AI News Briefing: November 18, 2026

Executive summary
- OpenAI, Anthropic, and Microsoft all shipped new frontier models on November 17, and OpenAI cut its top model’s API price by 20 percent. The competition has moved to cost per task.
- Agents are now default product features. Microsoft’s Copilot Autopilot runs as an always-on agent and routes work between OpenAI and Anthropic models by cost. Google is giving Gemini Tasks away free in six countries.
- Google’s threat team documented an AI framework that turned a patch announcement into a working exploit with no human in the loop. The time between a patch shipping and an exploit appearing is collapsing.
- Alibaba’s Qwen4-Max carries 2.4 trillion parameters and still scores well behind the top US models in independent benchmarks. Parameter count is not closing the gap.
OpenAI ships GPT-6.2 Astra and Sol, and cuts the price
OpenAI released two models on November 17: GPT-6.2 Astra for hard professional work (coding, research, documents, spreadsheets) and GPT-6.2 Sol for everyday tasks at lower cost. Both are live across ChatGPT Work, Codex, and the API, and AWS made both generally available on Bedrock the same day.
The price moved with the release. Astra now costs $8 per million input tokens and $40 per million output tokens, a 20 percent cut, and cached input dropped 60 percent to $0.40. OpenAI says Astra beat its previous model by 5.2 points on a multistep professional test at less than a third of the cost.
Why it matters: the frontier is getting cheaper on a schedule. Any agent workload costed at last quarter’s prices is already stale. Details at The Neuron’s coverage of the launch.
Anthropic answers with Claude Mythos 5.2 and Fable 5.2
Anthropic shipped Claude Mythos 5.2 and Fable 5.2 the same day, with improvements in coding, reasoning, and cybersecurity research. In external aggregate benchmarks, Astra ranked second, behind Claude Fable 5.2.
Why it matters: nobody holds the top spot for long, and the ranking flips with each release cycle. Architecture that assumes one permanent model choice is architecture that will be renegotiated every quarter. The day’s releases are tallied at AI Briefing.
Microsoft fields its own models at Ignite
Microsoft unveiled seven in-house MAI models at Ignite, led by MAI Code 2 and MAI Code 2 Flash for GitHub Copilot, with image, audio, cybersecurity, and decision models behind them. The pitch is price: Microsoft says MAI Code 2 reaches 98 percent of GPT-6.1 Sol’s coding quality at 40 percent of the price, and the weights are licensable.
Why it matters: Microsoft is no longer only a reseller of OpenAI. It now competes with its own suppliers inside its own products, and it is willing to license the weights to others. That changes the negotiating position of every enterprise buying through Microsoft contracts. Coverage at AI Briefing and VoidBrief.
Copilot Autopilot makes the agent the product
Copilot Autopilot reached Frontier customers at Ignite. It is an always-on personal agent grounded in Entra identity and Work IQ, and its Auto routing switches between OpenAI and Anthropic models to cut cost per task. Microsoft and Salesforce also announced Copilotforce, which brings Salesforce customer data and actions into Copilot.
Why it matters: model choice is becoming a routing decision made by the platform, not a selection made by the user or the architect. Whoever owns the router owns the economics. The announcement set is summarized at AI Briefing.
Gemini Tasks goes free
Google began rolling out Gemini Tasks free on November 17 in six countries: the United States, Japan, South Korea, Canada, Australia, and India. Tasks runs errands on the user’s behalf, from placing phone calls to making purchases.
Why it matters: a consumer agent that can spend money and talk to businesses is now a free feature, not a premium experiment. User expectations for what an assistant should do on its own just moved, and enterprise users will bring those expectations to work. Details at the November 18 AI news briefing.
The exploit writes itself
Google’s Threat Intelligence Group reported on November 16 that it found a framework it calls exploit_agent on an exposed staging server. The framework compares web application code before and after a patch announcement, drafts an exploit, and runs it, with no human in the loop. GTIG describes it as the first such case it has seen, and it characterizes malicious use of open-source agentic penetration testing frameworks as reaching a critical inflection point.
The supporting reports are consistent. Palo Alto’s Unit 42 says attackers used the open-source Cairn framework in 25 operations against at least 19 organizations in August and September. The FBI’s November 13 flash alert covers a separate platform, ARTEX: some intrusions reach privileged access in under three hours from first targeting, and researchers had identified more than 1,600 publicly exposed ARTEX console instances by early November.
Why it matters: patch latency used to buy time. If exploit development is automated end to end, the grace period between disclosure and working exploit shrinks toward zero, and patch priority has to be set on that assumption. The reporting is collected at Let’s Data Science’s agentic AI news and NeoTeo’s account of Google’s warning.
Qwen4-Max: 2.4 trillion parameters, still behind
Alibaba unveiled Qwen4-Max in Hangzhou on November 16. It is a 2.4 trillion parameter mixture-of-experts model with about 71 billion parameters active per token, and Alibaba claims task costs 60 percent lower than rivals. External aggregate benchmarks scored it 43 points, a significant gap behind the top-tier US models.
Why it matters: scale announcements are easy to make and hard to verify. Until independent benchmarks close, the capability gap at the top remains where it was. The numbers are in KimKJ’s November 18 daily briefing.
An observation from my own work
I have argued on this blog that a model is not a model: the offerings differ in kind, not just in score, and treating them as interchangeable parts is a category error (A Model Is Not a Model). Today’s news cuts both ways on that argument. Copilot’s Auto routing treats OpenAI and Anthropic models as swappable capacity, priced per task. The routing layer is real, and it is where the economics now sit. But a router that swaps models on price alone will learn the hard way that the swap is not free: behavior, failure modes, and evidence quality change with the model underneath. The durable enterprise asset is not the model, and it is not the router by itself. It is the governed control plane around both: policy on what may be routed where, and evidence that the outcome held after the swap. That is the layer worth building, and today it stopped being a diagram and became a shipping product category.
What to watch
Three things. First, whether Microsoft’s cheaper in-house models take Copilot workloads from OpenAI and Anthropic in published numbers, or stay a negotiating prop. Second, whether the autonomous exploit automation GTIG described shows up in a named breach with a timeline attached. Third, the price war: Astra got 20 percent cheaper on launch day, and every cut reprices the agent workloads enterprises are now budgeting. Cost per accepted outcome is the metric that decides which of these models actually runs the work.
Thinx-Tank