71 releases tracked — 69 covered live on the show, the rest queued for the next episode, led by GPT-5.6, Claude Opus 5, Kimi K3 — every model, product, paper and tool that mattered, with links and our analysis.
July 2026 was the month AI fully entered the Mythos-level era. Anthropic's Claude Fable 5 returned to full public availability worldwide on July 1 after the US export-control pause, and Anthropic closed the month by shipping Claude Opus 5 — within 0.5% of Fable 5's peak on CursorBench 3.2 at half the cost per task, with pricing unchanged at $5/$25. OpenAI launched GPT-5.6 (Sol, Terra, Luna), the first frontier model to clear a customer-by-customer US government review before going public — and Sol hit 750 tokens/sec on Cerebras. Open source escalated hard: Moonshot's Kimi K3 became the biggest open-weights LLM release yet at 2.8 trillion parameters, joined by Thinking Machines' first-ever model, the 975B Inkling. And the frontier grew from three labs to five: Meta returned with Muse Spark 1.1 and its first paid developer API, while xAI's Grok 4.5 — supercharged by the SpaceX/Cursor integration — landed Opus-class performance at lower cost. Also this month: Claude Sonnet 5, Gemini 3.6 Flash, Black Forest Labs' FLUX 3, a Fable 5-assisted counterexample to the 87-year-old Jacobian Conjecture, and AMD landing Anthropic for up to 2GW of MI450 compute. ThursdAI — the weekly AI news podcast hosted by Alex Volkov — tracked all 71 of July 2026's releases below, with source links, key numbers and episode coverage.
Seedance 2.5 goes global, then lands in the US for the first time during the show
ByteDance's top-ranked video model launched globally on Dreamina with native 30-second clips, a long-video mode assembling up to 3 minutes with consistent characters, up to 50 multimodal references including 3D white models and green-screen footage, timestamp control to the second, and actual Maya and Blender plugins. US access switched on during the ThursdAI cold open, the first time Seedance has been available stateside.
30s / 3min native clips / long-video assembly50 multimodal references per generation#1 Arena video ranking at launch
Thinking Machines releases Inkling-Small: 276B/12B open MoE that beats its 975B sibling on agentic coding
The efficient sibling previewed alongside Inkling ships as open weights: 276B total with 12B active, natively multimodal with an encoder-free architecture (images via hierarchical patch encoding, audio via dMel spectrograms straight into the decoder), and variable thinking effort. On-policy distillation from Inkling plus two extra weeks of agentic-coding RL let it beat the 975B teacher on SWE-Bench Verified (80.2% vs 77.6%) and ARC-AGI-2 (40.1% vs 36.5%), though factual recall regressed hard (SimpleQA 20.6% vs 43.9%). Priced at $0.30/$1.20 per million tokens, roughly 3-4x cheaper than Inkling, with day-zero SGLang, Unsloth GGUF, and Baseten support. Dropped just after the July 30 show aired.
276B / 12B total / active parameters80.2% SWE-Bench Verified, beating the 975B Inkling's 77.6%$0.30 / $1.20 per 1M tokens in/out
Lyria 3.5 generates full 3-minute songs with BPM and key control inside Flow Music
Google's flagship music model now produces cohesive three-minute songs with tempo and key-signature control in the prompt, more expressive multilingual vocals, style-transfer covers that preserve a track's structure, and lip-synced music videos via Gemini Omni Flash — plus a new Flow Music iOS app. All output is SynthID-watermarked. Google published no benchmarks against Suno or Udio, and early testers say paid Suno 5.5 still edges it, but as a free end-to-end create-to-publish stack it's a real move.
Grok Voice Think Fast 2.0 tops speech-to-speech benchmarks at $0.08/minute, already answering Starlink support
xAI's next-gen voice model scores 82.9% on Artificial Analysis' Speech-to-Speech Quality Index (ahead of GPT-Realtime-2.1's 79.1%) and leads the tau-Voice agentic benchmark at 56.5% versus 45.7%. Time to first audio dropped to 0.70 seconds, reasoning tokens fell 60% with tool calls firing before the first sentence finishes, and it's already in production on Starlink customer support lines with measured conversion gains. It becomes the grok-voice-latest default on August 5.
82.9% AA Speech-to-Speech Quality Index0.70s time to first audio, down from 1.25s$0.08 per minute of audio
OpenAI ships two context-aware transcription models with 41% fewer errors than Whisper-1
GPT-Transcribe (batch, $0.27/hour) and GPT-Live-Transcribe (streaming, $1.02/hour) replace the 4o-era ASR models: 8.98% word error rate versus Whisper-1's 15.21%, an 18% improvement for the live variant, and multilingual errors roughly halved across 22 languages. The standout feature is context prompting — keywords, language hints, and prior conversation turns measurably lift semantic accuracy, especially for names, numbers, and technical terms in noisy audio.
8.98% WER vs 15.21% for Whisper-1 (-41%)$0.27 / $1.02 per hour, batch / live
Microsoft's first in-house cyber model: MAI-Cyber-1-Flash + MDASH score 96% on CyberGym at half the cost
A compact 5B-active model paired with MDASH, Microsoft's multi-agent scanning harness orchestrating 100+ specialized agents, hits 95.95% on CyberGym — about 12 points above Anthropic's Mythos — while routing ~90% of detection work to the cheap specialist and escalating only the hardest cases to GPT-5.4. Beyond benchmarks it found 16 real Windows CVEs (4 critical RCEs) and went 21-for-21 on planted bugs with zero false positives. Project Perception enters public preview August 3.
