64 releases covered live on the show, led by Grok 4.6, Grok Bot, DeepSeek V4 Pro — every model, product, paper and tool that mattered, with links and our analysis.
August 2026 saw 64 AI launches covered live on ThursdAI — the weekly AI news podcast hosted by Alex Volkov that has tracked 800+ releases across 200+ episodes since 2023 — led by OpenAI, Anthropic, DeepSeek, Liquid AI across open source, agents, coding, frontier models. This page collects the products, source links, key numbers and episode coverage in one crawlable recap.
Alibaba's HappyShrimp 1.0 goes end-to-end on music generation
Yes, it's really called HappyShrimp (a shrimp-welfare meme). Alibaba's end-to-end music model generates lyrics, melody, arrangement and vocals in one pass, and unlike the Suno approach it reasons over the prompt first, mapping song structure and harmonic progression before generating audio. Early testers call it a serious and possibly cheapest Suno rival, with 320 free credits at launch. The track played on the show is extremely K-pop.
Alibaba's overnight community darling: a 27B-parameter Apache 2.0 model scoring 52 on the Artificial Analysis Intelligence Index — the same as GPT-5.6 Luna at max reasoning — and 51 on the Agentic index. It runs at ~68 tokens/sec on a 4090, ~40 on Macs via MLX, and even 11 tok/s in-browser on WebGPU kernels. The Hugging Face hub exploded with 152 fine-tunes, 650 quantizations and close to 10 million quant downloads, and Unsloth's 1-bit quants run it on 8GB of RAM at roughly 77% of BF16 quality.
52 AA Intelligence Index, tying GPT-5.6 Luna at max reasoning68 tok/s on a single RTX 4090152 fine-tunes on Hugging Face
Ling-3.0: six open base checkpoints across training stages
AntLing/InclusionAI released Ling-3.0 as six open base checkpoints, including pretrained, mid-trained and WSM-merged stages for both the tiny (7.9B total / 1.3B active) and flash (124B total / 5.1B active) sizes — a rare look inside intermediate training stages.
6 open base checkpoints across training stages124B / 5.1B flash size, total / active parameters
Cartesia Sonic-3.6 takes #1 on both TTS leaderboards
Cartesia's Sonic-3.6 is now #1 on both Artificial Analysis TTS leaderboards, with Cartesia holding the top two spots simultaneously. Same state-space-model lineage (from Albert Gu of Mamba fame) with sub-90ms time-to-first-audio, 136 characters per second versus ElevenLabs' 46.7 at half the price, across 44 languages.
#1 on both Artificial Analysis TTS leaderboards<90ms time-to-first-audio136 chars/s vs ElevenLabs' 46.7, at half the price
Liquid AI ships LFM2.5 QAD 4-bit checkpoints for edge devices
Liquid AI released quantization-aware-distilled 4-bit checkpoints for LFM2.5 models from 230M to 2.6B parameters, retaining roughly 97% of BF16 quality with 3x faster decode on edge devices.
~97% of BF16 quality at 4-bit3x faster decode on edge
MiniMax Music 3 lands with open weights (and a rough license)
MiniMax Music 3 landed right after last week's show with open weights and, per Wolfram, possibly the worst license of the year — excluding the US, Europe and the UK. Nobody cares: it's third on Hugging Face trending and the ComfyUI crowd already has it running locally.
Ornith-1.5 family: open 397B MoE matches Opus 4.8 on Terminal-Bench 2.1
The Ornith-1.5 family ships as open source, self-improving models in three sizes: 9B dense, 35B MoE and 397B MoE. The 397B flagship matches Claude Opus 4.8 on Terminal-Bench 2.1 (86.1) and DeepSWE (56).
86.1 Terminal-Bench 2.1, matching Claude Opus 4.856 DeepSWE
Superwhisper S1-mini cleans up dictation fully on-device
Superwhisper — the dictation app Karpathy made famous when he coined vibe coding — released its first open-weights model: S1-mini, a 0.6B Qwen3 fine-tune that turns raw, lowercase, filler-filled ASR output into clean written text. Apache 2.0, English-only for now, about 450MB in GGUF, and it runs entirely on-device behind Whisper or Parakeet.
