π€π» AI Daily Digest β August 5, 2026
OpenAI Crosses 1 Billion Active Users β and Codex Now Does 99.8% of Its Output Tokens
OpenAI crossed a number that took ChatGPT less than four years to reach. In a July 31 blog post, the company said its models now touch more than 1 billion active users and over 2 million businesses. The milestone came with a claim about usage depth: after about six weeks, users send roughly 50% more messages per day and use ChatGPT for about twice as many kinds of work. The post frames the growth as a consequence of falling prices β "when the cost of useful intelligence falls, more work becomes worth doing."
The pricing moves backing that line landed days earlier. OpenAI cut GPT-5.6 Luna by 80% (now $0.20 per million input tokens, $1.20 output) and GPT-5.6 Terra by 20% ($2/$12), while Sol pricing stayed put. Internal work on Sol is said to have cut end-to-end serving costs by 20% and lifted token-generation efficiency by over 15%; the company also credits better context management and retained reasoning for moving Sol's ARC-AGI-3 score from 13.3% to 38.3% while using six times fewer output tokens.
The stat worth holding onto: agentic work through Codex now accounts for 99.8% of OpenAI's weekly output tokens. The company is effectively describing itself as an agent company now, not a chatbot company β and the billion-user number is the growth story behind its march toward an IPO.
β OpenAI Β· TechRepublic
π OpenAI β Building Abundant Intelligence Β· TechRepublic
Gemini Robotics 2 Brings Whole-Body Intelligence to Robots
Google DeepMind announced Gemini Robotics 2 on July 30, a model family positioned as "the intelligence layer to power any kind of robot." The headline change over the first version is whole-body control: the VLA (vision-language-action) model coordinates movement from toe to fingertip instead of mostly handling the upper body. On an Apptronik Apollo 2 humanoid, the system walked, crouched, picked objects off shelves, and reasoned about multi-step tasks in real time.
Two sibling models widen the range. Gemini Robotics ER 2 is an embodied-reasoning model that watches video, builds a plan, and makes hundreds of decisions during tasks lasting several minutes β it can track progress, resume from the last correct step after an error, split long tasks across multiple robots, and call tools like Google Search to clarify ambiguous instructions. Gemini Robotics On-Device 2 runs locally on the robot, and DeepMind says it adapts to a new dual-arm platform with fewer than 200 real-world examples and a few hours of training. Safety got a benchmark of its own, ASIMOV-Agentic, for evaluating whether a robot refuses unsafe commands and asks for help when unsure; ER 2 also detects nearby people and can trigger a safety stop.
The numbers from the announcement give a fair picture of where whole-body dexterity actually stands. Lifting objects succeeded at 68.4% from a table, 45.7% from the floor, and 76.3% from a shelf; a 22-degree-of-freedom five-fingered hand removed light bulbs at 92%, but installing them landed at 36%. DeepMind is honest about the gap β human-level dexterity is the stated next target. The direction is clear either way: robot makers are consolidating on one general brain instead of a pile of task-specific controllers.
β Google DeepMind Β· The Paper
π Google DeepMind β Gemini Robotics 2 Β· The Paper Coverage
NVIDIA Opens the First Full-Duplex Voice Model That Calls Tools While Talking
NVIDIA released NemotronLabs VoiceChat on August 3 β an 11-billion-parameter speech model with open weights on Hugging Face, and the first open full-duplex model that calls tools mid-conversation. "Full duplex" means it listens and speaks at the same time: no wake word required, users can interrupt mid-sentence, and the model adjusts its output on the fly. The stack collapses the usual ASR β LLM β TTS relay into one streaming network: a Fast Conformer speech encoder feeds a Nemotron Nano v2 9B backbone, a TTS decoder emits speech, and a separate output channel produces tool-calling scripts without contaminating the spoken response.
The numbers put it in context. Turn-taking latency is around 450ms; barge-in resolution about 480ms; it ranks second among open full-duplex models on VoiceBench. On the BFCL-v3 tool-calling suite it averages 56.1% β but the breakdown is the honest part: 82.5% on picking the right tool, 44.2% on getting the arguments right, 33% pass@1. NVIDIA recommends no more than five tools per session and notes parallel calls are unreliable. This is a first-generation open attempt, not a drop-in replacement for a hosted voice API.
The strategic read matters more. Closed realtime voice APIs were the only place to see duplex conversation and agentic tool use working together; now there is an open reference implementation on vLLM (A100 through B200) that researchers can poke at. The catch is the OpenMDW v1.1 license, which limits use to research β so treat it as a blueprint to study rather than a stack to ship.
β NVIDIA Β· Artificial Analysis
π NVIDIA β NemotronLabs VoiceChat 11B (Hugging Face) Β· Artificial Analysis
OpenAI Says "Apple Is Getting This Wrong" in Trade-Secrets Fight
OpenAI published its most detailed response yet to Apple's trade-secrets lawsuit on August 3, under a title that does not mince words: "Apple Is Getting This Wrong." Apple sued on July 10, accusing former hardware executives β including Tang Tan, now OpenAI's chief hardware officer β of taking confidential information into OpenAI's consumer hardware business. The case sits in the Northern District of California, and Apple recently moved for a preliminary injunction.