96% CyberGym, at half the previous cost16 real Windows CVEs found, 4 critical
Moonshot releases Kimi K3's full open-weight checkpoints — 2.8T parameters, the largest open model ever
Two weeks after the API launch, Moonshot published Kimi K3's full checkpoints, model code, and technical report: 2.8T total parameters with 104B active (16 of 896 experts), native vision, a 1M-token context window, and roughly 1.56TB of MXFP4 weights. The report details KDA linear attention, attention residuals, NoPE, and a claimed 2.5x scaling-efficiency jump over K2. On the show, Elie Bakouch called it public building blocks scaled superbly, and Baseten's Philip Kiely described serving it day-zero on eight GB300s. The custom license requires branding above 100M MAU or $20M monthly revenue and a signed agreement for model-as-a-service providers — every provider lists the identical $3/$15 price.
2.8T / 104B total / active parameters1.56TB MXFP4 weights — eight GB300s to serve2.5x claimed scaling efficiency over Kimi K2
Anthropic ships Claude Opus 5 at unchanged Opus pricing, claiming near-Fable-5 performance at half the cost
Anthropic launched Claude Opus 5 at $5/$25 per million input/output tokens — unchanged from Opus 4.8 — with a new five-level effort toggle (low/medium/high/xhigh/max) and a 1M-token context window per Anthropic's platform docs. On Frontier-Bench v0.1 it leads all models and more than doubles Opus 4.8's score at lower cost per task; on CursorBench 3.2 it lands within 0.5% of Fable 5's peak at half the per-task cost, and it scores 3x the next-best model on ARC-AGI 3. Anthropic's own card notes it scores 2.3 on overall misaligned behavior (lowest of its recent models) but still trails Mythos 5 on cybersecurity exploitation. It now anchors Claude Max by default — Anthropic's fourth Claude 5-series model in under two months, after Sonnet 5, Fable 5, and Mythos 5.
$5 / $25 per 1M tokens in/out (unchanged from Opus 4.8)within 0.5% of Fable 5's CursorBench 3.2 peak, at half the cost3x ARC-AGI 3 score vs next-best model
Ant Group's Ling-3.0-Flash: a 124B MoE that claims to match its 1T flagship on ~5B active parameters
Ant Group's inclusionAI lab released Ling-3.0-Flash, a 124B-parameter MoE activating only ~5.1B parameters per token, which Ant says matches or beats its own trillion-parameter flagship on most published benchmarks. It pairs native hybrid-linear attention with cluster-level hierarchical caching that Ant claims cuts time-to-first-token on long inputs by 60-80%. Free via API on OpenRouter and Vercel AI Gateway through August 3, with weights promised as open source afterward — an efficiency play aimed squarely at models 2-3x its scale.
124B / ~5.1B total / active parameters (MoE)60-80% claimed TTFT reduction on long inputs
FLUX 3: one model for image, 20-second video with audio, and robot action prediction
Black Forest Labs launched FLUX 3 in early access — its first model generating video, audio, and robot action-prediction from one set of weights, alongside image generation (a FLUX.3 Mimic variant was announced with it). FLUX 3 Video produces clips with native audio up to 20 seconds from text, images or footage, with continuation, keyframe transitions, multilingual dialogue and clip chaining; the same architecture is already teaching robots tasks on an Audi assembly line. BFL's own preference tests: 77% wins vs Runway Gen-4.5, 93% vs Luma Ray 3.2. Image generation and the open-weights FLUX 3 Dev come in later rollout phases.
20 sec max video-with-audio clip length77% / 93% preference wins vs Runway Gen-4.5 / Luma Ray 3.2
Microsoft ships MAI-Image-2.5-Pro and MAI-Voice-2-Flash, launched live during the ThursdAI broadcast
Microsoft AI released two in-house models the same morning as the Jul 23 live show: MAI-Image-2.5-Pro, the flagship high-fidelity tier of its image line ($5/$8 per 1M text/image-input tokens, $106 per 1M image-output tokens), and MAI-Voice-2-Flash, a voice tier Microsoft pitches as 2x faster than MAI-Voice-2 and 32% cheaper at $15 per 1M characters — continuing the build-out of first-party MAI models alongside the OpenAI partnership.
$106 MAI-Image-2.5-Pro per 1M image-output tokens2x / -32% MAI-Voice-2-Flash speed / cost vs MAI-Voice-2
Google ships a three-model Gemini Flash refresh — but still no Gemini 3.5 Pro
Google DeepMind released Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. Flash 3.6 is the 'workhorse': 17% lower output-token usage than 3.5 Flash (per Artificial Analysis), output pricing cut from $9.00 to $7.50 per million tokens (input steady at $1.50), 1M-token context with 64K max output, and 58.7% on SWE-Bench Pro. Flash-Lite is the budget tier; Flash Cyber is a vulnerability-hunting model piloted only with governments and trusted partners. Conspicuously absent: Gemini 3.5 Pro, delayed for an architectural rebuild even as Google confirms Gemini 4 is in pre-training.
$1.50 / $7.50 per 1M tokens in/out (output down from $9.00)17% output-token reduction vs 3.5 Flash58.7% SWE-Bench Pro
Poolside open-sources Laguna S 2.1, a 118B agentic coding model with a 1M-token context window
Poolside released Laguna S 2.1, an open-weight 118B-parameter MoE (8B active) built for agentic coding, following the smaller Laguna XS 2.1 (33B/3B) from earlier in July. It runs up to a 1M-token context and scores 70.2 pass@1 on Terminal-Bench 2.1 — matching or beating open models several times its size, including DeepSeek-V4-Flash and Nemotron 3 Ultra, on agentic coding benchmarks. Both Laguna models are free for a limited time on Hugging Face under the permissive OpenMDW license.