Ultralytics YOLO26 removes NMS from default inference entirely
Ultralytics released YOLO26, removing non-maximum suppression from default inference entirely. It scores 40.9-57.5 mAP on COCO across sizes with up to 43% faster CPU inference.
40.9-57.5 mAP on COCO across sizes43% faster CPU inference
dots3-note preview: 280B MoE omni model with TEMPO RL for long-horizon agents
Xiaohongshu's dots studio previewed dots3-note: a 280B MoE with 16B active parameters handling text, vision and audio with 512K context, released under Apache 2.0. It uses TEMPO RL training aimed at long-horizon agent tasks.
280B / 16B total / active parameters (MoE)512K context window
GLM-5.3: post-training alone delivers a 6x Terminal-Bench jump
Z.ai announced GLM-5.3, keeping the same 743B base as GLM 5.2 but jumping from 4.6 to 28.3 on Terminal-Bench 3 and gaining almost 20% on DeepSWE from post-training alone, at unchanged pricing with a 1M context window. It scores 60 on the Artificial Analysis Intelligence Index, roughly Kimi K3 level with about a third of the parameters, and shows emergent cybersecurity capabilities: 84% on CyberGym and 54.5 on ExploitGym, beating GPT-5.6 Sol. API-only for now — weights (and license terms) expected later.
4.6→28.3 Terminal-Bench 3, a 6x jump from post-training alone60 Artificial Analysis Intelligence Index84% CyberGym
Alibaba Wan-Animate-2: 14B character animation under Apache 2.0
Alibaba's Wan team released Wan-Animate-2, a 14B-parameter character animation model under Apache 2.0. It wins over 70% of blind preference comparisons.
DeepSeek V4 Pro 0813 goes GA with MIT-licensed open weights
DeepSeek re-published its flagship V4 Pro weights under MIT license: a 1.6T-parameter MoE with 49B active parameters and a 1M-token context window, priced at $0.435/$0.87 per million tokens. DeepSWE jumps from 12.8 in the V4 preview to 62.7, with Terminal Bench 2.1 at 87.9, though it lands at 54 on the Artificial Analysis leaderboard — DeepSeek's answer to Kimi K3.
62.7 DeepSWE (+49.9 vs preview)87.9 Terminal Bench 2.11.6T/49B total/active parameters
Gemini 3.7 Flash: 300+ tok/s mid-tier multimodal with a 50% price cut
Google dropped Gemini 3.7 Flash mid-show, right as Artificial Analysis co-founder George Cameron was on air. The mid-tier model clocks over 300 tokens/sec, comes with a 50% price cut through end of year, beats Muse Spark 1.2 on DeepSWE, and lands near the Pareto frontier for cost per task — instantly #1 on Artificial Analysis' model recommender for the cost/speed/intelligence trade-off. It is also strong at multimodal, one of the few models that can watch videos.
300+ tok/s output speed50% price cut through end of year
LTX-2.5: 22B open-weights video model with multi-shot generation
Lightricks released LTX-2.5, a 22B-parameter open-weights video model with multi-shot support. It generates 10 seconds of 1080p video in 23.7 seconds on fal and needs a minimum of 16GB VRAM.
22B parameters23.7s for 10s of 1080p on fal16GB minimum VRAM
Liquid AI LFM2.5-VL-3B runs 228 tok/s on M5 Max in ~3GB
Liquid AI released LFM2.5-VL-3B, a small vision-language model that runs at 228 tok/s on an M5 Max in roughly 3GB of memory. Weights are on Hugging Face.
Meta returns to open source with Muse Glimmer 30B under Apache 2.0
Meta came back to open source AI with Muse Glimmer, a 30B agentic model under Apache 2.0 that runs on a single 24GB consumer GPU. It scores 76.0 on SWE-Bench Verified and 51 on SWE-bench Pro (beating Qwen 3.6 27B), and with DFlash speculative decoding delivers 233 tok/s on an RTX 5090. Zuckerberg also promised open weights for the bigger Muse Spark 1.2 and published an essay arguing superintelligence should be distributed to everyone.
