In partnership with
In Today’s Issue:
🔓 Meta reopens the weights, and promises its frontier model next
🏋️ An AI agent hacked a gym while booking a class
🧮 Claude pushed a stubborn Riemann bound from 41.6% to 67.2%
🤖 China's robot champion goes public at about $9 billion
🎬 LTX-2.5 ships open weights and faster-than-real-time video
✨ And more AI goodness…
⚡ The Signal
Meta just put its most permissive license ever on its best local model, and promised the frontier one next.
Muse Glimmer is a 30B model under Apache 2.0, downloadable today and running on a single consumer GPU, and Mark Zuckerberg says the weights to Muse Spark 1.2, the frontier model behind Meta's own coding agent, follow soon. Later the same day, Anthropic made Claude Sonnet 5's launch price permanent, keeping $2 and $10 per million tokens past the August 31 expiry it had published. OpenAI moved the other way, splitting its Daybreak security program into two tiers and gating the new GPT-5.6-Cyber model behind vetting. Three labs, one day, three answers to the same question: which layer do you give away to protect the layer that pays. Meta is the only one that decided the model itself is the giveaway.
All the best,
Kim Isenberg
(ABC News: Billy Draper)
🏋️ An AI Assistant Booked a Gym Class and Hacked It
A man in Australia asked his AI assistant to book a gym class. It found a hole in the booking software and removed a stranger from the waitlist.
Andrew was running the open-source OpenClaw agent on Anthropic's Claude. It found it could book months beyond the gym's limit, then reported that "The API has zero authorisations checks on cancelling other people's reservations" and that it had already tested that on the person at waitlist position #1. Asked to undo it, the agent answered "I can't add them back." The ABC calls it the first known autonomous cyber attack in Australia.
👉 tl;dr: Nobody wrote an exploit. An errand found one.
(Anthropic)
🧮 Claude Pushed a Riemann Bound From 41.6% to 67.2%
An unreleased research version of Claude improved a longstanding lower bound on the Riemann zeta function.
Nobody has proved the Riemann hypothesis, so mathematicians instead prove that at least some fraction of the zeta function's zeros behave as it predicts. Claude raised that fraction from 41.6% to 67.2%, combining recent work by Baluyot, Goldston, Suriajaya and Turnage-Butterbaugh with a 2000 paper by Bombieri. It took two sessions, 31 million output tokens, roughly 650 failed ideas and 60 coordinated subagents, and was checked by Anthropic mathematicians plus outside number theorists Brian Conrey and Dan Goldston.
👉 tl;dr: Anthropic does not expect these techniques to prove the hypothesis. The best known bound still moved.
(TechCrunch)
🛡️ OpenAI Splits Its Cyber Program Into Blue and Red
OpenAI restructured Daybreak, its security program, into two access tiers and put a purpose-trained cyber model behind the tighter one.
Daybreak Blue opens frontier models including GPT-5.6 Sol to approved defenders for incident response, malware analysis and patch validation. Daybreak Red gates the new GPT-5.6-Cyber behind stricter vetting for vulnerability research and exploit validation, and is limited for now to trusted partners including Accenture, IBM, CrowdStrike and Cloudflare. OpenAI's reason: "defenders have a narrowing window to prepare."
👉 tl;dr: The lab selling the defense also ships the capability that makes it necessary.
🎬 Watch This
Ezra Klein talks to Jasmine Sun, just back from reporting through the Midwest on every side of the data center fight. The episode's numbers carry it: New York Governor Kathy Hochul has imposed a one-year moratorium on data center construction, and more than 100 other local and statewide moratorium proposals are moving across the country. Klein's argument is that AI has finally run into something it cannot scale past, which is the ordinary politics of building very large things where people live. The real question: how much of the backlash is about the buildings, and how much is about AI itself.
"We're making Claude Sonnet 5's introductory pricing permanent."
– Claude (@claudeai), Anthropic's official product account
Sonnet 5 launched in June at $2 per million input tokens and $10 per million output tokens on the API, with that price published as running only through August 31. Making it the standing rate is what a price war looks like once it reaches the mid-tier models that carry most production traffic.
Mark Zuckerberg says Meta will "soon" release the weights for Muse Spark 1.2, the frontier model behind its Muse Code agent. No date has been named yet, but on Glimmer's terms that would put frontier-class weights in public hands for free.