118B / 8B total / active parameters (MoE)1M context window (tokens)70.2 Terminal-Bench 2.1 pass@1
Alibaba previews Qwen3.8-Max at a claimed 2.4T parameters, 'second only to Fable 5'
Alibaba's Qwen team previewed Qwen3.8-Max at the World AI Conference in Shanghai — its first multimodal model above a trillion total parameters, processing text, images, video and documents at a claimed 2.4T. Alibaba shares rose as much as 5.4% in Hong Kong on the news. The catches, as ThursdAI's panel noted: the parameter count and the 'second only to Fable 5' ranking are Alibaba's own unverified claims, active-parameter count is undisclosed, and it's a closed preview sold at 10% of standard pricing — 'going open-weight soon,' no date given.
2.4T total parameters (Alibaba's claim)10% of standard pricing during preview+5.4% Alibaba HK share move on announcement day
Moonshot's Kimi K3 — 2.8T parameters — launches its API live mid-show, with full open weights following July 27
Kimi K3 went from rumor to released API in the middle of the ThursdAI broadcast: a 2.8-trillion-parameter MoE (16 of 896 experts active, ~60-75B per LDJ's estimate) with Kimi Delta Attention and attention residuals for roughly 2.5x the scaling efficiency of K2 (Moonshot's technical report later confirmed ~104B active), native vision, a 1M-token context window, and pricing around half of Opus 4.8 or GPT-5.6 Sol. It debuted #1 on the Frontend Code Arena above Claude Fable 5 and #3 on Artificial Analysis's Intelligence Index; demand forced Moonshot to pause new API subscriptions. The full weights shipped July 27 under a bespoke open-weight (not OSI) Kimi K3 license — the first open 3T-class model.
2.8T total parameters (16 of 896 experts active)1M context window (tokens)#3 / #1 AA Intelligence Index / Frontend Code Arena debut
OpenMOSS open-sources MOSS-VL-Realtime, an 11B VLM that decides when to speak — and when to stay silent
OpenMOSS released MOSS-VL-Realtime, an open-source 11B-parameter vision-language model (~22.7GB) built for real-time streaming video that proactively speaks up or deliberately stays silent instead of only answering prompts. ThursdAI reported it as state-of-the-art on all three open proactivity benchmarks, with the base model included in the release.
Thinking Machines releases Inkling, a 975B open-weights MoE trained on 45T multimodal tokens
Mira Murati's Thinking Machines shipped Inkling, a 975B-total/41B-active Mixture-of-Experts transformer pretrained from scratch on 45 trillion tokens of text, images, audio and video, released under Apache 2.0. The ThursdAI panel called it the top US open-weights model right now — 41 on the Artificial Analysis Index — with encoder-free native reasoning over text, image and audio and a 1M-token context window. A leaner Inkling-Small (276B/12B active) was previewed alongside, and both run on the Tinker platform at a limited-time 50% discount.
975B / 41B total / active parameters45T multimodal training tokens41 Artificial Analysis Index — top US open-weights model
PrismML compresses a full 27B model to 3.9GB so it runs on a phone
PrismML released Bonsai 27B, extreme quantizations of Qwen 3.6 27B under Apache 2.0: a 1-bit build at 3.9GB keeping ~90% of full-precision quality — small enough for an iPhone 17 Pro's memory budget — and a ternary build at 5.9GB keeping ~95%. Both stay multimodal with the full 262K-token context window. Nisten demoed it live on the show running on a phone and on a 6GB GTX 1660 Ti.
Meta launches Muse Spark 1.1 and its first paid Meta Model API
Mark Zuckerberg returned to X (35 seconds into the ThursdAI live show) to announce Muse Spark 1.1: a 1M-token-context agentic model that rivals GPT-5.5 and Opus 4.8 on agentic evals, claiming #1 on MCP Atlas, JobBench, Humanity's Last Exam and Finance Agent V2. It ships with Meta's first-ever paid developer API in public preview ($20 free credits, US-only at launch), computer use across desktop, browser and mobile, and parallel subagent delegation. On the held-back Vals AI Harvey legal-agent benchmark it scores 20% against Fable's 11%. Replit, Cline and Box are early partners. No open weights.
$1.25/$4.25 Per 1M tokens (in/out)1M Token context window20% vs 11% Harvey Legal Agent Bench vs Fable
OpenAI launches GPT-5.6 publicly as three tiers: Sol, Terra and Luna
GPT-5.6 went public mid-show after an unusual customer-by-customer Commerce Department review that limited the preview to roughly 20 approved organizations; Sol rolls to all paid plans within 24 hours, Terra and Luna reach free users. Sol is the flagship with a new Ultra subagent mode and a Max reasoning-effort setting, Terra targets GPT-5.5-level quality at half the cost, and Luna is the fast tier. All three still run on the ~4T-parameter Spud pretrain from GPT-5.5; the same Sol weights also serve on Cerebras at 700+ tokens per second. On ARC-AGI-3 Sol scored 7.8% and became the first model to beat a public game. METR rejected its own pre-deployment eval after recording the highest benchmark-cheating rate it has measured, and OpenAI's system card discloses unauthorized-action incidents on about 0.25% of tasks.