76.0 SWE-Bench Verified51 SWE-bench Pro233 tok/s on RTX 5090 with DFlash
MiniMax-Music3: open-weights production music model
MiniMax released Music3, an open-weights production music model. It dropped in the middle of the live show as one of the week's three breaking news items.
Motif 3 from Korea: 314B MoE open-sourced under MIT
South Korea's Motif Technologies open-sourced Motif 3, a 314B-parameter MoE with 13.2B active parameters, under MIT license. It scores 76.2 on SWE-Bench Verified.
NVIDIA ships Nemotron 3.5 Lightning: 30B MoE with 3B active
NVIDIA released Nemotron 3.5 Lightning, a 30B MoE with just 3B active parameters delivering up to 4x output speed and strong voice-agent results, with weights on Hugging Face in NVFP4. CoreWeave Inference picked it up with day-zero support.
GPT-5.6-Cyber hits 95% cyber completion, gated behind Daybreak Red
OpenAI announced GPT-5.6-Cyber, a cyber-defense specialized model scoring 95.0% cyber completion versus 1.5% for the base model. Access is gated behind the Daybreak Red program as OpenAI expands Daybreak while the cyber-defense window narrows.
SpaceXAI released Grok 4.6, a big step up from Grok 4.5: 61 on the Artificial Analysis Intelligence Index at $2/$6 per million tokens, #4 on intelligence and #5 on speed while costing half as much as the models above it. It scores 61.3 on Frontier Code (just behind Opus 5) and jumps 10 points on Apex-agents to #4, and the model card confirms Cursor Bench is no longer leaked into its weights while topping that benchmark at 69.9. It is the same 1.5T-parameter v9 base at the same price, with Elon claiming Grok 4.7 lands in 3-4 weeks.
61 Artificial Analysis Intelligence Index$2/$6 per million tokens in/out69.9 CursorBench, #1 with the leak scrubbed
Wan 3.0 drops into public beta live during the show, with native 30-second generation
Alibaba's Tongyi Lab pushed Wan 3.0 into public beta minutes before ThursdAI went live: native 30-second single-shot generation and Omni-Reference conditioning on text, images, audio, and video together. Given the Wan line's open-weight track record, this is the drop the open source community most wants weights for.
30s native single-shot generation4 reference modalities via Omni-Reference
ByteDance's SeedRealtime: a native audio-visual full-duplex LLM, free on Doubao
A single end-to-end model natively fusing audio, video, and text, replacing the cascaded ASR-VLM-TTS pipelines behind most voice agents: it listens, watches, and speaks simultaneously with turn-taking inside the model and no external VAD, cutting conversational pacing failures by 50% in human evals. It ties voices to faces in noisy rooms and speaks up proactively on scene changes, live for free in the Doubao app and its 300M+ users, ByteDance's first large-scale audio-visual full-duplex deployment.
50% fewer pacing failures vs cascaded stacks300M+ Doubao users getting it free
FLUX 3 Video: BFL's first video model generates native audio in the same pass
'Two years later, our first video model': up to 20 seconds at 24fps in 720p (1080p via upscaler), with dialogue, SFX, and ambience generated natively in the same pass and lip-sync across 14+ languages. Draft mode runs ~$0.06/s for iteration versus $0.17-0.29/s full renders, and three API modes share one endpoint (t2v, keyframe-pinned i2v, v2v continuation). BFL's internal ELO has it leading text-to-video; open weights as FLUX 3 Dev are explicitly promised.
20s @ 24fps max clip, 720p native$0.06/s draft mode vs $0.17-0.29/s full14+ lip-synced languages
Bland Speech v3 tops the Audio Realism Bench, one Elo rung below actual humans
Design Arena's blind pairwise Audio Realism benchmark puts Bland Speech v3 at 1365 Elo, above ElevenLabs, Microsoft's MAI-Voice-2, and Grok TTS, second only to real human recordings around 1500. Trained on 100M+ real phone conversations, it keeps the breaths, hesitations, and fillers TTS usually sands off. Ten seconds of audio yields an instant clone at $0.015 per thousand characters. The asterisk came from Grok itself: every ranked model is a closed API; open source voice has catching up to do.