Meta's Big Open Source Comeback
The Takeaway
👉 Meta shipped Muse Glimmer on August 10: a 30B dense model under Apache 2.0, with no user cap and no "Built with Llama" branding requirement attached.
👉 It runs locally at under 20GB quantized, down from over 55GB at full precision, and landed day one on Ollama, LM Studio, vLLM and Together AI.
👉 It leads both size-class rivals on most of Meta's agentic suite, including 75.5 to 54.2 and 62.5 on MCP Atlas and 51.2 to 36.9 and 50.2 on SWE-Bench Pro.
👉 Zuckerberg says the weights to Muse Spark 1.2, Meta's frontier model, come next. On the Artificial Analysis Openness Index, Meta has already jumped from 5.0 to 8.0.
Meta spent the Llama years narrowing what it would give away. On Monday it reversed course, and picked about the most permissive license there is to do it.
Muse Glimmer is a 30-billion-parameter dense model, released August 10 under Apache 2.0, from the Meta Superintelligence Labs group that Zuckerberg credited by name in the announcement. Dense matters for a local model: every parameter fires on every token, so memory use is predictable and there is no expert-routing to tune. It takes text and images in, was trained on more than 100 languages, carries a 128K context window, and is tuned for the unglamorous parts of agent work: calling tools, chaining multi-step tasks, recovering when a step fails, and acting as an LLM-as-judge, the pattern where one model grades another model's output. Weights are on Hugging Face in BF16 alongside GGUF and ExecuTorch builds, with day-one support in Ollama, LM Studio, vLLM and Together AI. K-Quant compression takes it from over 55GB at full precision to under 20GB, which is one 24GB consumer GPU or a high-end Mac. Meta also ships DFlash, a speculative decoding drafter that has a small model guess ahead and the large one check, claiming 3.1x faster generation on an RTX 5090, 1.8x on an M5 Max and 1.5x on an M4 Max.
(Artificial Analysis)
Glimmer leads both size-class rivals on most of Meta's own agentic suite: 75.5 to 54.2 and 62.5 on MCP Atlas, 51.2 to 36.9 and 50.2 on SWE-Bench Pro, plus wins on DeepSearch QA, WildClawBench and GAIA2. Qwen3.6-27B holds two, OSWorld-Verified at 75.6 to 65.9 and TerminalBench 2.1. Artificial Analysis, which averages a fixed benchmark suite into one number, is more measured: Glimmer lands at 35 on its Intelligence Index, five points above Gemma 4 31B and a few behind Qwen3.6 27B at the same size, with an 82% hallucination rate against 49%, and 953 Elo on GDPval-AA v2 against a 1,000 human baseline. Read together, that is a top-tier local agent model doing the job Meta built it for.
(Artificial Analysis)
The license is the biggest part of this. Apache 2.0 replaces the Llama Community License, which capped free commercial use at 700 million monthly users, made every derivative carry a "Built with Llama" label, and shut EU companies out of the vision features. All of that is gone: fine-tune Glimmer on your own data, ship it in a product, rebrand it, and Meta has no say. Artificial Analysis, an independent benchmarking outfit, tracks precisely this with an Openness Index that scores four things out of a possible 18: how freely you can obtain and use the weights, and how much the lab publishes about its methodology, its pre-training data and its post-training data. Glimmer scores 8.0, against 5.0 for Llama 4 Maverick, and level with DeepSeek V4 Flash and GLM-5.2. That is Meta measured against its own past rather than a rival: Maverick shipped in April 2025 as a 400-billion-parameter mixture-of-experts model you needed server hardware to run. Sixteen months on, Meta's open model fits on one GPU and comes with far fewer strings. The 8.0 is all availability and methodology, though. Meta still publishes nothing about what Glimmer was trained on, which is why Olmo 3 32B Think, the Allen Institute model that does release its data, sits at 16.0.
Meta's money comes from advertising, so giving a model away costs it far less than it costs a lab that bills per token. OpenAI is gating GPT-5.6-Cyber behind vetting. Anthropic has just locked in a lower Sonnet 5 price to defend volume. Meta is buying the default position in the local agent stack that both of them are trying to charge for. The manifesto published alongside it, The Future Is For Everyone, promises personal agents running 24/7, a private mode Meta says it cannot read into, and independent board oversight of release criteria. Those are commitments rather than shipped features, and today's release is the first thing Meta has done that makes them credible.
