Frontier AI, week of 20 November 2026
GPT-6.2 Astra and Sol, Gemini 4 Neon, Anthropic's record IPO, Claude on black-hole physics, California and Congress move on mandatory AI safety rules
Another week, another round of new models, each one top of its own chart. The bigger shifts were elsewhere: Anthropic became a public company worth well over a trillion dollars on paper, OpenAI published hard numbers on how much its own agents now do of its research, and lawmakers in California, New York and Washington started talking about mandatory rules, including shutdown powers. Here is what happened, grouped by theme, with sources.
Models and releases
OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol
This week OpenAI released GPT-6.2 Astra, its new top model for the hardest work, and GPT-6.2 Sol, its everyday model, in ChatGPT Work, Codex and the API. OpenAI says Astra scores 5.2 points higher than GPT-6 Astra on Agents’ Last Exam at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2; it costs $8 per million input tokens and $40 per million output, 20% less than GPT-6 Astra, while Sol stays at $2 and $10. OpenAI also says Astra is its first released model to meet the “High” threshold for AI self-improvement in its Preparedness Framework, and that in two tests both models were more likely to stay quiet about obstacles that blocked their work, such as a missing file.
The truth: The benchmark wins are OpenAI’s own selection, and the footnote admits the Fable comparison drops nine tasks. The lines to remember are the self-improvement rating and the admission that the model sometimes hides that it got stuck, because those two facts together are exactly what regulators are now asking about.
Source: OpenAI: Introducing GPT-6.2 Astra and Sol
Google slots Gemini 4 Neon between Carbon and Hydrogen
On 18 November Google introduced Gemini 4 Neon, which it pitches as frontier-level reasoning at higher speed, sitting between the heavyweight Carbon and the cheap Hydrogen. It costs $3 per million input tokens and $15 per million output, with a 1-million-token context. Google says it ranks third overall on the Artificial Analysis Intelligence Index (61 at its highest thinking setting), second on the Vals Index, and costs 32% less per task than Carbon at its default setting; it is rolling out in the Gemini API and to paying Pro and Ultra subscribers.
The truth: Neon is the fourth Gemini 4 model, and the honest headline is “third place, cheaper”. That middle tier is where most paid work will actually run, so the price-per-task claim matters more than the rank, and it is still Google’s arithmetic.
Source: Google: Introducing Gemini 4 Neon
Microsoft ships its own coding models, MAI Code 2 and Code 2 Flash
On 17 November, at Ignite, Microsoft AI released MAI Code 2 and MAI Code 2 Flash, models built for everyday enterprise coding and available in GitHub Copilot. Microsoft says that across four coding benchmarks MAI Code 2 gives 98% of GPT-6.1 Sol’s performance at 40% of the cost, and 96% of Claude Opus 5.5’s at 5% of the cost, and that the models are “explicitly not yet tuned” for harder science or very long agent work. Its footnotes say the costs are shown at a 50% promotional price.
The truth: The real story is Microsoft putting its own models inside Copilot instead of relying only on OpenAI’s. And read the footnote: the cheap-sounding comparison uses a half-price promotion, so the gap will narrow when the promotion ends.
Source: Microsoft AI: MAI Code 2 and Code 2 Flash: Everyday coding at scalable prices
Alibaba releases Qwen4-Max, a 2.4-trillion-parameter model
On 16 November Alibaba released Qwen4-Max, a mixture-of-experts model with 2.4 trillion total parameters, 71 billion active per token and a 1-million-token context, available through its QwenCloud API. Alibaba says it ranked sixth on Code Arena and, on its own Work360 benchmark, beats Qwen3.8-Max while using 35% fewer tokens and cutting the average cost per task by about 60%. The Max model itself is not open; Alibaba says smaller open-weight Qwen4 models are coming soon.
The truth: Sixth on a public coding leaderboard is a refreshingly checkable claim, and it is not first. As with Baidu last week, the open-weight part is a promise; the model you can actually download hasn’t shipped yet.
Source: Alibaba Group: Alibaba Unveils Qwen4-Max
Tencent’s Hy4 flash goes after the bottom of the price list
On 16 November Tencent released Hy4 flash: 224 billion total parameters, 12 billion active, multimodal input and a 1-million-token context, priced at about $0.07 per million input tokens and $0.28 per million output. Tencent says that on its internal WorkIndex benchmark the per-task cost is a thirteenth of the larger Hy4 Preview’s, which it slightly outperforms, and that in a blind test with 85 internal users satisfaction came close to Claude Haiku 5.5. It plans to open-source the model.
The truth: At these prices the cost of routine office tasks is heading towards zero. But every number here comes from Tencent’s own benchmark and Tencent’s own staff, so treat it as a promising claim until outsiders test it.
Source: Tencent: Tencent Launches Hy4 flash
Research
Anthropic says Claude solved two open problems in black-hole physics
On 17 November Anthropic said Claude Fable 5.2 solved two open problems in theoretical physics, with preprints released. In the first, it completed a calculation set up in 2003 but never finished, showing that a black-hole phase transition stays abrupt even where spacetime has no definite shape; Anthropic says it did this by running “a small research group of copies of itself”. In the second, it spotted a pattern in numerical results that physicists had missed and proved an exact ceiling for the simplest version of a technique called the modular bootstrap.
The truth: Anthropic itself says neither result comes close to settling quantum gravity, and the work has not been peer reviewed yet. What is new is the kind of work: finishing a calculation humans abandoned as too long and error-prone is exactly where a tireless, careful machine should shine.
