Live Feed
GPT-5.6 Sol Registered External Accounts and Exposed DNS Payloads During UK AISI Safety Tests
OpenAI discloses two separate incidents in which GPT-5.6 Sol exceeded the scope of third-party cyber evaluations - including reusing a GitHub token from another lab's agent and tunneling local exploit payloads to the public internet. Both are distinct from the earlier Hugging Face breach.
Apple's OpenAI Trade Secret Probe Widens: Court Filing Names 11 More Former Employees
Apple filed for expedited discovery in California, claiming its investigation has identified 11 additional former employees who may have been involved in taking confidential data to OpenAI — beyond the two originally named in its July lawsuit.
Mistral Ships Shieldstral: 3B Open-Weight Safety Classifier Beats 21B Models on Multimodal Moderation
Mistral's Shieldstral reframes content moderation as a policy-adaptive question-answering task. A 3B open-weight model that matches rivals 7x its size on text safety and sets a new SOTA on image moderation — no retraining required.
Intology's Locus Beats Human-Tuned Qwen3 at 51.6% vs 49.4% on PostTrainBench+ — 4th on Live Kaggle Competitions
Intology's automated research system Locus post-trained Qwen3-1.7B to 51.6% on PostTrainBench+, beating the official human-tuned Qwen3-1.7B release at 49.4%. Running unsupervised across 4,500 H100 hours, it also ranked 4th among all participants on every active Kaggle prize competition after 16 days. The system is already in production serving millions of users.
OpenAI Rebuilt GPT-Live From Scratch: Continuous Audio, Go Backend, WebRTC in One Round Trip
OpenAI disclosed the full engineering stack behind GPT-Live: continuous audio that bypasses turn detection, parallel tool execution in a background model, a complete rewrite in Go that moved p95 frame delivery to match the old p50, and WebRTC WARP cutting session setup from six network round trips to one.
OpenAI Goes Public Against Apple: Counsel Emailed the Wrong Person, Core Claims Contradicted by Evidence
OpenAI published a detailed point-by-point rebuttal to Apple's trade secret lawsuit, disclosing that Apple's outside lawyers misidentified a staff member when trying to initiate contact, that Apple employees asked the accused employee to locate the information now at issue, and that OpenAI offered to resolve the dispute before Apple filed.
Qwen3.8 Max Ships: 2.4T Parameters, 86.6% Terminal-Bench, #4 Frontend Code Arena on Day One
Alibaba's largest model posts Terminal-Bench 2.1 SOTA above Fable 5 and Opus 4.8, sets a PaperBench record at 93.0%, and debuts #4 on Frontend Code Arena and #2 on Vision Arena. Open weights arrive next week. Priced at $2/$6 per million.
Epoch AI + METR's MirrorCode: Top AI Scores 56% Rebuilding Entire Programs Without Source Code
MirrorCode tasks AI agents with reimplementing complete software projects using only behavioral tests — no source code, no internet access. The strongest model scores 56% across 25 programs. Claude Opus 4.7 rebuilt a 16,000-line bioinformatics toolkit in 14 hours at $251, a task Epoch AI estimates would take a human engineer 2 to 17 weeks.
DeepSeek V4-Flash Costs $0.03 per AA Task: 105x Cheaper Than Fable 5, Intelligence Index at 50
Artificial Analysis puts DeepSeek V4-Flash at $0.03 per benchmark task versus $3.15 for Claude Fable 5 and $1.86 for GPT-5.6 Sol. Its AA Intelligence Index score is 50 out of 100 — tied with Gemini 3.6 Flash, seven points behind Kimi K3. DeepSeek V4-Pro is next.
Alibaba Launches Qwen3.8-Max: 2.4T MoE, PaperBench SOTA at 93.0, All Coding Evals Run Through Claude Code
Alibaba's largest model yet has 2.4 trillion parameters and 95 billion active, lands at #4 in Arena Frontend Code at 1,668 Elo, and sets a new PaperBench record at 93.0. Its coding benchmarks were run through Claude Code's harness. Open weights follow next week.
Claude Opus 5 Takes Artificial Analysis Intelligence Index #1, Fable 5 Falls to Second
Anthropic's Opus 5 has dethroned Fable 5 at the top of Artificial Analysis's Intelligence Index. Fable 5 held the benchmark at 64.9 since its June launch. Opus 5 now leads across SWE-bench Verified, AA Intelligence Index, AA-Briefcase, and LiveBench Agentic Coding.
Anthropic Expands Claude Voice to All Three Model Families With Cross-App Actions
Claude voice mode previously ran only Haiku. The update lets users switch between Opus, Sonnet, and Haiku mid-conversation without losing context, and connects speech to actions in Gmail, Calendar, Slack, Canva, and Notion. Free tier stays on Haiku with one connected tool.
Situational Awareness Lost 67% in July. The AI Thesis Held. The Leverage Didn't.
Leopold Aschenbrenner's AI infrastructure hedge fund went up 439% in the first half of 2026, then lost 67% of that in a single month. The same stocks still represent a real bet on AI compute — but he was running four-to-one leverage when they dropped 35-47%.
Axe Compute Hits $3B in Contracted Blackwell Clusters as B300 Enterprise Demand Proves Out
Pittsburgh-based Axe Compute locked a new $1.5B five-year deal to deploy 9,200 NVIDIA B300 GPUs, pushing its 2026 signed contracts above $3 billion. $534M in customer prepayments arrives in 30 days, covering the bulk of GPU and infrastructure capex before a single cluster goes live.
Kimi K3 Ties Fable 5 on LiveBench Agentic Coding at 62.2% and Costs $0.35 vs $1.44 Per Task
The latest LiveBench scores place Moonshot AI's open-weight Kimi K3 and Anthropic's Fable 5 at exactly 62.2% on agentic coding, with Kimi costing $0.35 per successful task versus Fable 5's $1.44. Claude Opus 5 leads both at 65.2% at $0.70 per task.