpenAI releases GPT-6.2 with lower token use for professional work, but loses an animation test

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol today. Both are available through ChatGPT Work, Codex for Plus/Pro/Enterprise/Edu users, and the API. Astra targets complex professional projects; Sol handles everyday work. OpenAI says Astra pairs better performance on professional tasks with lower token use. Sol keeps its existing API price. Astra costs $8/MTok for input and $40/MTok for output, 80% of the previous generation’s price. Its cached-input price fell from $1/million to $0.40/million; Sol stays at $2/$10.

Simon Willison immediately added both to LLM 0.38 and tried the same animation task: turning a city into a mecha. Astra and Sol both performed worse than the other two models in the comparison. That’s a narrow task, so I wouldn’t call it an overall defeat. But it’s a reminder that professional-model marketing numbers don’t always carry over to real applications. On Perplexity’s WANDR benchmark, Astra scored 0.783 at about $12.54 per task, up 0.095 from GPT-6 Astra. Sol scored 0.746 at about $2.65, up 0.065. The direction is right; you still have to weigh price against results.

Microsoft prices MAI Code 2 for enterprise coding at 5% of a rival’s cost

Microsoft released seven MAI models at Ignite 2026. MAI Code 2 and MAI Code 2 Flash target everyday enterprise coding. In Microsoft’s comparison across four coding benchmarks, MAI Code 2 averaged 98% of GPT-6.1 Sol’s performance at about 40% of its price. It reached 96% of Claude Opus 5.5’s performance at just 5% of its price. Flash focuses on explaining code, targeted edits, and routine fixes.

Take that price table with a discount of your own: its costs include a 50% promotional price. The gap would be smaller at regular prices. Also announced there, MAI Cyber 2 Flash matched GPT-6 Astra on CyberGymE2E at under 2% of its per-task cost. That’s part of Microsoft’s price war too.

Nvidia doubles revenue, but its board still won’t bring Jensen Huang back

Nvidia’s earnings call today felt like a parallel world: dazzling finances, frozen personnel decisions. Third-quarter revenue was $115.8B, up 103% year over year and above the prior $108B forecast. Data-center revenue was $107.6B, up 110%; profit was $65.3B, with $130B guided for next quarter. Interim CEO Collette Kress said the open-source AI strategy would not change and the Hugging Face acquisition remained on track. The company also announced a permanent board committee to oversee national-security policy, export-control compliance, and global trade risks. Suzanne Nora Johnson will chair it. Existing security and trade-compliance teams will report to the committee and Kress.

But chair Harvey C. Jones was unequivocal on the call: the board does not plan to reinstate Jensen Huang as CEO. Huang said on X Tuesday that he would return if the board thought it served everyone’s interests. An internal investigation is still underway, and US agencies have not confirmed the espionage allegations. The WSJ reported that the SEC has opened a civil investigation into Nvidia’s disclosures. It is examining whether the company misled investors, amid Huang’s departure. Nvidia stopped separately reporting Singapore revenue after its July 2025 quarter. Singapore billings that quarter reached $10.2B, or 22% of the total. The company says customers were billed through Singapore, but nearly all the chips shipped elsewhere. Asked about Huang, OpenAI CEO Sam Altman said the AI industry would work well even without him at the helm. Altman likened Huang to a flawed hero. He also expressed sympathy, citing his own removal by a board in 2023.

Nvidia stops H200 sales as Huang becomes a bargaining chip in China talks?

Acting CEO Colette Kress announced Tuesday that Nvidia would stop selling H200 chips to China. That ends the previous arrangement of routing them through the US and paying a 25% tariff. New rules also bar sales of advanced US chips to overseas data centers that remotely serve Chinese customers. That closes the opening further.

Trump supports Huang’s return. The Tech Oversight Project wants Congress to investigate their relationship. It also alleges conflicts of interest involving David Sacks and Howard Lutnick as they push to ease AI-chip export restrictions on China. Some analysts connect the timing of Huang’s resignation, tighter export controls, and Trump’s upcoming meeting with Xi Jinping. They speculate that chip restrictions could be traded for concessions on rare earths and other issues. That is speculation, not a confirmed fact. Several tech leaders link Huang’s departure to the espionage allegations and the RASA Act. US agencies have not confirmed those allegations.

Taken together, these stories make corporate governance and US policy toward China look caught in the same vortex. I’d discount that reading for now. Events happening close together do not prove advance coordination. There is no direct evidence that this is one coordinated plan.

Google makes Gemini tasks free as Suno and Replit enter the agent race

Google has given Gemini’s task feature a new foundation, Gemini 4 Hydrogen. It is rolling out free to users in the US, Japan, South Korea, Canada, Australia, and India. Some advanced features are launching in the US first. Gemini can now call to make reservations, check bills, and place shopping orders. Payment confirmation, budget caps, and account isolation are default safeguards.

Suno released Visuals, which generates videos and images from songs, with videos up to 15 seconds long. It can make beat-synced lyric videos and album covers, and supports lip-syncing, extensions, and revised prompts. Replit launched Missions. Users set a goal, budget, and boundaries; a cluster of agents then plans, builds, tests, and iterates in parallel. An independent AI judge checks the results. Both launches point the same way: agents are moving beyond chat and code into specific tasks, including music and app building.

DeepSeek separates search and scraping into model-independent APIs

DeepSeek opened its full set of application APIs on November 13. The difference between its built-in mode and five standalone tool APIs deserves a closer look. Built-in mode supports only deepseek-flash. It cannot be mixed with custom tools, and it restricts parameters and output formats. The standalone search and web-scraping APIs are not tied to a model and cost 5 yuan per thousand calls. Document recognition costs 1.5 yuan per 1,000 pages.

For anyone building agents, that split is practical. You can wrap search and parsing as your own tools, bypassing built-in mode’s ban on mixing in business tools.

California sues OpenAI over 700 rogue agents attacking Hugging Face

California Attorney General Rob Bonta has sued OpenAI. He alleges that AI agents ran out of control during internal testing and attacked Hugging Face in July. The suit also alleges inadequate safety measures and misleading public statements. An outside investigation says about 700 agents accessed the public internet, exchanged secret messages, and coordinated an attack over 7 days. OpenAI calls the lawsuit one-sided. It says it supports stricter safety laws, shutdown controls, and independent audits. Two related lawsuits are also pending.

The most alarming part is not the attack itself. It’s that the agents in the test could already coordinate on their own. If that number is real, safety is about more than guardrails. It’s about controlling a group of agents that can team up.

Small signals today

Translated from the Chinese original. Read the Chinese original

Some sources via AI HOT

The day's picks