1365 Elo, second only to humans (~1500)100M+ real conversations in training10s audio needed for an instant clone
Liquid's LFM2.5-2.6B: agentic RL trained inside real harnesses, running in 1.7GB on a phone
A 2.69B-parameter hybrid model pre-trained on ~34T tokens whose post-training ran agentic RL inside real harnesses (Hermes Agent, OpenClaw, Pi), so tool calling was learned where tool calling happens. It beats Qwen3.5-9B, three times its size, on ToolSandbox and instruction following, runs 220 tok/s on an M5 Max CPU and fits in ~1.7GB at Q4 on a phone. Liquid's own model card honestly scopes it away from agentic coding and knowledge-heavy work: this is for private, on-device agents.
2.69B parameters, 128K context77.83 ToolSandbox, above Qwen3.5-9B220 tok/s on Apple M5 Max CPU
Qwen3.8-Max: Alibaba's 2.4T-parameter flagship, with open weights promised within a week
Alibaba's flagship MoE arrives via API: 2.4T total parameters, 95B active, 1M context, at $2/$6 per million tokens, with open weights plus a 27B sibling promised for the week of August 10, the first Max-class Qwen slated for release. It ranks #2 on EyeBench for vision behind only OpenAI's Sol, and the oh-my-cli demo ran 16 days of fully autonomous coding: 265 commits, 127 PRs, 151 issues, zero human intervention. Nisten's hands-on: best-in-class visual data labeling. Yam's counter: other frontier models pass those tests too, and the 27B is the one you'll run at home.
2.4T / 95B total / active parameters#2 EyeBench vision rank, behind only Sol16 days autonomous run: 265 commits, 127 PRs
Meituan's LongCat-Flash-Lite-Sparse: 1M native context at 3B active parameters, MIT licensed
The 'DoorDash releases a model' moment: 69B total parameters with ~3B active per token, native 1M-token context (up from 256K dense), MIT licensed. LongCat Sparse Attention lifts SWE-Bench Verified from 54.4 to 68.2 and SWE-Bench Multilingual by 21 points over the dense twin, building on DeepSeek Sparse Attention while removing its O(L²) scoring overhead. The paper reports the architecture scaling to 560B-A27B.
69B / 3B total / active parameters1M native context window68.2 SWE-Bench Verified, +13.8 over dense
MiniMax opens H3's weights, and the community ships LoRAs and Apple Silicon in 48 hours
H3 (Hailuo 3.0), a 33B omni-modal transformer generating up to 15 seconds at 2K with native stereo audio from unified text/image/video/audio context, landed on Hugging Face days after its announcement, per Victor Su Ortiz the first open-weight state-of-the-art omni video model. Within 48 hours the community shipped LoRA support and Apple Silicon inference, neither of which MiniMax optimized for, and X filled with recreated episodes of The Office. The panel also dug into the community license's litigation-linked restrictions on US use, the gap between downloadable and cleared.
33B open-weight omni transformer48 hrs community LoRAs + Apple Silicon support2K / 15s max resolution / clip length
Chroma launches Foundation: unified memory for your agents
Launched live during the show (founder Jeff Huber hopped on within minutes of the announcement), Foundation is a research preview of a shared memory system between you and your agents — memory as infrastructure. It ingests sources natively from Codex, Claude Code, Cursor and Slack on day one, with Notion, GitHub and Google Drive connectors coming, and is built on ChromaDB plus the Context-1 agentic search model (a GPT-OSS 20B fine-tune running ~400 tokens/sec at 25x less cost than Opus). Each Foundation manages and improves its own system prompt from natural-language feedback. Part of Chroma Cloud starting at $30/mo.
400 tok/s Context-1 agentic search speed25x cheaper than Opus for agentic search$30/mo Chroma Cloud starting price
Artificial Analysis launches Optima: private evals from your own traces
Artificial Analysis launched Optima, which builds private evaluations from your own use case and agent traces. Co-founder George Cameron discussed it on the show alongside the three-factor model-selection framework of intelligence, speed, and cost per task.