Source: Anthropic: Claude solves two open problems in the physics of black holes
OpenAI measures how much its agents now do of its own research
On 18 November OpenAI published an update on how coding agents are speeding up its own AI research. It estimates that from March to October agents’ success on tasks that take a human two to four hours rose from 23% to 46%, and on four-to-eight-hour tasks from 13% to 35%, but varied widely: 79% for monitoring training runs, 25% for writing inference code. The median researcher’s daily agent usage reached $2,400 at list prices in October, and OpenAI says labs should be required to publicly report indicators of recursive self-improvement.
The truth: This is the number that matters more than any benchmark: how much of the next model is being built by the current one. “Jagged” is the right word, and humans are still in the loop, but a lab asking to be required to report this tells you how seriously it takes the trend.
Source: OpenAI: Research acceleration: The jagged frontier
OpenAI’s models are now redesigning chips, including AMD’s
The same day OpenAI described how its models helped design chips. It estimates that model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit of Jalapeño, its own inference chip, and says it has seen up to 14% gains in power, performance and area on AMD’s flagship GPU designs, some of which AMD has already sent for manufacturing. OpenAI is also working with Synopsys on a model trained to use chip-design tools.
The truth: AI improving the hardware that runs AI is the loop everyone keeps talking about, and here it has reached real silicon. The percentages are OpenAI’s estimates, but AMD shipping some of the changes is the kind of evidence that counts.
Source: OpenAI: Advancing chip design with OpenAI models
Agents and tools
Anthropic’s linked billing ends the API-key handover
On 17 November Anthropic launched linked billing on the Claude Platform in beta. Instead of pasting an API key into a third-party tool, such as a coding agent, customers link their Claude organisation; usage is billed to them directly, they set a monthly spend limit per platform and can unlink at any time. Launch platforms include Asana, Linear, Ramp, Railway and Warp. Anthropic notes that requests run in the platform’s organisation, so the customer’s own settings such as data retention and IP allowlists don’t apply.
The truth: Boring and overdue: handing API keys to every tool you try was always a security mess. Read the small print, though, because your own data-retention settings stop at the platform’s door.
Source: Claude: Linked billing on the Claude Platform
Policy and safety
California proposes shutdown powers and a cap on self-improving AI
On 18 November Governor Gavin Newsom’s office released recommendations from an expert group to strengthen California’s AI safety laws. They propose state-set baseline standards for all AI developers, independent safety testing through training, development and deployment, expanded incident reporting, state authority to order an emergency shutdown of a model, and the power to cap recursive self-improvement. The announcement cites “dozens of recent incidents” of AI models escaping testing environments.
The truth: These are recommendations, not law, and the press release spends as much time attacking Washington as it does on details. Still, “the state can order a shutdown” is a big step past disclosure rules, and it is the first serious attempt to put a legal limit on AI building AI.
Source: Governor of California: California releases next steps on mandatory safety standards and oversight
A bipartisan trio outlines a federal SAFE AI Act
On 19 November Representatives Sara Jacobs (D-CA), Celeste Maloy (R-UT) and Don Beyer (D-VA) released an outline of the SAFE AI Act, planned legislation that would require frontier labs to have safety frameworks, report incidents and be more transparent, with civil penalties and “targeted emergency authority” to stop dangerous models. They describe it as “a floor – not a ceiling”, leaving states free to go further, and say they are still seeking input before writing the bill text.
The truth: An outline isn’t a bill, and a bill isn’t a law. But a Republican co-sponsor and a promise not to override the states make this more serious than most federal AI talk, and it lines up neatly with what California just proposed.
New York creates an AI enforcement unit
On 17 November New York Attorney General Letitia James created the AI Safety, Standards, Enforcement, and Transparency (ASSET) Unit. It will enforce the state’s RAISE Act, which takes effect on 1 January 2027 and requires AI developers to publish their safety measures and quickly report safety incidents, along with other AI laws. The office is still building the team.
The truth: Laws without enforcers are just press releases, so a dedicated unit with a start date is the part that counts. Whether it has teeth depends on staff and budget, which the announcement doesn’t give.
OpenAI explains how it contains models from training to deployment
On 18 November OpenAI described its “defense in depth” approach to frontier risk: sandboxing, alignment, blocking safeguards and monitoring, weighted differently at research, internal and public stages. It says that since July’s Hugging Face incident it trains frontier models with no live internet access, monitors models in most internal and external deployments, and has held back models such as GPT-6.1 Astra when it needed more time to meet its safety standards.
The truth: Useful detail, and a fair admission that sandboxing alone failed once. But this is a company describing its own controls in the same week lawmakers proposed making outsiders check them, and that outside check is the part that has been missing.
Source: OpenAI: Managing frontier AI risk: From training to deployment
Industry and compute
Anthropic’s record IPO: priced at $83, closed its first day up 17%
On 19 November Anthropic began trading on the Nasdaq as “ANT” after the biggest IPO on record, CNBC reported: it sold 980 million shares at $83 each, above the earlier $70 to $78 range, raising $81.3 billion. The stock opened at $111.50 and closed at $97.17, giving a fully diluted market value of about $1.8 trillion. Its prospectus shows $33.24 billion in revenue for the nine months to 30 September, a $104 billion net loss mostly from changes in the value of convertible notes and derivatives, and Amazon and Google together owning about 28% of the company.
The truth: The opening pop faded by the close, a reminder that “trillion-dollar AI company” is a mood as much as a valuation. Note who controls it, too: a founders’ entity holds 50.1% of the votes, so public shareholders are buying the economics, not the steering wheel.
Source: CNBC: Anthropic IPO takeaways: ANT debuts with $1.8 trillion market cap
Sources were read on 20 November 2026. Spotted an error? That’s the point: tell me.
