What an AI Summer we’ve been having!

If you looked away for a week, you missed three moves from three labs that do not agree on what winning even looks like.

DeepSeek put V4-Flash into public beta with a model that beats its own flagship and plugs directly into OpenAI's Codex tooling. Alibaba shipped Qwen 3.8-Max, 2.4 trillion parameters, and is open-sourcing it. OpenAI revealed the name of its next model family, Astra, by claiming it solved ten math problems nobody had cracked in a decade, with no release date attached.

Cheap and compatible. Massive and open. Unavailable and unmatched. Three bets, one week. Here is each one, and the part of each that is not being reported carefully.

DeepSeek's small model now beats its big one. V4-Flash-0731 outscores V4-Pro-Preview on all nine agent and coding benchmarks DeepSeek published, hitting 82.7 on Terminal Bench 2.1 against 72.1, and jumping from 7.3 to 54.4 on DeepSWE.

The detail worth your attention is that the architecture did not change. Same 284-billion-parameter mixture of experts as the preview build, identical size. DeepSeek only redid the post-training. A sevenfold jump on DeepSWE came from training method rather than scale, which is a real data point about where the remaining headroom actually sits.

The strategic move is not the benchmarks though. V4-Flash natively supports OpenAI's Responses API and is adapted for Codex, so one configuration exposes it across Codex CLI, the ChatGPT desktop app, and the VS Code extension. DeepSeek did not ask developers to leave OpenAI's tooling. It made itself the cheaper engine inside it, at reported input pricing around $0.14 per million tokens.

The caveat: every one of those benchmarks is DeepSeek's own, run on an in-house harness it has not released, and two of the test sets are proprietary and unrunnable by anyone else. The improvement over its own preview is apples to apples and real. The cross-vendor ranking is not verified.

Alibaba's flagship is 2.4 trillion parameters with 95 billion active, a one-million-token context, and native vision, priced at $2 and $6 per million tokens with caching at $0.25. Alibaba claims 93.0 on PaperBench ahead of both GPT-5.6 Sol and Claude Opus 4.8, and 86.6 on Terminal Bench 2.1.

The news is the license, not the launch. Alibaba spent most of this year keeping its strongest models proprietary, and it is now open-sourcing at Max scale for the first time, with weights for both 3.8-Max and a smaller 3.8-27B checkpoint due within days. That 27B model is the one most teams will actually run, since a 2.4-trillion-parameter checkpoint is a datacenter artifact rather than something you put on a workstation. The stock rose 7 percent in Hong Kong on the announcement.

Alibaba's own showcase runs are genuinely striking if they hold up: ten-plus days of self-evolving development from an empty repository to production, 500 turns of chip-design optimization, a year of simulated e-commerce strategy.

The caveats stack up here too. The benchmarks are Alibaba's, unverified, and one independent aggregator that scores across many sources currently places Qwen 3.8-Max 31st of 215 overall even while ranking it first on reasoning. The license terms have not been published, which is the detail that decides whether open weights means genuinely open or open with commercial conditions attached. Watch for that when the weights land.

OpenAI's release was not a release. It confirmed the name of its next model family in the third paragraph of a blog post about mathematics, with no product attached.

The claim is that an internal version of Astra produced new solutions to ten problems in mathematics and theoretical computer science that had been open for at least a decade, spanning high-dimensional geometry, group theory, quantum complexity, sphere packing, and coding theory. One construction proves that non-sofic groups exist, answering a question Gromov posed in 1999.

What separates this from a press release is that OpenAI formalized the underlying logic in the Lean theorem prover and published it, so the proofs are machine-checkable rather than something you take on trust. Most launches lead with a benchmark table. This one led with artifacts experts can inspect.

Astra is described as multiple agents coordinating over hours or days on a single problem. Altman demonstrated it to policymakers in Washington, and it is expected to be among the first models submitted for federal pre-release review. There is no release date, no pricing, and no availability, and OpenAI has not said whether it becomes GPT-6.

The caveat is the honest one: ten proofs is a real result and a narrow one. Mathematics is a domain where verification is unusually clean, which is exactly why it makes such an effective demonstration. It tells you nothing about how the model handles your codebase, and nobody outside OpenAI can check.

Anthropic responded within hours claiming Fable solved five comparable problems, so the math-proof race is now a marketing surface too.

Line them up and the strategies are almost mutually exclusive.

DeepSeek is competing on switching costs. Cutting prices is a move anyone can match. Making your model run natively inside a competitor's developer tooling removes the friction that usually protects an ecosystem, and that is much harder to answer.

Alibaba is competing on distribution. Ship at frontier scale, price at $2 and $6, give away the weights, and become the substrate other people build on. Moonshot did this with Kimi K3 two weeks ago at 2.8 trillion parameters. It is a coordinated bet across the Chinese labs that being everywhere beats being best.

OpenAI is competing on the ceiling. Astra's pitch is not that you can afford it or download it. It is that it did something nobody else has done, verifiably. That argument is aimed at governments and enterprises, not at developers comparing per-token pricing.

Worth noticing who is open here. The two models you can download are both Chinese. The American one has no release date and closed weights. Eighteen months ago that sentence would have read the other way around.

The stories moving fast beyond today's headlines:

Three frontier releases landed in seven days with three incompatible strategies. DeepSeek put V4-Flash into public beta, where the same 284-billion-parameter architecture as its preview build, re-post-trained rather than scaled, now beats the larger V4-Pro-Preview on all nine published agent benchmarks including 82.7 on Terminal Bench 2.1 against 72.1 and a jump from 7.3 to 54.4 on DeepSWE, and it natively supports OpenAI's Responses API with Codex adaptation at roughly $0.14 per million input tokens, making it a drop-in inside OpenAI's own tooling. Alibaba released Qwen 3.8-Max, a 2.4-trillion-parameter sparse mixture of experts with 95 billion active parameters, a one-million-token context and native vision at $2 and $6 per million tokens, claiming 93.0 on PaperBench ahead of GPT-5.6 Sol and Opus 4.8, with open weights for it and a smaller 3.8-27B checkpoint due within days in Alibaba's first Max-scale open-sourcing, sending the stock up 7 percent. OpenAI confirmed its next model family is called Astra inside a math research post claiming ten decade-old open problems solved, including proving non-sofic groups exist, with the logic formalized in Lean so the proofs are machine-checkable, but no release date, pricing, or availability, and Anthropic countered within hours that Fable solved five. Every benchmark here is vendor-run, Qwen's license terms are unpublished, and DeepSeek used an unreleased in-house harness with two proprietary test sets. The signal is that nobody competed on raw intelligence this week: DeepSeek chose compatibility, Alibaba chose licensing, OpenAI chose an untestable capability claim, and the two models you can actually download are both Chinese.

Stay building. 🤖