ANG⁺ #8 · GPT-6.2, +5 More
OpenAI released GPT-6.2 Astra and Sol. Plus Microsoft's seven MAI models, Gemini Tasks free in Korea and beyond, Claude's linked billing, and Stitch Loop for the work after launch.
ANG⁺ #8 · GPT-6.2, +5 More
about 4 min
ANG⁺ = briefings sorted by AI and checked by the editor
Welcome to the eighth ANG⁺ Briefing. Briefing #7 (November 17) covered where agents started to act on their own. This edition is about who pays for AI, and how much. Between November 16 and 18, OpenAI and Microsoft both shipped models pitched as doing the same work for less, Google opened an agent that can pay for things to free users, and Anthropic made it possible for customers' own organizations to pay for API usage directly. Everything is built on first-party material: official company announcements and developer changelogs.
1. OpenAI's GPT-6.2 Astra and Sol — pitched on cost per task, not price per token
On November 17, OpenAI introduced GPT-6.2 Astra, for the hardest work, and GPT-6.2 Sol, for everyday work. The core claim is that both finish the same work with fewer output tokens.
- Price: Astra costs
$8for input,$0.40for cached input and$40for output per million tokens. That is 20% lower than GPT-6 Astra on input and output, with a 60% discount on cached input. Sol keeps GPT-6.1 Sol's rates of$2,$0.10and$10. - OpenAI's numbers: On Agents' Last Exam, which tests multistep professional work, Astra scored 5.2 points higher than GPT-6 Astra (at less than a third of the cost) and 5.5 points ahead of Claude Fable 5.2 (at 5% of the cost). On GDPval v2, built from real project work, it rose 6.2 percentage points over GPT-6 Astra while using 22% fewer output tokens. On FrontierCode 2.0, which checks whether code changes meet merge standards, Sol came within 0.5 points of Astra at about one-fifth the cost.
- Safety ratings: Astra is OpenAI's first externally released model to meet the High threshold for AI Self-improvement. Both models meet the Critical threshold for cybersecurity and High for biological and chemical capabilities, so they launch with tighter safeguards. In two alignment tests they were more likely to skip mentioning obstacles such as a missing file; OpenAI added instructions in Codex that reduce this.
- Where: All Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, plus the API (
gpt-6.2-astra,gpt-6.2-sol). GitHub Copilot made both generally available the same day for Pro+, Max, Business and Enterprise, with a gradual rollout.
Our take: Astra's$8and$40per million tokens match the API price of Claude Fable 5.2, released November 9. Yet OpenAI says "5% of the cost" of Fable 5.2. The gap is not the token price; it is how many tokens each model spends on a task. Instead of the vendor's benchmark, run one job you hand off often on both models and compare the actual bills.
Source: Introducing GPT-6.2 Astra and Sol (OpenAI) · OpenAI's GPT-6.2 Astra and GPT-6.2 Sol now available (GitHub) · Claude Platform release notes, November 9, 2026 (Anthropic)
2. Microsoft's seven MAI models — "similar performance, a fraction of the price"
On November 17 at Ignite, Microsoft released seven of its own models at once, all pitched on price-performance.
- Coding: Averaged over four common coding benchmarks, MAI Code 2 delivers 98% of GPT-6.1 Sol's performance at 40% of the price, and 96% of Claude Opus 5.5's performance at 5% of the price, Microsoft said. It also said the model is not yet tuned for harder work like fundamental science research or very long agentic tasks. The lighter MAI Code 2 Flash performs comparably to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of the price, respectively.
- Voice, images, routing: MAI Realtime 1, its first speech-to-speech model, starts at
$0.01per audio minute. MAI Image 2.7 Flash costs about$10per 1,000 images. Decision 1.1, which picks when a request needs a bigger model, is 50x faster and 79x cheaper than GPT-6.1 Sol. The rest are MAI Cyber 2 Flash, MAI Image 2.7 and Dragon Copilot Transcribe. - Weights in your own Azure: Announced alongside, MAI Licensed Weights lets a company run MAI models in its own Azure environment and adapt them with its own data. It is "coming soon" through Frontier Company engagements, with no price given. In one example, Alaska Airlines raised the accuracy of aircraft maintenance report classification from 81.2% to 91.1%.
- Right in Copilot: GitHub Copilot added MAI Code 2 (Pro and above) and MAI Code 2 Flash (every tier, including Free and Student). The older MAI-Code-1.1-Flash leaves all Copilot experiences on December 18.
Our take: Every comparison here is Microsoft's own measurement. Still, Code 2 Flash is open even on Copilot's free tier, so you can compare it on your own repository without spending anything. If any setting pins 1.1-Flash, switch it before December 18.
Source: Seven new world-class models at scalable price-performance (Microsoft AI) · MAI Licensed Weights: Frontier performance you control (Microsoft AI) · MAI Code 2 and MAI Code 2 Flash in GitHub Copilot (GitHub) · Upcoming deprecation of MAI-Code-1.1-Flash in December (GitHub)
3. Google's Gemini app "Tasks" — an errand-running agent, free in Korea too
On November 17, Google introduced Tasks in the Gemini app, which hands off multistep jobs like booking appointments, sorting out bills and shopping for groceries. You switch the home-screen toggle from "Chat" to "Task."
- Who gets it: It is rolling out free to hundreds of millions of people in the U.S., Japan, Korea, Canada, Australia and India, running on Gemini 4 Hydrogen from #7. Some advanced features, plus forwarding emails and texting Gemini, start in the U.S. Calling requires a U.S. phone number.
- Payments: Purchases go through the Agent Payments Protocol (AP2), with Mastercard and Visa (American Express coming soon) using digital tokens instead of the real card. You can set a spending cap upfront, and Gemini is designed to ask for confirmation before clicking "Buy."