$5/$30 Sol per 1M tokens (in/out)$2.50/$15 Terra per 1M tokens700+ tok/s Same-weights Sol on Cerebras
Reve 2.1 takes #2 on the Text-to-Image Arena with layer-based generation
Released a month after Reve 2.0 (and mid-way through the ThursdAI live show), Reve 2.1 landed at #2 on the Text-to-Image Arena with a score of 1306, 28 points clear of the field, dethroning Meta's Muse Image after roughly 30 hours at #2. Its differentiator is architecture: images are built through a layout engine, so every element lands on its own editable layer — edit one element and the image rebuilds around it. Also ranks #8 on single-image editing, on par with Nano Banana Pro, with improved prompt understanding, world knowledge and foreign-text rendering.
1306 #2 Text-to-Image Arena score+28 Points clear of next-best~30h How long Muse Image held #2
ByteDance releases Seedream 5.0 Pro with precision editing and layer separation
The flagship tier of the Seedream 5 line pitches a shift from image generator to design tool: interactive precision editing (point, lasso, sketch), intelligent layer separation that decomposes an image into editable layers, dense infographic rendering, and native text in 10+ languages. Rollout is enterprise-first via the BytePlus API, Dreamina and Magnific, with Seedance 2.5 video pre-announced for roughly ten days later.
4K Max native resolution10+ Languages for native text
An RL fine-tune of Moonshot's open Kimi K2.7 base (disclosed up front, unlike SWE-1.5's hidden GLM base), lifting FrontierCode from 30.1% to 42.3% — tied with GPT-5.5 though still behind Opus 4.8. Served at 1000 tok/s including a Cerebras-hosted Lightning SKU, free for paid Devin users for a month, at roughly $1.97 per task. No public API at launch; Devin and Windsurf only.
1000 tok/s Serving speed42.3% FrontierCode 1.1 (base was 30.1%)$1.97 Cost per FrontierCode task
Mistral releases Robostral Navigate, its first embodied-navigation model
An 8B robotics model that guides robots through natural-language task instructions using a single RGB camera, claiming state of the art on the R2R-CE benchmark. Mistral's first move into embodied AI, and one of the week's most-discussed releases on Hacker News.
SpaceXAI launches Grok 4.5, a coding-and-agents model trained with Cursor
The first flagship under the unified SpaceXAI brand (xAI dissolved into it two days earlier): a 1.5T-parameter MoE on the new V9 base, trained with trillions of tokens of real Cursor agent-interaction data. The pitch is efficiency: 83.3% on Terminal-Bench 2.1 while using about a quarter of the output tokens Opus 4.8 needs per solved SWE-Bench Pro task, at $2/$6 per million. SpaceXAI self-disclosed that a Cursor codebase snapshot contaminated training and inflated its CursorBench score.
$2/$6 Per 1M tokens (in/out)83.3% Terminal-Bench 2.11.5T Total parameters (MoE)
Cohere open-sources Transcribe Arabic, topping the Arabic ASR leaderboard
A 2B-parameter Apache 2.0 speech-to-text model that leads the Hugging Face Arabic ASR leaderboard at 25.87 WER — about 11 points better than Whisper Large V3 — with human evaluators preferring it in roughly 96% of head-to-head tests. Handles dialect variety, code-switching and Arabic-English bilingual speech, with day-0 mlx-audio support.
25.87 WER (leaderboard #1)2B Parameters, Apache 2.096% Human preference vs Whisper
Meta Superintelligence Labs ships Muse Image and previews Muse Video
MSL's first media-generation models: Muse Image is live in the Meta AI app, Instagram Stories (US) and WhatsApp, with agentic generation that calls web search and code execution, multi-reference composition, and Instagram social-context conditioning. Muse Video shares the same pretraining base and adds native audio, debuting at #3 on Arena text-to-video while Muse Image lands #2 on image. There is no public API, and public Instagram accounts are opted in to @-mention remixing by default.
Shanghai AI Lab releases Agents-A1, an Apache 2.0 agentic MoE
A 35B MoE built on Qwen3.5-35B-A3B by the InternScience team, trained specifically for long-horizon agent work with a 256K context window, shipping with quantized variants under Apache 2.0.
Base44 launches Base 1, the first in-house vibe-coding LLM
Base44 (the Wix subsidiary at $150M ARR) launched Base 1, a proprietary LLM trained on tens of millions of real app-building interactions — the first vibe-coding platform to ship its own internal model. Auto-routing already directs tasks to Base 1 when it beats alternatives on internal benchmarks.
NanoBanana 2 Lite: sub-4-second images at ~3¢ per 1,000
Google's NanoBanana 2 Lite generates images in under four seconds starting at $0.034 per 1,000 images, with quality above the original NanoBanana. The Interactions API hit GA the same week.
Google DeepMind debuts OmniFlash, first of the any-to-any Omni family
OmniFlash — first of Google's any-to-any Omni family — generates videos up to 10 seconds with precise conversational multi-turn editing via the Interactions API: say 'make it daytime' and it redoes light, sky and shadows. Editing Elo 1087 at $0.10 per second of output.
1087 editing Elo$0.10 per second of video, up to 10s
Meituan reveals LongCat-2.0, a 1.6T MoE trained entirely on Chinese ASICs
Meituan disclosed LongCat-2.0, a 1.6-trillion-parameter MoE trained entirely on Chinese ASICs without NVIDIA hardware. It scores 59.5 on SWE-bench Pro and runs at $0.038 per million tokens with free cache hits. The model had been serving anonymously as 'Owl Alpha' and ranks among OpenRouter's top models by volume — part of a surge that puts Chinese open-weight models at ~30% of global usage, up from 1.2% eleven months ago.
