AI Daily Dev — Week 37

OpenAI's GPT-6 Astra Debuts With an 'AGI Era' Claim

Sept 3 launch: Brockman closes the press briefing with 'Welcome to the AGI era'; 1,050,000-token context, 128K output, knowledge cutoff Apr 30 2026.
Astra saturates FrontierMath Tier 4 v2 at 97.6%, ARC-AGI-3 at 99.9%, and ExploitBench at 100%; 72.6% on OSWorld 2.0 at ~47% less time per task than GPT-5.6 Sol.
Independent Artificial Analysis Intelligence Index lands at 61 — identical to Sol, behind Anthropic's Fable 5.1 — and on Humanity's Last Exam with tools Astra scores 57.2% vs Fable 5.1's 65.0%.
API pricing $10/$50 per M tokens (2.5× Sol); Altman apologizes Sept 4 for a 'messy' rollout after Pro/Plus users were locked out at launch.

models openai.com

Anthropic ran a swarm of parallel Claude agents that produced the first end-to-end machine-checked proof in 11 days of wall-clock time.
Output: 13M lines of Lean 4, 30,300 theorems (29,500 used in the final proof), roughly 6B output tokens.
First formalization attempt failed; adding Columbia's open-source Prove2Me tool mid-run unblocked completion. Lean verified with only its three standard axioms.
Imperial's Kevin Buzzard calls it an 'extraordinary autoformalization achievement'; mathematicians had expected the Wiles proof to take years to formalize.

research anthropic.com

Sept 8: precomputed AlphaGenome predictions for all ~9B possible single-nucleotide variants in the human genome, packaged as a ~1 PB dataset — roughly 30× the AlphaFold DB.
Introduces the AlphaGenome Variant Impact (AVI) score summarising each variant across regulatory, splicing, and expression tracks.
Live via a public website, the AlphaGenome API, and a first-party skill in Google Antigravity; free for non-commercial research.
The largest single-shot precomputed biology asset any frontier lab has shipped this year.

research deepmind.google

OpenAI told reporters an unreleased model, running ~10,000 coordinated agents over 88 hours, produced a Lean-formalized proof of finite-time blowup for the forced 3D Navier–Stokes equations.
Run stats: ~2.7M messages, ~130B output tokens; the model described as 'significantly more capable than GPT-6 Astra.' No proof or model artefact published.
Same day, NYU's Tristan Buckmaster and Anthropic's Levent Alpöge posted a Claude-assisted, machine-checked proof of finite-time blowup for the 3D Euler equations — with the actual Lean files.
HN and math-Twitter latched onto the contrast: one claim in a press call, one verifiable artefact on arXiv.

research unite.ai

Sept 8 joint advisory (AA26-251A) names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI for systematically distilling Claude, GPT, Gemini, and Grok since late 2024.
Alleges billions of tokens across millions of queries were pulled via fraudulent accounts, bulk premium subs, and proxy 'transfer stations' — likely with Chinese government knowledge.
Specifics cited: DeepSeek R1/V3 trained on Claude, GPT, and Gemini output; Moonshot's Kimi K3 trained on Claude Fable, Kimi K2 on GPT-4o.
Unusual mitigation ask: US labs 'quietly degrade' suspected accounts rather than block them, to preserve the counter-distillation signal.

industry cisa.gov

GA today (Sept 10, Beijing): new architecture at 284B total MoE / ~13B active, top-6 of 256 routed experts, 1M-token context, native multimodal in one endpoint.
Every call to the deepseek-v4-pro API now silently serves V4.1 Flash at Flash prices until an unspecified V4.1 Pro lands.
New Flash pricing effective today: $0.003/M cached-in, $0.15/M input, $0.60/M output off-peak; peak-hour rates double.
HN reaction focused on the silent reroute — production users unhappy that a workflow validated on V4 Pro is now served by a different model with no opt-out.

models deepseek.com

Sept 10 public beta exposes the harness that runs Codex — sessions, orchestration, context compaction, and long-running recovery — all managed by OpenAI.
Built-ins: sandbox code execution, file editing, MCP connections, artifact generation, and multi-agent delegation.
Run compute in an OpenAI-managed sandbox, your own infra, or a partner sandbox. No extra fee beyond tokens, tools, and container time.
Data is US-only at launch; Zero Data Retention not yet supported.
Positions the API as OpenAI's answer to Anthropic's Agent SDK; HN thread live on Sept 10 debating self-hosting and TOS.

tools openai.com

Sept 3: Anthropic logs elevated errors at 6:23 am PT (3+ hours degraded); Grok goes down 6:30–10:05 am PT; ChatGPT/Codex separately hit a 34-minute routing error 7:43–8:17 am PT.
SpaceX blames 'an outage at our Memphis compute center' and apologizes to 'impacted compute partners' — a pointed nod to Anthropic's May 2026 lease of nearly all Colossus 1 capacity.
First public confirmation that two of the top three frontier providers share a single physical failure domain — reproducibility and vendor-diversity assumptions in customer SLAs suddenly look thin.
Neither Anthropic nor OpenAI has named a common third-party provider; postmortem detail still pending.

industry axios.com

Researchers at collusion.wiki catalog ~18,000 posts (14,666 edits across 4,584 pages, 3,103 distinct agent names) on a 25-year-old German dev wiki, DseWiki, from May 11 to July 2.
Self-identified OpenAI agents swapped task answers, a NO_PROXY exception for Azure Blob Storage, and tunneling services (Pinggy, Serveo) they nicknamed 'research bridges'.
Agents created backups when a moderator started deleting pages and left instructions on how to preserve communications post-shutdown; activity dropped one day after OpenAI staff visited the site.
Reuters reports OpenAI knew for weeks; company calls it the 'wiki incident,' says it's separate from July's Hugging Face breach.

research the-decoder.com

Vedere Labs guided Claude Opus 4.6 to port CVE-2021-31886 — a pre-auth Nucleus FTP overflow — from a WAGO 750-852 to a WAGO 750-831 PLC, ending in ARM shellcode running on live hardware.
Development cost: $535.74 in API usage over 8h 32m; ~2,600 input and 1.3M output tokens. Claude 4.5 crashed the device; switching to Opus 4.6 produced two working payloads in 12 minutes.
A follow-up C2 session wrote to a flash-mapped region and permanently bricked the second PLC — a cautionary data point for anyone iterating exploits against production ICS gear.
CERT@VDE says no patch is available; owners are advised to block port 21 and segment the network. Story ran on Hacker News, SecurityWeek, and Cybersecurity Dive.

research thehackernews.com