OpenAI Revealed Its Next Major Model in a Research Post

OpenAI announced its next frontier model family inside a research post about mathematics, four days after showing it to Washington.

Astra cleared ten problems that had not moved in decades. Nobody outside OpenAI can run it, test it, or tell you what it will be called. Altman demoed it to policymakers and regulators before any public audience, and Astra is expected among the first models submitted under the federal pre-release review framework the administration was finalizing that same week. Read the order again. Regulators saw the model before the market did.

Brussels arrived at the same place from the opposite direction and took five years to do it. The EU AI Act became enforceable law on Sunday, complete with fines and a complaint channel any competitor can pull.

Then there is the part of the market that answers to neither. Alibaba reopened its flagship weights days after Moonshot opened a bigger model, which tells you more about competitive pressure than it does about openness.

In today’s AI news:

• OpenAI reveals its next model family through ten math proofs

• The EU AI Act goes live after five years, and complaints are now a weapon

• Qwen3.8-Max opens its weights, and loses the comparison that counts

• Today’s top tools + quick news

News: OpenAI published ten new results in mathematics and theoretical computer science produced by Astra, an internal build of its next major model. Every problem had gone at least a decade without movement, and OpenAI estimates the tokens needed to find them would cost roughly $2,000 at Sol API rates.

Details:

• The headline result is the first explicit construction of a non-sofic group, closing a question open since Mikhail Gromov introduced soficity in 1999.

• Astra also disproved Connes’s 1980 rigidity conjecture, proved Ehrhart’s volume conjecture, and resolved Erdős problems 146, 180 and 183.

• Every result ships with a machine-checkable Lean 4 certificate on GitHub under Apache 2.0, and the repository reports a “sorry” count of zero.

• Mathematician Thomas Bloom, who runs erdosproblems.com, called the results significant, rating the constructions above May’s unit distance counterexample.

• Altman demoed Astra to policymakers and regulators in Washington before any public showing, and OpenAI has not decided whether it ships as GPT-6, GPT-5.7, or under the Astra name.

Why it matters: OpenAI has not decided whether Astra ships as GPT-6, as GPT-5.7, or under its own name, and that choice is a claim about whether this is a generational break or an increment. Astra is a new model class alongside Sol, Terra and Luna, built for long-running multi-agent work, which is exactly the capability Chief Scientist Jakub Pachocki said last summer OpenAI was chasing: systems that hold a problem for hours or days rather than seconds. Astra is expected among the first models submitted under the federal pre-release review framework the administration was finalizing that same week, which means the first audience for frontier capability was a regulator rather than a customer.

News: As of August 2, the EU AI Act is live law rather than a compliance calendar item. Getting here took five years: proposed in April 2021, in force since August 2024, phased in ever since. What changed this week is that transparency now reaches anyone shipping generated output into Europe.

Details:

• Chatbots must tell users they are not human. Deepfakes require labels. Generated or altered content must carry machine-readable marks that detection tools can read.

• This is general application, not day one. Prohibited practices have been enforceable since February 2025, and GPAI rules, governance and penalties since August 2025.

• 83 provider and 152 deployer signatories joined the Code of Practice. Signing commits you to tamper-evident signed metadata, a free public detection tool, and contractual bans on stripping marks.

• Article 85 lets any natural or legal person file a complaint with no standing or harm requirement, and authorities must take it into account. Competitors qualify. Transparency breaches carry €15M or 3% of global turnover.

• The Digital Omnibus pushed high-risk deadlines to December 2027 but left transparency alone. Existing systems get until December 2 for machine-readable marking.

Why it matters: Europe drafted this in April 2021, eighteen months before ChatGPT shipped, and spent five years reaching enforcement. That head start is now a live experiment in whether being first to regulate compounds into trust and market access or into a tax that pushes frontier work somewhere else. Watch European labs specifically, because Mistral and Aleph Alpha both signed the code and now build under obligations their American and Chinese competitors do not carry. The signal to track over the next year is who copies the framework and who deliberately routes around it, because that answer decides whether Brussels wrote the template or the cautionary tale.

News: Alibaba made Qwen3.8-Max broadly available and confirmed open weights ship this week. The 2.4T mixture-of-experts model finally has published numbers, and the comparison that matters is not Fable 5. It is Kimi K3, the other Chinese frontier model, which open-weighted on July 27.

Details:

• Qwen3.8-Max scores 86.6 on Terminal-Bench 2.1, ahead of Claude Opus 4.8 and Fable 5 at 84.6, behind GPT-5.6 Sol at 88.8. It trails Fable 5 badly on repository work, 67.7 to 80.0 on SWE-bench Pro.

• Against Kimi K3 it loses the clean comparison: K3 takes FrontierSWE 81.2 to 73.5. K3’s 88.3 on Terminal-Bench comes from its own harness, and Artificial Analysis measured 85 on a neutral one.

• The timeline is the tell. Qwen3.6-Max in April was the first Qwen flagship to ship closed-weights only, breaking an Apache 2.0 tradition. Alibaba promised open weights again on July 19, three days after K3 launched.

• Moonshot disclosed K3’s 104B active parameters and shipped a 1.56TB repository with 8-to-32 GPU serving recipes. Alibaba has disclosed no activated-parameter count, so serving cost cannot be modeled at all.

• Self-hosting is not for you. 2.4T weights at 4-bit need roughly 1.2TB of VRAM and eight H200s give 1.13TB. K3’s own serving recipes start at 8 GPUs and scale to 64, with its API at $3/$15 against Fable 5’s $10/$50.

Why it matters: Kimi K3 moved the ceiling, arriving at 2.8T against DeepSeek V4-Pro’s roughly 1.6T as the previous largest open-weight release, and landing in the tier Alibaba had closed three months earlier. Alibaba announced Qwen3.8 with an open-weight promise on July 19, three days after K3 launched and eight days before K3’s weights actually shipped. That is not a values decision from a lab already open-weighting everything below its flagship. The checkpoint still needs a multi-node datacenter, so the wide majority of people who use this model will reach it through a paid API on somebody else’s hardware, exactly as they would a closed one. Until an open flagship runs on hardware a startup actually owns, open weights describe the license, not the access.

• 🔏 Content Credentials (C2PA) - The provenance standard the EU code just made table stakes

• 🧠 Qwen3.8-Max - Alibaba’s 2.4T flagship, open weights landing next week

• 🔬 Lean 4 - The proof assistant now verifying frontier-model mathematics

• ⚙️ QM - Y Combinator’s MIT-licensed multiplayer agent harness, the one it runs the firm on

• ✨ Gemini Spark - Google’s always-on agent, now global except the EEA, UK, Switzerland and Nigeria

• Minnesota’s nudify ban took effect August 1 after a judge denied xAI’s restraining order on timing, not merits. The First Amendment question is untested, with a hearing set for August 19 and $500,000 per violation on the line.

• Google pulled Nano Banana 2 from Google Earth within a day after users generated fake crashes and craters at real coordinates. Every image carried a SynthID watermark, and it did not help.

• Apple capped bug bounty submissions with a 30-day cool-off after AI slop swamped triage. It promptly locked out a real macOS root exploit worth up to $200K.

• Snapchat cut fully AI-generated video from Spotlight recommendations and monetization, and the penalty applies even when creators disclose the AI.