Earnings, models and agents all landed on the same Tuesday. Nvidia reported its first quarter under an interim chief, OpenAI and Microsoft both pushed new coding models to developers, and Google handed a shopping and phone calling agent to a very large free audience.
1. Nvidia Posts a $115.8B Quarter and Leaves China Out of Its Outlook
What happened. Nvidia reported third quarter fiscal 2027 revenue of $115.8 billion, up 103% from a year earlier, with data center sales of $107.6 billion. It guided to about $130 billion for the fourth quarter and said that outlook assumes no data center compute revenue from China. The release, attached to an 8-K filed with the SEC on Nov 17, quotes Colette Kress as interim CEO and CFO and lists the planned Hugging Face acquisition as pending.
The hype check. The fourth quarter guide includes an extra week, so part of the jump is calendar, not demand, and gross margin is guided down to 71% to 72% from 74.8%. The release mentions a CEO search and an independent board committee investigation in its forward looking statements but says nothing about what that review covers.
Verdict: Supported. The numbers are in a filed exhibit, and they answer the question we flagged in our preview of Nvidia’s first report without its founder, though the leadership questions remain open.
2. OpenAI Ships GPT-6.2 Astra and Sol to Developers
What happened. OpenAI added two API models, gpt-6.2-astra and gpt-6.2-sol, with Astra aimed at hard reasoning and coding and Sol at everyday coding and agent work. Both run on the Responses API and Chat Completions, but tool calling needs the Responses API. A companion note says the two hour and eight hour background completion windows in the Responses API beta are priced 30% and 40% below standard rates for these models.
The hype check. The developer notes also change how reasoning is kept when a conversation’s earlier turns are edited and how the top level instructions field is handled, which means existing apps may need a migration pass before switching. OpenAI says the models are better at professional work and coding at lower cost, but no independent benchmark results were available at launch.
Verdict: Promising but unverified. The models are live for developers; the quality claims still need outside testing.
3. Microsoft’s MAI Code 2 Bets on Cheap, Good Enough Coding
What happened. Microsoft AI released MAI Code 2 and MAI Code 2 Flash in GitHub Copilot, with Flash also in a limited preview in Microsoft Foundry. Microsoft says MAI Code 2 reaches 98% of GPT-6.1 Sol’s score at 40% of the cost, averaged across four coding benchmarks, and that Flash matches smaller rival models at a fraction of their price.
The hype check. The scores come from Microsoft’s internal test harness at maximum reasoning effort, the costs are shown at a 50% promotional price, and no raw scores, context window or model size are given. Microsoft itself says the models are not tuned for very long agentic work or research grade problems.
Verdict: Promising but unverified. The positioning is clear, but every headline number is self reported.
4. Google Makes Gemini Tasks Free, Including Calls and Purchases
What happened. Google is rolling out Tasks in the Gemini app at no cost in the US, Japan, Korea, Canada, Australia and India, letting the assistant book appointments, review bills, call businesses and shop using connected email, calendar and payment methods. It runs on Gemini 4 Hydrogen, and each task runs in an isolated virtual machine tied to the user’s account.
The hype check. An agent that logs in to accounts and spends money is a large target for prompt injection, and the safeguards Google lists (confirmation before buying, spending caps, link filtering) are described by Google, not tested by outsiders. Some advanced features start in the US only, and the post does not say how long task data is kept.
Verdict: News. The launch is real; how well the purchase and login guardrails hold up at this scale is the open question.
5. Nvidia Pitches Nemotron for Science as an Agent Kit for Labs
What happened. At SC26, Nvidia announced Nemotron for Science, which pairs its Nemotron 3.5 Super VL open model with an open chemistry dataset and post training recipes so researchers can build science agents on their own hardware. Nvidia says an agent at Argonne National Laboratory planned and tested hypotheses from one prompt and found two new binding sites, and a KTH agent repaired 81% of broken GROMACS setups in one overnight run.
The hype check. The announcement gives no license terms or release date, and the results are hand picked case studies rather than benchmarks. Claims that an agent found new binding sites need lab validation and a paper before they mean much.
Verdict: Promising but unverified. The pitch of running models on local systems is sensible, but the evidence so far is anecdotal.
Also on our radar
- Meta’s Muse agent launched in Australia with local connectors such as Canva, Domain and Ticketek, free for most uses.
- Amazon SageMaker AI endpoints can now route requests to the instance already holding a matching prompt cache, and AWS reports up to 97% lower median time to first token in its own Llama 3.1 70B tests.
- The TeleChat team described Xing 4, a model series trained on Huawei Ascend chips, and released only its 29B mixture of experts model while keeping the larger 285B version as a case study.
What to watch
Anthropic’s IPO roadshow, which opened on Nov 10 with a range of $70 to $78 for 825 million Class A shares under the Nasdaq ticker ANT, should end in a final price soon. The number it settles on, and how Nvidia’s results shape demand for AI stocks around it, will be the clearest test yet of how public markets value a frontier lab.
Sources
- NVIDIA, NVIDIA Announces Financial Results for Third Quarter Fiscal 2027 (Exhibit 99.1)
- NVIDIA, Form 8-K, filed Nov 17, 2026
- OpenAI, API changelog
- Microsoft AI, Introducing MAI Code 2 and MAI Code 2 Flash
- Google, Put Gemini to work for you
- NVIDIA, Ian Buck Kicks Off NVIDIA’s Week at SC26 as Science Enters the Age of Agents
- Meta, Muse Is Now Available in Australia
- AWS, Amazon SageMaker AI real-time endpoints now support KV cache aware routing
- Hugging Face Papers, TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks
- Anthropic, Anthropic launches roadshow for proposed initial public offering
By Willa Ravenwire · The AI Bandwagon