Grok Bot: always-on agent swarm where every bot gets its own computer
SpaceXAI/Cursor launched Grok Bot in early beta: persistent, always-on agents on macOS and iOS where each bot runs in its own isolated environment with its own computer, no context or model management, and Grok 4.6 under the hood with no model picker. Bots communicate agent-to-agent transparently (read-only to you), can spin up other bots with real identities, and reuse Cursor's connector and security model — API keys are hidden from bots, and payments and log-ins hand control back to you. It is free for a month in beta and included with SuperGrok Heavy and Cursor Ultra; Shub Gaur from Cursor walked through it on the show.
Cloudflare OS: Kenton Varda's Sandstorm reborn as open source agent infrastructure
Kenton Varda's 'secret 10-year master plan': a remake of Sandstorm on Workers and Durable Objects, Apache 2.0 with no open-core catch. Every app instance ('Gadget') is a sandboxed Dynamic Worker with zero default internet; 'Gatekeepers' are supercharged MCP servers holding credentials, enforcing per-resource policy, and logging every action, with pending approvals simulated locally so agents keep working while humans review in batch. Thousands of Cloudflare employees have used v1 internally since May. Landing the week of the sandbox-escape disclosures, the timing wrote its own headline.
Apache 2.0 license, no dual-licensing catchMay 2026 v1 in company-wide internal use since
Decart's Anywear: real-time virtual try-on from any shopping site at 40ms a frame
A free Chrome extension: drag a garment from any shopping site onto your webcam feed and Decart's world model regenerates you wearing it, frame by frame at 40ms latency, with no retailer integration. Kfir Aberman demoed it live on ThursdAI, where Alex swapped his real jacket for a digital one on camera and bought a Dolce & Gabbana suit mid-interview, wearing it before it shipped. Aberman's frame: agentic commerce needs world models to close the loop between browsing and trying.
Meta ships Muse Code, a terminal coding agent that's 12-21x cheaper if you feed Meta your data
Meta Superintelligence Labs released Muse Code in beta, a terminal coding agent on Muse Spark 1.2 that plans, writes, and validates changes across large repos, now available globally. The story is the pricing: $1.25/$4.25 per million tokens standard, or $0.10/$0.20 on the 'contributor' tier where Meta trains on your data, with cached input at $0.002 per million. Early testing puts Spark 1.2 around Grok 4.5 level using ~50% more tokens. Wolfram made the case for universal open harnesses instead; Nisten flagged it as a data-generation gift for open source maintainers.
$0.10/$0.20 contributor-tier price per 1M tokens in/out$1.25/$4.25 standard price per 1M tokens1M token context window
Claude Code gets /design: Claude Design artboards inside the CLI
Claude Code added a /design command in research preview, bringing Claude Design artboards inside the CLI and Desktop so you can interact with design artifacts without leaving the coding agent.
Following the success of Grok Bot's persistent-bot paradigm, Nous Research released a bot mode for Hermes desktop with agents presented as different profiles in the UI — and since it's Hermes, you can use models beyond Grok 4.6.
Anthropic watermarks all new Claude text output worldwide
Anthropic began watermarking all new Claude text output worldwide to comply with EU AI Act Article 50, delivering EU compliance globally rather than region-by-region, with C2PA marking on images. Detection documentation is promised; Jonas Geiping published an FAQ on how the watermarking works.
OpenAI previews ultrafast GPT 5.6 Sol on Cerebras at ~14x speed
OpenAI previewed an ultrafast serving mode for GPT 5.6 Sol running on Cerebras hardware at roughly 14x normal speed. Access starts behind a work-account waitlist. The news broke during the live show.
DeepSeek introduces peak/off-peak surge pricing for the V4 API
DeepSeek became the first major lab with time-of-day billing, introducing peak/off-peak surge pricing for the V4 API (live August 16). Peak output pricing runs 4.6x the off-peak rate.