The response attacks the timeline. Apple claimed it contacted OpenAI in February and got no reply; OpenAI released emails showing Apple's outside counsel mixed up two Asian surnames and sent the note to the wrong person, then admitted the error. OpenAI says it then heard nothing for five months until the lawsuit landed. On former engineer Chang Liu, OpenAI published iMessage records showing Apple employees contacted Liu after his January departure to retrieve project files β and argued his continued access came from Apple's failure to clean up iCloud sharing permissions, which the company later reframed as "residual access."
OpenAI calls the injunction request "built on false information" and says it neither possesses nor wants Apple's trade secrets. Whatever the merits, this reads as a turf war over the next generation of native AI hardware β Apple defending its supply-chain moat while OpenAI needs to show its hardware effort stands on its own. The court has not ruled on the injunction.
β OpenAI Β· MacObserver
π OpenAI β Apple Is Getting This Wrong Β· MacObserver
Tuya AI Coding Turns One Sentence Into an App That Talks to Real Devices
Tuya Smart, the AIoT cloud platform listed on NYSE and HKEX, launched Tuya AI Coding on August 3 β an AI-native no-code platform that turns natural language into a deployable app. Type "I want an app that tracks my smart pet feeder and warns me when food runs low," and the platform generates the UI, the interaction logic, and the backend in one pass: database, API gateway, user authentication, and device management, wired into Tuya's cloud infrastructure across 200-plus countries.
The differentiator is the hardware layer. Most AI app generators stop at a webpage or mini-program. Tuya's version plugs natively into its device ecosystem β 100,000-plus SKUs of connected hardware β so a generated app can actually control real devices, read device stats, and trigger scenes. The target audience is deliberately non-technical: designers, product managers, founders, freelancers, students. The company says the dev cycle goes from months to minutes and the technical barrier drops by 90%.
The timing lines up with a broader shift. Gartner expects 75% of new enterprise applications to be built with low-code or no-code tools by the end of 2026. The interesting bet is this: as AI app generation commoditizes the frontend, the last mile that separates a demo from a product is connection to the physical world β and that is exactly the moat an IoT company already owns.
β Tuya Β· εζΉ+ ε ±ε°
π Tuya Developer β AI Coding Kit Β· Tuya AI Coding Β· εζΉ+ ε ±ε°
Preprint: Long-Horizon Agent Training Teaches Habits That Transfer Across Benchmarks
A preprint posted July 31 (arXiv:2608.00181) asks a question that benchmark tables usually dodge: when you train an agent on long-horizon tool use, does it learn anything that transfers? The authors post-trained an open-weight MoE model, Qwen3.5-122B-A10B, on 363 Model Context Protocol (MCP) tasks across 27 categories using a two-stage SFT-then-RL pipeline. Crucially, no external-benchmark task or grader entered training, and no external score influenced the reward.
The results argue yes. At greedy pass@1, the trained model beats the base on five external evaluations: Toolathlon +9.6pp, Ο2-Bench +5.3pp, BFCL-V4 +3.5pp, SWE-Bench Pro +5.8pp, and Terminal-Bench 2 +2.8pp. The striking one is SWE-Bench Pro β software-engineering performance improved even though the training collection contained no software-engineering tasks at all.
The paired-trajectory analysis identifies four behavioral shifts that appear across office workflows and code alike: more careful local-goal formation, building goal-relevant working state, keeping parent goals stable through local repairs, and verifying completion. The claim, put plainly, is that RL on multi-tool long-horizon tasks changes how an agent works, not just what it knows β and that way of working transfers beyond the training domain. That reads less like task memorization and more like habit formation.
β arXiv
π arXiv β Cross-Benchmark Generalization in Long-Horizon Agents
LiveEdit: Real-Time Video Editing With Text Instructions, Open-Sourced
A team from Tsinghua and HKUST released LiveEdit (arXiv:2606.26740), a framework for streaming video editing with text instructions β accepted at ECCV 2026, with training code and models open-sourced. The target is live scenarios: livestream effects, video conferencing, augmented reality, where you cannot wait for the whole video before editing starts.
The technical problem is that video diffusion models rely on bidirectional spatiotemporal attention β they need future frames before they can process a given one. Streaming means the model only sees the present and the past, and the paper shows that naively truncating future frames thins out attention across a longer history, breaking the local temporal prior and producing flicker and drift. LiveEdit processes incoming video in a causal, chunked way β 12.66 FPS at 4 steps per video chunk β while keeping edited regions accurate and unedited regions consistent. It also avoids recomputing static background tokens over and over, which is the main cost of real-time inference.
The paper's framing is worth keeping in mind: generated video can come from noise, but editing must preserve the original structure, lighting, and motion. That constraint is why streaming editing is a separate problem from streaming generation, and why most "real-time" tools to date have been anything but. This one is open source, so it is directly testable.
β arXiv Β· ζΊε¨δΉεΏ
π arXiv β LiveEdit Β· ζΊε¨δΉεΏε ±ε°
Top comments (0)