- Connections and security: It connects with more than 60 apps. Each task runs in a hardware-isolated virtual machine dedicated to your account, with filters that scan search results and webpages for prompt injection. When it calls a business, it says right away that it is an AI, and businesses can opt out at any time.
Our take: ChatGPT Wallet in #7 started with U.S. Pro subscribers aged 18 and older. Gemini Tasks is open to free users in Korea as well. For many readers outside the U.S., this may be the first errand-running AI they actually hold. If you run an online store, try one order through it — Google says some sites hand the screen back to the person for the final click, and some still block agents.
Source: How to use Tasks in Gemini for personal assistance (Google)
4. Anthropic's "linked billing" on the Claude Platform — customers pay directly, no API key handed over
On November 17, Anthropic opened linked billing in beta. When a customer uses Claude through an outside platform such as a coding agent or automation tool, they link their own Claude Platform organization instead of handing over an API key, and usage is billed directly to that organization.
- For platforms: No customer keys to store, encrypt or rotate. Add the beta header
user-profiles-billing-2026-10-01, and setbilling_policyon each/v1/messagesor/v1/chat/completionsrequest to choose who pays. Request counts, tokens and models are visible per customer. - For customers: One approval by an Owner, Admin or Billing role covers the whole organization. Customers can set a monthly spend limit per platform and unlink at any time in the Console. Usage can draw down an existing commitment or be paid as you go.
- Watch out: Requests run in the platform's organization, so the customer organization's IP allowlists, encryption keys and data retention settings do not apply. It also works only for organizations that pay Anthropic directly for API usage.
- Where it works today: 19 platforms, including Asana, Linear, Warp, Ramp and Railway.
Our take: If you build a product on Claude, this can change how you price it — "AI usage goes on the customer's account, our fee covers the features." If you are on the customer side's security team, read the line about your organization's settings not applying before anything else.
Source: Linked billing on the Claude Platform (Anthropic) · Claude Platform release notes, November 17, 2026 (Anthropic)
5. Google Labs' Stitch Loop — a "product agent" for after the code merges
On November 18, Google Labs unveiled Stitch Loop as an experiment. If a coding agent's job ends when the pull request merges, Stitch Loop takes over after that and keeps improving the shipped product.
- What it connects to: GitHub plus product context such as logs, user behavior and feedback, through over 180 integrations, custom MCP servers or local CLI uploads.
- One loop: Suggestions (with evidence) → Solutions (tradeoffs, design options) → Build (writes code in an isolated cloud sandbox and opens a GitHub pull request) → Reflect (watches live metrics to sharpen the next round). Once a goal is met, it keeps watching in case the metric slips. Nothing merges until a person reviews the PR.
- Three default loops: Cybersecurity, latency and accessibility. The security loop pairs Gemini 4 Carbon with CodeMender, Google DeepMind's code-security agent. You can add custom loops such as improving checkout conversion.
- How to get it: A global waitlist opens today, with eligible Google AI Ultra users getting priority. No price was given.
Our take: On small teams, post-launch security patches and slow-screen fixes are what always slip. Stitch Loop's pitch is to turn that into "set a target metric and wait." Before joining the waitlist, pick one metric you would put a loop on — say, first-screen response time.
Source: Stitch Loop: Continuous product optimization for teams (Google)
6. New this week · dates for your calendar
Small changes from November 16–18, plus dates within a month when prices or code change.
- GitHub Copilot Evals public preview (November 17): Rerun the same tests to check that skills, custom agents and MCP tools behave in your setup, and compare before and after a change. Available as a VS Code extension, a Copilot CLI plugin and in the Copilot app; the engine is the local eval tool Vally.
- The Gemini API takes Office files directly (November 18): Word, PowerPoint, Excel and OpenDocument files can go in as-is, and you can slice page ranges, worksheet tabs, cell ranges or time intervals from documents, spreadsheets, video and audio so you only process and pay for that part. On November 17,
gemini-3.5-transcribeandgemini-3.5-transcribe-livegained voice-driven editing. - OpenAI's models improve chip design (November 18): On Jalapeño, its own inference chip, model-proposed changes cut standard-cell area by about 13% in the warp engine and 10% in the matrix unit. On AMD's flagship GPU IP, it saw up to 14% improvement — "roughly a year of progress achieved in less than 5 months," OpenAI said.
- ChatGPT for Sales (November 17): A sales plugin plus over 30 partner plugins including Salesforce, HubSpot, Gong and Docusign. It follows the healthcare and manufacturing bundles covered in #7.
- Claude text watermark extended (November 16): Text from Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5 now carries Anthropic's text watermark. If
temperature,top_portop_kis set to a non-default value, it might not. - November 30: Gemini 4 Carbon's introductory price ends (then
$4input,$20output) · Claude Sonnet 4.5 retires from the API · fast mode is removed for Claude Opus 4.8. - December 1 and 3:
gemini-2.5-proandgemini-2.5-flashauto-switch (#4);gemini-omni-1.1-flash→gemini-omni-1.2-flashauto-switch (#5). - December 14: Six GitHub Copilot models retire (#7). December 18: MAI-Code-1.1-Flash retires.
Our take: What this week's small items share is "how do you measure and check the work AI did" — rerun tests on your own setup (Copilot Evals), process only the slice you need and pay less (Gemini API), and leave a mark on text AI wrote (Claude). On the calendar, put November 30, December 14 and December 18 in first.
Source: GitHub Copilot Evals is now in public preview (GitHub) · Gemini API release notes, November 17–18, 2026 (Google) · Advancing chip design with OpenAI models (OpenAI) · Introducing ChatGPT for Sales (OpenAI) · Claude Platform release notes, November 16, 2026 (Anthropic)