AI news, November 18, 2026: GPT-6.2 Astra and Sol, Microsoft's seven MAI models, Google warns of autonomous attacks
Five stories from Monday and Tuesday, November 16โ17, 2026 (US time) for solo makers who build with AI. Two are new model lineups that change what you pay per task, two are security changes that shorten the time you have to patch, and one puts daily limits on Apple's cloud AI features that your apps may rely on. Figures are as the companies and outlets state them.
OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol, with cheaper Astra tokens
OpenAI introduced GPT-6.2 Astra, its new frontier model, and GPT-6.2 Sol, its everyday model, saying both reach better results on hard tasks while using fewer output tokens. They are available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and in the API as gpt-6.2-astra and gpt-6.2-sol; OpenAI's changelog notes that tool calling needs the Responses API. Astra's standard API price is $8 per million input tokens, $0.40 cached and $40 output, which OpenAI describes as a 20% cut on input and output and a 60% cut on cached input versus GPT-6 Astra, while Sol keeps GPT-6.1 Sol's $2, $0.10 and $10 rates. The Verge adds that OpenAI says Astra is better at slides, spreadsheets and documents and at coding websites, games and 3D work.
Why it matters for makers: if Codex or an API app already runs on Astra, moving to 6.2 should lower the cost of each task, and heavy prompt caching gets the biggest discount.
Sources: OpenAI, OpenAI API changelog, The Verge
Microsoft releases seven MAI models aimed at low-cost, always-on agents
At Ignite, Microsoft AI released a family of seven models that it pitches on price-performance for agents running around the clock. The lineup includes MAI Code 2, which Microsoft says reaches 98% of GPT-6.1 Sol's coding quality at about 40% of the price, the cheaper MAI Code 2 Flash, MAI Cyber 2 Flash for security work, the MAI Image 2.7 editing model, and MAI Realtime 1, a voice model starting at $0.01 per audio minute. Microsoft also says its second-generation Maia 200 accelerator is now in production running its largest models. All comparisons are Microsoft's own benchmark claims.
Why it matters for makers: if coding-agent bills are your biggest AI cost, MAI Code 2 and Code 2 Flash are worth a side-by-side test on your own repo before trusting the headline numbers.
Source: Microsoft AI
Google says attackers now run patch-to-exploit pipelines with no human in the loop
Google Threat Intelligence Group's latest AI Threat Tracker says bad actors now use autonomous LLM frameworks that watch software updates, work out which flaw a patch fixed, and build an attack for it without human involvement. Newsweek reports that researchers found an exposed server running such a system, dubbed exploit_agent, along with stolen data, and that other frameworks named CAIRN and PentAGI showed up in intrusions and scanning. GTIG's chief analyst John Hultquist said attackers are mostly still going after "known, unfixed issues", just with much greater speed and scale. Newsweek also reports cases where stolen cloud accounts were used to run AI workloads, with some victims facing more than $100,000 in charges in a single day.
Why it matters for makers: the gap between a library patch and a working exploit is shrinking, so turn on automated dependency updates and set billing alerts or caps on any cloud account that holds AI API keys.
Anthropic shortens its vulnerability disclosure window from 90 to 30 days
Anthropic's Frontier Red Team says it will publicly disclose high- and critical-severity vulnerabilities it reports to vendors and maintainers after 30 days instead of 90, and other issues after 45 days, for reports sent on or after January 1, 2027. It will also escalate to an outside coordinator after 15 days without a reply, and wait 30 days instead of 45 after a patch before publishing full details; the seven-day window for actively exploited bugs stays the same. Anthropic's reason is that models now find bugs fast and a flaw one model finds is increasingly likely to be found by another. Maintainers of critical open-source projects can apply to its free OSS Scanner service, and eligible maintainers can apply for six months of Claude Ultra at no cost.
Why it matters for makers: if you maintain an open-source package, expect less time to fix reports from AI-assisted researchers, and make sure your security contact is easy to find so escalation doesn't happen by default.
Sources: Anthropic, Help Net Security
Apple adds daily limits to Siri AI and cloud-based Apple Intelligence, with more through iCloud+
A new Apple Support page says some Siri AI and Apple Intelligence features that run on server models have daily limits per device, starting with the 27.3 release. The categories include Siri AI, photo tools such as Clean Up and Image Playground, AFM Cloud models in Shortcuts, and third-party apps that use Apple Foundation Models in Private Cloud Compute. Limits reset around midnight local time, and iCloud+ plans from 200 GB give up to 2x the standard limit, rising to up to 20x on the 12 TB plan. Apple does not publish the standard limit itself, and the iCloud+ increase is not offered in South Korea, the EU, the UK or Brazil.
Why it matters for makers: if your iOS app calls Apple's cloud foundation models, plan for requests being paused when a user hits the daily cap, with a clear message or an on-device fallback.
Sources: Apple Support, The Verge