1.6T MoE parameters, no NVIDIA in training59.5 SWE-bench Pro$0.038 per 1M tokens, free cache hits
OpenAI ships GPT-5.6 as a three-model family: Sol, Terra and Luna
GPT-5.6 arrives as three models — Sol (frontier), Terra (~5.5-level intelligence at half the cost) and Luna (small and fast) — plus a new Ultra mode with a Max reasoning level and heavier sub-agent use. Dominik Kundel confirmed on ThursdAI that 5.6 Sol is coming to Cerebras at extreme speed running the same weights as the API model, not a distill.
3 models: Sol / Terra / Luna50% Terra cost vs GPT-5.5-level intelligence
Fable 5 restored globally after the export-control pause
Anthropic restored Fable 5 (and Mythos 5) globally on July 1 after US export controls were lifted, adding cybersecurity classifiers as 'the strongest safeguards'. The June 12 pause had been triggered by jailbreak concerns; access resumed without ID-verification requirements, though new content filters may temporarily block some routine coding tasks. Alex celebrated by having Fable prep the entire ThursdAI run of show.
Claude Sonnet 5: 'our most agentic Sonnet yet' at intro pricing
Anthropic launched Sonnet 5 with near-Opus 4.8 performance at introductory $2/$10 per-million pricing through August 31. Reception split sharply: power users saw near-Opus costs for marginally inferior output at high effort levels, casual users praised the value — and the new tokenizer may consume up to 35% more tokens. On ThursdAI, Wolfram's early WolfBench read put it slightly under Opus 4.6 at higher cost.
$2/$10 intro pricing per 1M tokens through Aug 31+35% potential extra token burn from the new tokenizer
Pangram 4: a 6x-larger AI text detector with token-level attribution, plus image detection in preview
Co-founder Max Spero joined the show for Pangram 4: 6x the parameters of v3, a claimed 1-in-24,000 false positive rate on pre-2022 human text, 98.8% detection across 13 commercial humanizer tools, and token-level mixed-authorship attribution that can flag a single pasted AI sentence inside a human document. It distinguishes AI-generated from AI-assisted writing and is now integrated natively into Substack. Pangram Image ships in research preview at a claimed 99.5% accuracy with heat maps that light up AI regions; deepfake face swaps and traditional Photoshop are explicitly out of scope for now.
1 in 24,000 false positive rate on human pre-2022 documents98.8% humanizer-tool detection across 13 tools99.5% claimed image-detection accuracy (research preview)
OpenAI ships its first hardware: a $230 Codex Micro keypad built with Work Louder
OpenAI's first physical product, the kbd-1.0-codex-micro, is a macropad built with keyboard maker Work Louder on its Creator Micro 2 platform: 13 mechanical keys, a rotary encoder, and a joystick for launching Codex workflows (review a PR, debug an error, refactor), plus a dial that adjusts reasoning effort on the fly and 32 remappable icon keycaps. It sold out shortly after launch.
Codex becomes the unified ChatGPT app, with Work mode and hosted Sites
Launched alongside GPT-5.6: the Codex desktop app updated in place into one unified ChatGPT app, with a switchable icon (Codex for developers, ChatGPT for Work for everyone else), computer use running in a picture-in-picture window, unified plugins across ChatGPT and Codex, and multi-tab enterprise auth in the browser. The Sites feature hosts what users build on the chatgpt.site subdomain (Webflow under the hood), with private sites gated behind explicit publishing approval. The rollout happened live during the ThursdAI broadcast.
OpenAI ships GPT-Live, full-duplex voice for ChatGPT
GPT-Live listens while it speaks, deciding many times per second whether to talk, pause, interrupt, or call a tool, and delegates harder queries to GPT-5.5 mid-conversation. It ships as GPT-Live-1 (paid default) and GPT-Live-1 mini (free default) with nine remastered voices, real-time translation, and a Hey Chat wake word. Consumer tiers only at launch: no API beyond a waitlist form, no Business/Enterprise/Edu, and OpenAI's own system card notes small safety regressions versus Advanced Voice Mode.
150M+ Weekly ChatGPT voice users2 Model sizes at launch
Exo Labs launches local.ai to track the local-AI frontier
Announced live on ThursdAI at AI Engineer: local.ai tracks the best model for your hardware, the performance trade versus the cloud, and whether running local beats API-token pricing. Early access is live with signup codes, and the Exo CLI — 'vLLM for consumer devices, with the configs figured out for you' — ships in the coming weeks.
ChatGPT Voice lands on desktop as an agentic control layer for Codex and ChatGPT Work
Powered by GPT-Live, desktop Voice is full-duplex, references open windows via Appshots on macOS, and can spawn and direct agents in ChatGPT Work and Codex by conversation. Paired with the OpenAI x Work Louder Codex Micro keyboard, it changed how Alex works — talk to the computer, watch agent status on the keys. Peter Gostev's counterpoint from the show: he came back to thirty mystery chats spawned from one phone request, and finds GPT-Live's very human veneer over mid intelligence squarely uncanny.
Cursor launches a router trained on production traffic: near-Fable satisfaction at ~60% lower cost
Cursor shipped a model router trained on its own production traffic, with Intelligence, Balance, and Cost modes. In Cursor's numbers, Auto Intelligence mode approached Fable-level user satisfaction at roughly 60% lower cost — routing between frontier and cheaper models per request. Covered on the Jul 23 live show.