Why it matters: Apache 2.0 on a 30B agent model that runs on one GPU resets the floor for everyone building local agents. They can start from Meta's weights for free, fine-tune on their own data, and ship without a usage cap or a license review. The bigger prize is Muse Spark 1.2: if Meta puts frontier weights out on the same terms, every closed mid-tier model has to justify its price against something free.
Sources:
🔗 https://www.meta.com/thefutureisforeveryone/
🔗 https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
🔗 https://artificialanalysis.ai/articles/muse-glimmer
The GTM Playbook Behind Warmly's Acquisition
Warmly ran pipeline, outreach, and lead scoring on autopilot for hundreds of startups — before a single sales hire.
HubSpot acquired them for it. Now the cofounders are walking you through the exact system, live, before they disappear into product. Join the Builder Session on August 12.
Leave with an agentic GTM stack you can replicate this week. Plus HubSpot Credits when you join HubSpot for Startups.
The chart: LTX's launch benchmarks for LTX-2.5, the open-weights video model it is releasing today. The bars are a blind preference test: people watch two clips from the same prompt, positions randomized, and vote without knowing which model made which. LTX-2.5 takes 67%, ahead of Seedance-2.5 at 65%, then Gemini Omni Flash (55%), MiniMax-h3 (50%), Seedance-2.0 (44%), Wan-2.6 (42%) and Flux-3 (28%).
The lesson: Two points separate the top two, so on raw quality this is a tie at the frontier. The companion table is where LTX-2.5 breaks away: Seedance 2.5, Kling 3.0 Pro, Veo 3.1 and Grok 1.5 are API-only, with no downloadable weights, no on-prem deployment and no fine-tuning on your own footage. LTX ships all three, plus 10 seconds of 720p video in 6.8 seconds self-hosted on two GB200s, which is faster than real time. For a studio that cannot send client material to someone else's server, that decides the purchase.
The caveat: These are LTX's own launch numbers, and LTX labels the preference results "preliminary results" it expects to move as evaluation expands. A vendor picks the prompts, the comparison set and the evaluator pool, so treat the 67% to 65% margin over Seedance 2.5 as a tie until an independent arena runs it. The open-weights, on-prem and fine-tuning rows are structural claims, and checkable the moment the weights are public.
🤖 China's Robot Champion Goes Public, and the Margins Are Already Slipping
⚡ Bottom line: Unitree Robotics opened subscriptions on August 10 for a Shanghai STAR Market listing valuing it near $9 billion.
💡 Why it matters: It is the first profitable humanoid maker to open its books to public markets, and it is Chinese.
🔎 What it means: Quarterly filings will show what humanoid economics really look like, replacing demo videos as the sector's price signal.
Unitree priced its STAR Market IPO on August 6 at 150.80 yuan a share, selling 40.45 million shares, or 10% of its post-offering capital. That raises about 6.1 billion yuan, roughly $904 million, at a valuation near 61 billion yuan, about $9 billion. Subscriptions opened August 10 and close August 12, with CITIC Securities leading. Unitree will be the first humanoid robot maker to list on a mainland Chinese exchange.
Unitree's lineup, from H-series humanoids to the quadrupeds that still carry the business. (Gasgoo)
Unitree is profitable, which almost nobody else in humanoids is. 2025 revenue was 1.699 billion yuan at a 60.13% gross margin on core businesses, over 40% from outside China. It shipped more than 5,500 humanoids that year, more than any other company, on top of 33,000+ quadrupeds cumulatively. The robot dogs are why the humanoids exist: they sell today, at volume, and fund the rest.
A Unitree G1 handling a pan in a company demo. Manipulation is still where humanoid demos break. (Unitree)
The prospectus carries the number that should slow everyone down. Unitree forecasts first-half 2026 revenue of 1.05 to 1.13 billion yuan, up 35.6% to 45.4%, alongside adjusted net profit of 236 to 283 million yuan, down 6.4% to 22.0%. Revenue climbing while adjusted profit falls is what it costs to move from quadrupeds into humanoids, where R&D per unit is higher and volumes lower. The offer price already implies a diluted 2025 price-to-earnings ratio of 219, and nearly half the proceeds go to embodied intelligence models rather than hardware.
Figure, 1X and Physical Intelligence are valued privately, on projections. Unitree will report quarterly, against a peer set that can read its margins line by line. Humanoids have been priced on demo videos so far, and filings are a harder audience.
The best voice models, now with full orchestration. Build real-time voice and chat agents on one low-latency stack: any LLM, your tools and knowledge, testing, Guardrails, and omnichannel deployment.