Anthropic released Claude Fable 5.1 and Mythos 5.1
They’re actually the same underlying model, but Fable is available to everyone while Mythos has fewer safeguards for cybersecurity and biology and is only available to vetted organizations. (read the announcement)
Fable got a pretty big science/coding bump. It scored 52.6% on Terminal-Bench-Science vs 24.7% for Fable 5, and Anthropic says Mythos is even better at some of the harder research stuff. It designed protein binders with a ~50% hit rate across 12 targets and sped up GPU kernels for biology models by up to 2.5x.
Pricing is still $10/$50 per million input/output tokens, but cache reads are 75% cheaper. Anthropic says that makes typical workloads ~25% cheaper and highly agentic ones up to ~45% cheaper.
Some fun Fable stuff already:
- COLD WATCH is Ethan Mollick’s retro space-survival game built with Fable 5.1. A pretty fun way to see what a long-running coding agent can actually build.
- Cat Doom is exactly what it sounds like… a Doom-style FPS with cats. Fable 5.1 built the thing live on stream.
World Labs has a new world model called Atlas
It can take images, video and camera movements and turn them into a 3D world you can move around in. You can basically give it a few photos, tell the camera where to go and generate new image/video frames from that exact viewpoint. (see the example here) (another one here)
They showed it generating a 1-minute 1440p video from seven reference images, and it can reconstruct explicit 3D from just one or a few images. More images = less guessing about what the world actually looks like.
The robotics use case is pretty cool too. Atlas can take a few photos of a real space and simulate what a robot’s cameras/depth sensors would see while moving through it. So instead of scanning every training environment with expensive equipment, you could potentially just take some photos and generate the rest.
It’s all one model, trained from scratch, that mixes ideas from LLMs and video models. World Labs says Atlas can perceive, generate and reason about both virtual + physical worlds.
Early access is coming in the next few weeks. (read the blog here)
Meta has a new real-time speech model
Muse Voice Transcribe is Meta’s first real-time audio perception model. It does speech-to-text, figures out who’s speaking and knows when someone has stopped talking, all in one model.
The clever bit is it decides how long to listen before writing each word. Meta says this gets it pretty close to the sweet spot between speed and accuracy.
It’s trained across 70+ languages, can handle people switching languages mid-sentence and apparently hour-long conversations with 20+ speakers. And it costs something like $3 per 1,000 voice minutes.
It’s already being used for dictation in Meta’s desktop app and voice input in Muse Code. Live on the API now too. (read the announcement by Zuck)
Anthropic Launched Fable 5.1 — A Leader’s Curse
Anthropic launched Fable 5.1 and Mythos 5.1 today. Terminal-Bench jumped from 24.7% to 52.6%. Mythos 5.1 hit 60.9%. GPT-5.6 Sol is at 22.4%. Restrictions loosened: safety false positives down 60%, defensive vulnerability analysis now allowed. Price unchanged from Fable 5. Cache Read dropped 75% to $0.25, cutting overall cost 25%. Agent workloads down 45%. But one task still cost $5.69. GPT-5.6 Sol ran it for $0.88. Anthropic also locked the history context to block distillation and added invisible watermarks.
Anthropic widened the gap. Fable 5.1 is “uncontested” for now. The cost cuts, the relaxed safety, the locked history. Every move protects the same thing: an intelligence premium the market is refusing to pay. Fable 5 captured 6% of enterprise usage. Customers bought the cheaper Opus 5. Anthropic can push the model another 10%. Competitors only need 90% of the performance at a tenth of the price. The leader’s curse is running faster on a treadmill nobody is watching.
Anthropic built the best model in the world, but still priced it six times higher than the alternative, and spent the entire launch defending a premium the market already walked away from.
An Empty Building Just Filed An IPO
SB Energy, an AI data center company backed by Softbank, OpenAI, and Nvidia, filed for an IPO. It has zero revenue from data centers. None of its data centers are operational. It lost $3.2 billion in the first half of 2026. The $139 million it did earn came from an old energy business. Its S-1 mentions OpenAI 306 times, nearly as many as Softbank itself. Nvidia is providing $105 billion in financing and the company wants to raise $5 to $7 billion from the public.
SB Energy is not a data center company. It is a bet on OpenAI wearing a data center costume. The revenue is zero. The customer is one. The product is not built. The risk factors read like a confession: AI adoption could stall, the technology could become obsolete, the community could revolt, and the hyperscalers could cut spending. Every worst-case scenario is in the document. The company filed it anyway. The AI boom has reached the point where you can IPO a building that does not exist, for a customer that has not moved in, and call it infrastructure.
Zero revenue. One customer. $3.2 billion in losses. The S-1 is the most honest risk disclosure document ever filed.
Drop Out If You Have No Token
A computer science professor at Nanjing University, China, told his class: students without paid API tokens should drop out immediately. The line was in red on the slide. Jiang Yanyan, an associate professor, tried to get sponsors to cover the cost. He failed. He told the students to pay themselves. The course is called Generative Software Engineering. He recommended $15 of DeepSeek V4 Flash tokens. He said the point was not burning money. The point was that a CS degree in 2026 without real API experience is a degree in the past.