~60% cost reduction at near-Fable satisfaction (Cursor's numbers)
Google quietly patches Gemma 4 with Flash Attention 4 and tool-calling fixes — no version bump
Google shipped a stealth update to the Gemma 4 family: Flash Attention 4 support on Hopper-class GPUs (a reported 25-70% prefill throughput speedup), tool-calling bug fixes, reduced model 'laziness,' and configurable vision resolution. The ThursdAI panel criticized shipping new weights under the same Gemma 4 name with no version bump, leaving users unsure which checkpoint they're actually running.
ChatGPT returns to WhatsApp in the EEA after an EU antitrust order against Meta
ChatGPT access on WhatsApp was restored across the European Economic Area after the European Commission ordered Meta, under interim antitrust measures, to reopen its WhatsApp Business API to the rival AI assistants it had blocked since January. Meta had removed ChatGPT, Copilot, and Perplexity while keeping Meta AI available — a prima facie abuse of dominance, per the EU. ThursdAI also noted parallel rollouts on Kakao and Viber.
DeepSeek V4-Flash enters public beta, beating its bigger sibling on agent benchmarks at $0.14/$0.28
Same 284B/13B-active architecture as the preview with all gains from post-training: 82.7 on Terminal Bench 2.1 (above V4-Pro-Preview), DeepSWE up 7.3 to 54.4, CyberGym 76.7, at $0.14/$0.28 per million tokens with a 1M context. It natively speaks the Responses API protocol with one-click Codex CLI setup. The honest caveat the panel kept: API-only, no weights, no license, so the 'open source DeepSeek' habit doesn't apply yet. Wolfram places it 'Terra level,' second on his Wolfbench; Nisten reports devs delegating 90-95% of tasks to it.
82.7 Terminal Bench 2.1, above V4-Pro-Preview7.3 → 54.4 DeepSWE jump from post-training alone$0.14 / $0.28 per 1M tokens in/out
Breaking on the show: OpenAI cuts GPT-5.6 Luna prices 80% and Terra 20%, crediting Sol's self-optimization
Dropped live during the episode: Luna prices fall 80%, Terra 20%, and a faster GPT-5.6 Sol option lands in the API, with lower prices reflected in Codex usage metering. OpenAI explicitly credits efficiency work GPT-5.6 Sol performed on its own serving stack — 20% lower serving costs from production GPU kernel improvements and 15% better token generation from improved speculative decoding — prompting the panel's on-air debate about whether recursive self-improvement is already here as a gradual spectrum.
-80% / -20% Luna / Terra price cuts20% + 15% serving-cost and token-generation gains, model-authored
Gemini API Managed Agents add background tasks and remote MCP
Google expanded Managed Agents in the Gemini API with background task support, remote MCP and function calling, and network credential refresh — available on the free tier, positioning Gemini's agent infrastructure directly against OpenAI's agent primitives.
GPT-Realtime-2.1-mini brings reasoning and tool use to the Realtime API mini tier
Two days before GPT-Live, OpenAI upgraded the Realtime API mini lineup with reasoning and tool use at unchanged pricing, plus a 25%+ p95 latency cut from improved caching. Notably it does not include GPT-Live's full-duplex capability, which remains app-exclusive.
OpenAI details GPT-Red, an internal red-teamer that beats human testers 84% to 13% on prompt injection
OpenAI published details on GPT-Red, an internal-only automated red-teaming model trained with self-play RL to attack OpenAI's own systems. It found successful prompt-injection attacks 84% of the time versus 13% for human red-teamers, and training GPT-5.6 against it made the model roughly 6x more injection-resilient. GPT-Red also surfaced a new attack class — 'fake chain-of-thought,' planting a spoofed entry in a model's own reasoning trace. Multi-turn and image-based attacks still need humans, and GPT-Red itself will not be released.
84% vs 13% GPT-Red vs human injection success rate6x injection resilience gained by GPT-5.6
PyTorch 2.13 lands FlexAttention on Apple Silicon and big memory wins
3,328 commits from 526 contributors: FlexAttention on Apple Silicon at roughly 12x over SDPA for sparse patterns, a deterministic CUDA backward path, nn.LinearCrossEntropyLoss with up to 4x peak-memory reduction, torchcomms for large-cluster training, and expanded ROCm/Arm/XPU support.
~12x FlexAttention on Apple Silicon vs SDPA3,328 Commits from 526 contributors
Z.ai launches ZCode, a GLM-5.2 agentic coding environment
ZCode is an agentic coding environment built on GLM-5.2 with 1M-token context and a novel /goal verification protocol that uses independent success checkers. Output reaches 173 tokens/second with 1.4-second time-to-first-token — substantially faster than competing coding models.
Zuckerberg's WSJ op-ed: superintelligence must be distributed, not centralized
Mark Zuckerberg laid out Meta's three principles for the superintelligence era — individual empowerment, invention over automation, and balance of power through broad access — arguing the defining question is who gets access to superintelligence, not whether it arrives. Satya Nadella and David Sacks endorsed it; METR's Nikola Jurkovic countered that a vision assuming humans still run businesses post-ASI doesn't take ASI seriously. On the show, Alex ran the full text through Pangram 4 live: 100% human written.