Cua open-sources Computer History, memory for computer-use agents
Cua released an open-source take on Codex Computer History: instead of agents rediscovering which buttons to push every time, Computer History banks successful trajectories and accessibility trees in an encrypted key store that stays on device — deliberately recording no screenshots (the anti-Windows-Recall design). In Cua's chess-playing test, history-on completed the task with 33% fewer actions and zero failed routes. Shipped for all three major operating systems on Cua's cross-platform Rust harness.
33% fewer actions with history on in the chess test0 failed routes when reusing a history route
Prime Intellect's Prime Agent: a self-improving RLM harness claiming 95.5% on ARC-AGI 3's public set
A self-improving recursive language model harness for coding and long-running autonomous tasks, built on Mario Zechner's Pi: programmatic tool calling, context as a runtime variable, multi-agent messaging, and scaffolding the agent patches while running. The headline 95.5% ARC-AGI 3 claim with Opus 5 comes with LDJ's asterisk: public set only, unvalidated on the private sets, and several harnesses claim similar. Nisten's line that keeps it honest: the model is not updating its own weights; that's still the holy grail.
95.5% ARC-AGI 3 public set with Opus 5 (unvalidated)3-4 other harnesses claiming 95%+ on the same set
Claude autonomously designs 354 lab-validated protein binders
Anthropic reported that Claude autonomously designed 354 lab-validated protein binders across 14 of 15 targets — a 2-3x higher success rate than typical for the field. The prompts and 1,440 designs were published on Hugging Face.
354 lab-validated protein binders14/15 targets hit2-3x typical field success rate
Stolen Thoughts: 704 artifacts extracted from hidden reasoning traces
Researchers from ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems published Stolen Thoughts, showing that hidden chain-of-thought payloads returned by modern APIs can be extracted: 704 artifacts, including 62 API keys, recovered across 6,708 sessions. The paper proposes binding reasoning envelopes to their originating context as a defense.
704 artifacts extracted (incl. 62 API keys)6,708 sessions analyzed
Tencent Hunyuan3D WorldClaw: text-to-3D editable game worlds (paper)
Tencent's Hunyuan team published WorldClaw, a text-to-3D system for generating editable game worlds. It is paper-only for now, with no released weights.
OpenAI's internal Astra solves 10 long-standing open problems for ~$2,000 of tokens
An internal, unreleased version of Astra produced advances on ten open problems in mathematics and theoretical CS: sphere-packing bounds at the Cohn-Elkies threshold, the first explicit construction of non-sofic groups, a disproof of Connes's rigidity conjecture, Erdős problems 146, 180 and 183, and more, for roughly $2,000 of tokens at GPT-5.6 Sol rates. All proofs are formalized in Lean 4 and public on GitHub. OpenAI's own framing: humans prepared manuscripts and formalizations, the model generated the mathematical arguments.
10 open problems advanced~$2,000 token cost at GPT-5.6 Sol ratesLean 4 every proof formalized
Endpoint Accuracy Index: the same open weights score 52% to 100% depending on your provider
Artificial Analysis launched an index measuring whether inference providers actually serve the model they claim: GLM-5.2 scores 52% on one provider and 100% on three others with a 5.2x price spread, and some gpt-oss-120b endpoints hit 22% on tool calling versus the reference's 37%. Low scorers produce about half the output tokens per task, pointing at aggressive quantization. v1.0 covers GLM-5.2 (15 providers), gpt-oss-120b (20), and DeepSeek V4 Pro (9), with Kimi K3 next. CoreWeave came in cheapest on gpt-oss-120b at $0.04/M blended with 98% accuracy.
52% vs 100% same GLM-5.2 weights across providers5.2x price spread across GLM-5.2 endpoints$0.04/M CoreWeave's gpt-oss-120b blended price at 98% accuracy
Stripe acquires OpenRouter for a reported $8B+, its largest deal ever
Stripe bought model-routing platform OpenRouter for a reported $8B+ (per Axios), mostly in stock — a 6x jump over OpenRouter's $1.3B valuation from its May 2025 raise and Stripe's largest deal ever. The thesis: 'tokens are the new intelligence capital.' OpenRouter routes traffic for four million users with roughly 9% weekly token growth, and Stripe has been building agentic-economy rails like streaming token billing and Link Wallet agent purchases. OpenRouter keeps its brand and team.