Token is no longer a tool. It is a prerequisite. The rest of the CS department is still teaching compilers. Jiang told his students what the industry already decided: the interview does not ask about your textbook. It asks what you built with your token. A professor had to burn his own credibility to say what the university would not. The 15 bucks is not a fee. It is the price of admission to a degree that still means something.
A professor told his students to drop out if they had no token. It was not a threat. It was the most honest thing a CS professor said in 2026.
The Climate Cheat Just Died
Google and NASA JPL released MAPL-EMIT, a Vision Transformer that detects methane leaks from space. NASA’s EMIT satellite already monitored methane, but its matched-filter method missed half the plumes and could not separate overlapping sources. MAPL-EMIT was trained on 3.6 million synthetic plumes. It scored 84% recall and found 50% more plumes than NASA’s own benchmark. It disentangles overlapping plumes and traces each back to its source. It detected 24 of the world’s 25 largest landfills. The model, data, and code are open.
Methane is 30 times more potent than CO2 and drives 25% of global warming. The defense was always the same: prove it was mine. The old system could not. The new one can separate your plume from your neighbor’s, name the facility, and publish the record. The 125 countries that signed the Global Methane Pledge now have a cop in the sky. The cop is free, automatic, and does not need a warrant.
The cheapest climate cheat was blurry satellite data. AI sharpened it, and half the leaks that were invisible yesterday are public today.
Tesla Reboots the Past
Tesla plans to re-launch the Roadster, Semi, and Powerwall in September, all products originally announced roughly a decade ago. Analysts suggest the move is primarily aimed at boosting the stock price rather than signaling meaningful production capacity for these long-delayed vehicles.
Anthropic Ships Fable 5.1
Anthropic released Claude Fable 5.1, calling it their best model for coding, data analysis, computer use, and long-running agentic work. The model comes with a 75% cache read price cut to $0.25 per million tokens, positioning it as a direct challenger to OpenAI’s upcoming Astra release.
SB Energy Files IPO, Leans on OpenAI
SoftBank’s SB Energy has filed for an IPO, with its prospectus revealing the AI data center infrastructure company is ‘substantially dependent’ on OpenAI and NVIDIA. The filing highlights growing concentration risk in AI supply chains as major players build infrastructure around a handful of model providers.
NVIDIA DLSS 5 Drops September 3rd
NVIDIA’s DLSS 5 launches September 3rd with NBA 2K27 as its first supported title. The new AI upscaling architecture requires RTX 50-series GPUs at minimum, sparking controversy over backward compatibility as the company pushes users toward its latest hardware generation.
Manus Resumes Solo After Meta Deal Dies
Chinese AI agent startup Manus has resumed independent operations after its $2 billion acquisition deal with Meta collapsed. The company is now seeking fresh funding to continue building its autonomous AI agent platform, marking a significant setback for Meta’s AI agent ambitions in China.
OpenAI Astra Hits ‘Critical’ Cyber Level
OpenAI is preparing to release Astra, its most capable AI model yet, which has reached the ‘Critical’ threshold for cybersecurity capabilities under the company’s Preparedness Framework. Select partners will get early access to shore up defenses before broader availability, marking OpenAI’s first acknowledgment of a model achieving this danger tier.
Google Launches Pics, an AI Design Tool
Google has launched Pics, an AI-powered image generation and editing tool for Workspace users that directly competes with Canva. The product integrates generative AI into the enterprise productivity suite, signaling Google’s push to own the creative workflow layer alongside docs and slides.
Félix Raises $200M From a16z, GenCat
Latin American WhatsApp-based remittance startup Félix has raised a $200 million Series C co-led by Andreessen Horowitz and General Catalyst. The round underscores continued top-tier VC appetite for fintech infrastructure serving cross-border payment corridors in emerging markets.
Sequoia-Backed Empirik Launches at $21M
Empirik, incubated by Sequoia Capital, has launched with $21 million in funding to build AI-powered infrastructure observability that predicts outages before they happen. The startup targets enterprise reliability teams looking to move from reactive incident response to predictive prevention.
HiBob Raises $166M at $400M+ Revenue
HR software company HiBob has raised $166 million after its annualized revenue surpassed $400 million. The cloud-based HR platform continues consolidating the mid-market enterprise HR space, competing against established players like Workday and BambooHR.
Aranya Turns Bare Metal Into AI Clusters
Aranya has raised $11 million in seed funding for software that transforms bare-metal servers into AI training and inference clusters in under 48 hours. The startup targets the fragmented GPU compute market where organizations struggle to efficiently orchestrate heterogeneous hardware for AI workloads.
SciFin Debuts With $44M for Revenue AI
SciFin has launched with $44 million in funding to help B2B revenue teams close what it calls the ‘context gap’ between customer intelligence and sales execution. The platform uses AI to unify scattered customer data into actionable revenue workflows.
Lilly’s Biotech Deals Fuel VC Returns
Eli Lilly’s aggressive 2026 acquisition spree across multiple biotech startups is generating strong exit returns for venture investors, boosting confidence in the biotech funding ecosystem. The pharma giant’s dealmaking pattern signals sustained demand for innovative therapeutic assets.