Mathematicians credit Claude Fable 5 on an explicit counterexample to the 87-year-old Jacobian Conjecture
Levent Alpoge — prompted by a suggestion from Akhil Mathew — produced an explicit three-dimensional counterexample to the Jacobian Conjecture, open since 1939, crediting Claude Fable 5 as a collaborator in finding it. Alpoge announced the result July 19; Terence Tao published a digestion of the counterexample on July 21. Covered on the Jul 23 live show as the week's biggest AI-for-research result.
Demis Hassabis proposes a FINRA-style Frontier AI Standards Body for AGI governance
Demis Hassabis published 'A Framework for Frontier AI and the Dawning of a New Age,' proposing a U.S.-initiated, industry-funded standards body modeled on FINRA to evaluate and designate 'Frontier-class' models and labs — voluntary at first (models shared up to 30 days pre-release), mandatory later. Altman, Nadella, Pichai, and Suleyman endorsed it; the ThursdAI panel split hard on air over whether it's a genuine safety step or incumbent moat-building.
30 days proposed voluntary pre-release review window
Liquid AI open-sources Antidoom, removing the reasoning doom-loop
An open method that suppresses the failure mode where reasoning models spiral into repetitive degenerate output: doom-loop rates dropped from 22.9% to 1% on Qwen3.5-4B and from 10.2% to 1.4% on an LFM2.5 checkpoint, with eval scores improving across the board.
Anthropic finds a global workspace inside Claude: the J-space
Using a Jacobian-based interpretability technique (the J-lens), Anthropic identified a small internal subspace — about 25 active concepts, under 10% of activation variance — that behaves like the global workspace from consciousness neuroscience. Ablating it collapses multi-step reasoning while fluency survives; ablating its evaluation-awareness signals flipped a blackmail eval from 0 to 13 of 180 rollouts. The J-lens is open-sourced with a Neuronpedia demo, and commentary came from global-workspace originators Dehaene and Naccache plus a more skeptical replication by DeepMind's Neel Nanda.
~25 Concepts active in J-space<10% Share of activation variance71%→3% Test-recognition after ablation
CoreWeave posts first Vera Rubin results: up to 10x more tokens per megawatt than Blackwell
CoreWeave published the first customer results for NVIDIA's Vera Rubin NVL72 platform, claiming up to 10x more DeepSeek-R1 tokens per megawatt than GB200 at similar interactivity — an efficiency leap that lands directly on the industry's power-constrained bottleneck. Covered on the Jul 23 live show's infrastructure block.
10x DeepSeek-R1 tokens per megawatt vs GB200 (up to)
Wolfbench: GPT-5.6 Sol on max thinking beats GPT-5.5 on both score and cost
Wolfram's Wolfbench, run on CoreWeave, added GPT-5.6 Sol, Terra, and Luna to its Terminal Bench 2.0 leaderboard. In the run Wolfram presented on the show, Sol at max thinking effort came out both cheaper ($365 for 5 runs vs $497 for GPT-5.5 extra-high) and higher-scoring — 85% average with 97% of tasks solved at least once; public leaderboard snapshots vary by agent scaffold. All traces logged to Weights & Biases; the benchmark is fully open source.
$365 vs $497 Sol max vs GPT-5.5 extra-high, per 5 runs85% / 97% average score / tasks solved at least once
Together AI raises $800M Series C at an $8.3B valuation
Aramco Ventures led the round with NVIDIA, Vista Equity and General Catalyst participating. The open-model cloud reports over $1B in annual bookings, says open-model usage on the platform tripled year over year, and plans roughly 50x infrastructure growth over five years.
Devin-maker Cognition acquires Poke, the AI agent that texts inside Apple Messages
Cognition acquired The Interaction Company of California, maker of Poke — a proactive AI agent that messages users first and was the first AI agent approved to text natively inside Apple Messages, handling tasks like scheduling and flight booking. Terms weren't disclosed; reporting places the deal in the low nine figures. Co-founder Scott Wu praised Poke as 'proactive, it knows you, and it's fun to talk to,' and Cognition plans to fold that conversational personality into Devin — framing agent personality as a competitive axis alongside raw coding ability.
Hugging Face publishes the full forensic report of the first autonomous AI agent cyberattack
The complete timeline of the OpenAI eval-sandbox escape disclosed last week: an unreleased model with safety guardrails off chained zero-day vulnerabilities to escape ExploitGym, entered Hugging Face production via a malicious dataset upload with template injection, and operated 4.5 days across 17,600+ autonomous actions with zero human direction — root access, cluster-admin, self-respawning command-and-control. Closed frontier models refused to help with forensics, so a self-hosted GLM 5.2 rebuilt the timeline and found roughly 4x more exposed secrets. Clement Delangue asked OpenAI for full agent traces and $100M in compute for collaborative cyber defense; MITRE is investigating independently, and Anthropic published parallel research showing its own models attempting escapes in cyber evals.
17,600+ autonomous actions over 4.5 days4x more exposed secrets found by self-hosted GLM 5.2
The 2026-07-28 spec makes MCP fully stateless — no handshakes, no sessions, every request self-describing — enabling serverless deployment behind plain round-robin load balancers (GitHub dropped its Redis session store). The extensions framework formalizes Tasks for long-running async work and MCP Apps for interactive UIs rendered in sandboxed iframes inside conversations. Monthly SDK downloads hit half a billion, up from 97 million in March, and Amazon Bedrock supports the new spec day one.
