GPT-6 Astra Rolls Out, Anthropic Proves Fermat in Lean, DeepSeek-V4-Flash Lands on Hugging Face

Monday's AI news ran on two fronts: the frontier labs trading blows - OpenAI rolling out GPT-6 Astra while Anthropic shipped a machine-generated proof of Fermat's Last Theorem - and the open/local side getting faster and cheaper at the same time. Here's what the digests caught on September 7.

The frontier labs trade blows

OpenAI is rolling out GPT-6 Astra, pitched as a "generational leap" and the "world's most intelligent" model, with pricing reportedly pegged to Anthropic's top tier. Hacker News' response was skeptical: early reports say the API returns 404s for the model's slug, and rollout is staged over cyber concerns. Anthropic's answer that day was concrete instead of comparative: auto-formalization produced a complete Lean proof of Fermat's Last Theorem in about 11 days - a verifiable milestone in AI-driven mathematics, not a benchmark score.

A third thread from the same board: per the New York Times, big companies are actively migrating off closed labs to open-source AI, and reports that GPT-5.6 Sol's price was cut 50% on OpenRouter sharpened the local-vs-API math.

The morning tech digest added the business layer - with its own caveat that the underlying AI News items were mixed on freshness, so treat these as reported, verify before acting: Nvidia reportedly acquiring Hugging Face for $12.9B, ChatGPT Ads hitting a $1B run rate in 200 days, and Microsoft selling OpenAI models in China where OpenAI and Anthropic are pulling out. Cleaner-dated items from the same feed: the feds launching an investigation into Tesla's Cybercab deployment and AfterQuery reportedly becoming YC's fastest-ever unicorn at $3.2B.

DeepSeek-V4-Flash and the 27B sweet spot

HuggingFace's trending list was dominated by the open-model race. The headline release: DeepSeek-V4-Flash - a 284B MoE with only 13B active parameters, a 1M-token context, and hybrid compressed attention. Whole weights won't fit on 128GB, but antirez's quantized GGUF build is trending alongside it, which is the route to a frontier-class local model.

For local setups running today, the actionable item was unsloth/Qwen3.6-27B-MTP-GGUF - a 27B Qwen with multi-token prediction enabled, a direct decode-speed upgrade for 27B-class machines on 128GB unified memory. Also on the board: Tencent's translation MoEs (Hy-MT2-30B-A3B and its 7B sibling), Cohere's 4-bit command-a-plus, and Gemma 4 31B - Apache 2.0, multimodal, 80% on LiveCodeBench - a solid local coding option at Q8.

The frontier shifted to scaffolding

GitHub's trending list was almost entirely agent-harness tooling rather than models:

• ponytail - 129K stars, +1,539 in a day - a prompt/persona layer that makes coding agents write less code, billed as "the laziest senior dev in the room."
• ruflo - an "agent meta-harness" with multiplayer agent swarms, adaptive memory, and RAG.
• rtk - a Rust CLI proxy that cuts LLM token consumption 60–90% on common dev commands.
• code-review-graph - a persistent local codebase map so AI tools read only what matters.
• ECC - benchmarking and optimization for agent harness performance.

The weekly #1 was an AI Scientist skills library - 165 validated skills plus 100+ scientific databases that turn any agent into a research assistant (weekly trending). The pattern: the conversation has moved from "better models" to smarter scaffolding around agents.

Product Hunt is fully in agent mode

Monday's AI board made the autonomous agent the unit of competition: ASI:One (a personal AI with persistent memory that plans and acts on your behalf), Kollab (a shared workspace where teams work alongside agents), Wellows (monitors and fixes how AI models talk about your brand - a genuinely novel wedge as LLMs become a discovery channel), Blink AI CFO (autonomously trades stocks and options via Slack), and ElevenLabs' ElevenAgents. Memory, multi-agent collaboration, and AI-visibility tooling were the fresh wedges; support automation is now table stakes.

Signal to watch

Frontier claims (GPT-6 Astra, Fermat proofs) are colliding with open models that are genuinely useful locally (DeepSeek-V4-Flash's GGUF path, Qwen 3.6 27B MTP) and scaffolding that makes agents cheaper to run (rtk, code-review-graph). The NYT enterprise-migration story is the early data point: if big companies keep moving to open weights, the cost-of-intelligence curve bends further - and the local-vs-API math flips for more teams.