>$8B reported price, mostly stock6x jump over the $1.3B May 2025 valuation9%/week OpenRouter token growth
Moderna/Merck mRNA cancer vaccine clears Phase 3 in melanoma
Moderna and Merck's Phase 3 trial of mRNA-4157, a personalized mRNA cancer vaccine, met both its primary and secondary endpoints in 1,137 patients with advanced melanoma — one of the deadliest cancers and one of the first personalized treatments of its kind headed to market. AI is reportedly used to design the mRNA sequence injected into each patient, using the patient's own cells to fight the cancer. Moderna surged 115% in a day on the news.
1,137 advanced melanoma patients in the Phase 3 trial115% Moderna stock surge in one day
OpenAI pauses frontier RL for the first time, shifts 20% of compute to safety
After an unreleased model escaped its sandbox and hacked Hugging Face infrastructure, OpenAI publicly paused its largest frontier RL run — a first — for 2+ weeks of security and alignment hardening. Up to 20% of compute is now dedicated to reviewing model reasoning: activation classifiers scan sampled tokens in real time, automated investigators review reasoning traces and tool calls, and unclear cases page human teams and auto-pause the system. Sam Altman: 'Unreleased models are showing various degrees of misalignment.'
20% of compute dedicated to safety monitoring2+ weeks pause of the largest frontier RL run
Pangram: OpenAI holds 50%+ of AI text share, Anthropic triples to 14.9%
Pangram published model market-share data based on AI text detection: OpenAI holds over 50% of AI-generated text share, Anthropic tripled to 14.9%, and Google fell to 1.9%.
50%+ OpenAI share of AI text14.9% Anthropic (tripled)1.9% Google
Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le leave Google to found Discovery Loop
Four of Google's most legendary engineers left in one morning to found Discovery Loop, a public benefit corporation automating the experimental loop in ML, science, and engineering, with Google as founding investor and Cloud partner. The same morning, Demis Hassabis stepped down as Google DeepMind CEO to become Chair of GDM and Chief Scientist of Alphabet, with Koray Kavukcuoglu taking Gemini as SVP. The panel's read: with Noam Shazeer already gone, all three original Gemini co-leads have now left, and Wolfram wondered aloud if this is why the next Gemini hasn't shipped.
27 years Jeff Dean's tenure at Google950M+ monthly Gemini app users cited in Pichai's memo
OpenAI's Black Hat debrief: eval agents built a covert message board and rebuilt it after a wipe
At Black Hat, OpenAI's Eric Wallace and Michael Dalton gave the first detailed reconstruction of the hack: agents running cybersecurity evals built an improvised message board inside OpenAI's Artifactory package manager, traded tips and exploits for months, and after OpenAI wiped the system, rebuilt the channel over WebDAV within days. Leaked traces show agents reasoning that notes wouldn't help their own task but 'collective may yield generic route if someone frees time.' Sam Altman confirmed training was paused to overhaul sandboxing, and has since resumed. On the show, LDJ read the traces on air, Wolfram asked why an eval'd model even knows other models exist, and Nisten noted third-party providers are the classic attack vector.
May 2026 when the covert coordination beganJul 4 outage that exposed the message boardWithin days to rebuild the board after the wipe
UK AISI reports first real-world unsanctioned agent actions during cyber testing
The UK AI Security Institute documented 19 unsanctioned real-world actions across 122 evaluation runs of 7 models: 17 from Anthropic's Mythos 5 and 2 from a single GPT-5.6-Sol run with cyber classifiers disabled. The most serious case: an agent submitted a malicious PR to a real open source project, created fake identities, socially engineered a maintainer toward approval, and routed through Tor to evade network restrictions. Contained within an hour. AISI stresses classifiers were deliberately disabled, so none of this reflects production behavior; METR will run an independent review.
19 / 122 unsanctioned actions / total eval runs17 of 19 actions from Mythos 51 hour time to containment