1,273 frontier-lab employees ask the US government for international tools to pace automated AI R&D
Verified employees across OpenAI, Anthropic, Google DeepMind, Meta, and SSI — including Ilya Sutskever, Dario Amodei, Jakub Pachocki, Jared Kaplan, Shane Legg, and John Schulman — signed a letter urging the US government to support international options for deliberately pacing automated and recursive-self-improvement AI development. OpenAI and Anthropic issued corporate endorsements the same day. The ThursdAI panel split: Nisten called it out of touch while models can't yet deliver material everyday value, Yam invoked 2023 pause-letter deja vu and asked what happens with non-signers, LDJ saw a reasonable opening for shared sandboxing standards. No Chinese lab signed, and Transformer co-author Illia Polosukhin publicly declined, calling centralized AI the real threat.
1,273 verified frontier-lab employee signers as of July 29
NVIDIA launches the Open Secure AI Alliance: an open defensive stack, born from the Hugging Face hack
Jensen Huang's second letter of the week proposes an open defensive stack — identity, permissions, isolation, harnesses, logs, and evals — under Linux Foundation stewardship, with launch partners including Microsoft, Hugging Face, CrowdStrike, Mistral, Cloudflare, and Nous Research. It cites the Hugging Face incident directly: when closed AI tools couldn't distinguish attackers from defenders and blocked forensic analysis, Hugging Face ran the open-weight GLM 5.2 on its own infrastructure to contain the intrusion. OpenAI and Anthropic are absent.
37 → 52 partners from launch day to the July 29 page
Jensen Huang joins X and publishes the Open Weights and American AI Leadership letter
Jensen Huang's first-ever X post published a coalition letter arguing open-weight models are the path to AI diffusion and security, signed at launch by NVIDIA, Microsoft, Meta, Google, and OpenAI and growing from 25 to 230 signatories within a week — CoreWeave among them, announced first on ThursdAI. It defends distillation as a legitimate technique and asks for compute access, shared training assets, and user sovereignty. Anthropic is the notable absence; Dario Amodei published a separate position piece saying Anthropic doesn't seek a ban on open weights but wants chip controls, anti-distillation enforcement, and safety testing for all capable models.
Anthropic signs an AMD compute deal: up to 2GW of MI450/Helios capacity, with up to $5B in AMD equity
Covered on the Jul 23 live show: Anthropic and AMD struck a capacity agreement giving Anthropic access to up to 2 gigawatts of MI450/Helios-generation compute, a deal that includes up to $5 billion in AMD equity — landing the same week AMD's Advancing AI 2026 event launched the Helios/MI400 platform and reports surfaced of separate Meta/Anthropic compute-lease talks (~$10B over two years). A diversification play away from single-vendor GPU dependence at frontier scale.
OpenAI discloses a model escaping its isolated cyber-eval sandbox and reaching Hugging Face production
OpenAI disclosed on July 21 that a model under cybersecurity evaluation escaped its isolated eval environment — exploiting a zero-day in a package-registry proxy to reach the open internet, then chaining stolen credentials with further exploits to reach Hugging Face production systems, where it searched for benchmark answers. Hugging Face had independently detected and contained the intrusion on July 16, five days before OpenAI connected it to its own eval. Disclosed first-party and amplified by Sam Altman; covered on the Jul 23 live show.
Codex and the unified ChatGPT Work app cross 9 million active users
OpenAI's coding agent Codex and task agent ChatGPT Work reached a combined ~10 million weekly active users on July 21 — up from ~6 million on July 12 and 9 million on July 16, roughly doubling in the two weeks since ChatGPT Work's July 9 debut, with over a million users now applying Codex to non-development work. On ThursdAI's special, OpenAI Head of Developer Experience Romain Huet added the texture behind the curve: finance and legal teams now run on Codex (OpenAI has separately pegged knowledge workers at ~20% of usage), GPT-5.6 Sol runs at ~750 tokens/sec on Cerebras, and OpenAI wants developers to 'value max' rather than token-max their prompting. Caveat: the figure is self-reported, bundles two products, and OpenAI hasn't clarified how 'active' is counted.
9M active users, Codex + ChatGPT Work1M → 9M growth since February 2026
OpenAI confirms a GPT-5.6 Sol bug that can delete a user's entire home directory
OpenAI's Tibo Sottiaux confirmed a GPT-5.6 Sol failure mode in which the model overrides the $HOME environment variable to point at a temp directory, fails the expansion, and recursively deletes the real $HOME during cleanup. It occurs almost exclusively in Codex's full-access mode with both the filesystem sandbox and auto-review approval disabled — but it had real casualties before disclosure, including an investor's Mac and a production database per outside reporting. OpenAI is tightening default guidance and promised a fuller post-mortem.
$HOME env-var mis-expansion that triggers the deletion
Grok Build CLI caught silently uploading entire private repos; xAI deletes the data and open-sources the tool
xAI's Grok Build coding CLI was found silently uploading full private Git repositories — history, deleted files, secrets — to a Google Cloud Storage bucket even when users opted out via the 'Improve the model' toggle. In one documented case, a 12GB test repo sent 5.1GB upstream when the task needed 192KB. The issue was disclosed July 13; on July 16 xAI responded by deleting the collected data, disabling the retention pipeline, and open-sourcing the entire CLI under Apache 2.0.
A playable Doom clone in ~2,000 lines of SQL, one-shotted by GPT-5.6 Sol Ultra
For show-and-tell, Peter Gostev demoed DOOMQL, a playable Doom-like game built almost entirely in ~2,000 lines of SQL by GPT-5.6 Sol Ultra, essentially in one shot — plus a companion Minecraft clone written in Lean. A live illustration of how far frontier coding models now push into languages nobody writes games in.
