AI News: September 13, 2026

1. Nvidia Is in Talks to Anchor Anthropic’s IPO With up to $10 Billion

Nvidia is negotiating to put up to $10 billion into Anthropic’s initial public offering as an anchor investor, according to Reuters. Anthropic is targeting roughly $100 billion raised at a $2 trillion valuation, which would make it the largest offering on record, and the deal is expected to close before the US midterm elections in November. The circularity is the part worth watching: Anthropic runs on Nvidia GPUs and committed to $30 billion in Azure compute on Nvidia chips in 2025, so a large share of any Nvidia investment returns to Nvidia as chip orders. Nvidia is separately backing around $300 billion in data center financing guarantees while investing in the AI labs that buy its hardware. Anthropic’s revenue went from about $9 billion annualized at the end of 2025 to over $65 billion by July 2026. Source

2. Altman Said OpenAI Will Adopt Embedded Evaluators Too

OpenAI CEO Sam Altman endorsed the core of Dario Amodei’s pacing proposal, saying “I agree with Dario that we need to pace the frontier” and confirming OpenAI would take on embedded third-party evaluators. Elon Musk also backed the position. The agreement is notable because embedded evaluators with publication rights and employee-level access is a concrete, verifiable commitment rather than a statement of principle, and two of the three largest frontier labs now claim to accept it. Critics counter that a safety agenda set by the leading labs functions as regulatory capture and sidesteps AI’s present-day harms. Source

3. Sam Altman Called a 2026 OpenAI IPO “Ill-Advised”

OpenAI will not go public this year. Altman told Fortune editor-in-chief Alyson Shontell that a listing right now would be “ill-advised,” citing AI safety concerns, and that the company will go public “when we’re ready” because “we’ve got a lot of stuff to do.” OpenAI filed confidentially for an IPO in June 2026 and had hired bankers and lawyers targeting a Q3 or Q4 2026 debut, but expectations have moved to 2027 amid tech sector volatility and the fallout from its rogue agent security incident. The timing contrast with Anthropic, which is racing to list before November, is now hard to miss. Source

4. GPT-6 Astra Cleared 7 of 100 Robot Manipulation Tasks Where the Baseline Cleared Zero

Robocurve ran OpenAI’s GPT-6 Astra and AI2’s MolmoAct2 on StationeryBench, five desk manipulation tasks including uncapping markers, pouring paper clips, and passing rulers between arms, across 200 trials on dual-arm YAM robots. Astra completed 7 of 100 tasks outright with a median progress score of 46 out of 100. MolmoAct2 completed none, with a median of 12. Cornell and Google DeepMind researcher Yoav Artzi called it a “step change in spatial reasoning” and reported near-human accuracy from Astra on 3D tasks in the unpublished REMAP benchmark, while suspecting OpenAI trained on a large volume of 3D data such as Blender scenes. The absolute numbers are still low, REMAP is not published, and Astra trails humans across other scenarios. Source

5. A KAIST and Naver Study Found Reasoning Operations Have Distinguishable Activation Signatures

Researchers at KAIST and Naver AI Lab had Qwen2.5-7B, Qwen3-8B, and Gemma4-31B solve math problems step by step, segmented the solution paths, used GPT-5 to label each segment with the reasoning operation it performed, then trained classifiers on internal activations. Eight operations including extraction, decomposition, formula recall, deduction, and computation produced reliably separable activation patterns in every model tested, with results replicated on Llama-3-8B. Separability peaks in the middle layers and falls off early and late. Identical wording maps to different representations depending on context, steps build on preceding context rather than forming in isolation, and the patterns stay identifiable even when the model gets the answer wrong. That last point is what makes it relevant to oversight, since models frequently omit parts of their actual reasoning from the text they emit. Source

6. Context Compaction Destroyed the Code an Agent Used to Build Its Answer

Simon Willison asked ChatGPT Work with GPT-6 Astra (Max) to generate looping 5K and 10K running routes from his home address, and got what he asked for: two routes as embedded D3 visualizations plus downloadable GPX and GeoJSON, built by geocoding through Nominatim and pulling local roads and trails from the Overpass API. The Python that produced it was never visible in the UI, and when he asked for it afterward the thread had been compacted and the system could not retrieve it. His argument is that any system doing context compaction should preserve pre-compaction content and expose it through a tool call, because otherwise the agent’s work is unauditable by the person who requested it. Source