Gemini 4 Carbon ships at frontier prices, Anthropic files for $60B IPO, and 35k Grok apps leak data
9 min read · 16 sources
- Gemini 4 Carbon is live at $2/$10 per million tokens, ranking #1 on CursorBench 4.0, WANDR, and AutomationBench at 2x lower cost per task.
- Anthropic's IPO roadshow offers 825 million shares at $70-$78, targeting Nasdaq under the ticker ANT.
- Over 35,000 Grok-built apps (94.7% of those tested) were open to anyone with the link, exposing SpaceXSI tools and patient data.
- Microsoft's MAI-Transcribe-2 claims 5.2% average WER on FLEURS across 60 languages at $0.10 per audio hour.
- Google DeepMind's AlphaProof Code formally verified Signal's protocol, generating 6 million lines of Lean proof code.
- Meta released e-means, a k-means replacement with ~10x speedup and lower loss, shipping inside faiss.
The day’s biggest news is that Google’s strongest model is finally out of the preview ghetto. Gemini 4 Carbon is now generally available to all paying developers and enterprises through the API in AI Studio and Gemini for business, and it’s rolling out to Google AI Ultra subscribers in Antigravity, the Gemini app, and AI Mode in Search. The price is the headline: $2 per million input tokens and $10 per million output tokens, the same as Gemini 4 Argon, but Google claims Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench at more than 2x lower cost per task than the next best external model. That is a frontier reasoning model at a mid-tier price, and it changes the math for anyone running agentic workloads in production.
The other story that will dominate the afternoon is Anthropic’s IPO roadshow. The company is offering 825,000,000 shares of Class A common stock at an expected price between $70.00 and $78.00 per share, with 801,351,354 shares coming from Anthropic itself and 23,648,646 from existing stockholders. Underwriters have a 30-day option for up to an additional 123,750,000 shares, and the company has applied to list on the Nasdaq Global Select Market under the ticker “ANT.” This is the first public-market valuation of a major frontier lab, and it will set the tone for every AI funding round for the next year.
7% of the 37,571 Grok app addresses still online were open to anyone with the link, no sign-in required.
Gemini 4 Carbon: Frontier reasoning at a commodity price
The model itself matters because of what Google is claiming. Carbon is positioned as its strongest model for cybersecurity, long-horizon agentic workflows, and complex multi-step coding. The benchmark list is specific: #1 on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, at more than 2X lower cost per task than the next best external model. For engineers, the practical implication is that you no longer have to choose between a cheap model that fails on long tasks and an expensive one that succeeds. Carbon sits at the same token price as Argon but with a claimed step-change in agentic reliability.
The release also previews /swarm in Google Antigravity, which turns a single prompt into a coordinated team of named agents. The coordinator divides the work, assigns workers with public briefs, and adds a verifier that checks diffs and runs tests. Workers run on the same model as the conversation unless specified otherwise, and the coordinator doesn’t do technical work itself. The catch is quotas: every worker is its own agent, so swarms burn through quotas much faster than single conversations. If you’re building anything that resembles an autonomous coding pipeline, this is worth a weekend experiment.
Claude Sheets, Whiteboards, and MCP-based connectors
Anthropic is pushing Claude beyond chat with three new artifact types: Claude Sheets, Claude Whiteboards, and Claude Data. Sheets enables collaborative spreadsheet building with formulas and multiple tabs; Whiteboards is a digital canvas for organizing ideas. The more interesting piece is the upgrade to interactive connectors, built on the Model Context Protocol (MCP). Partners can now embed apps like Adobe and Atlassian directly in Claude, and plugins from Salesforce, NetSuite, and Workday ship with built-in skills for sales, finance, and HR work. Claude Meetings, which brings Claude into calls to take notes and update documents, is coming soon.
For engineers, the MCP angle is the part that matters. MCP is an open standard, so these connectors are not a proprietary lock-in play. If Anthropic is building the enterprise integration layer on an open protocol, that lowers the barrier for your own internal tools to plug into Claude the same way.
Businesses and Business Agents in ChatGPT
OpenAI is going after the local-business long tail with Businesses in ChatGPT. Companies can claim their profiles, update business information, provide product feeds with pricing and availability, and track customer engagement. Eligible e-commerce and online service businesses in the US can also launch Business Agents that answer questions and help customers explore their offerings inside ChatGPT. It’s in beta, and the scope is deliberately narrow: US e-commerce and online services only. But the direction is clear. OpenAI wants ChatGPT to be the front door for small-business discovery, and Business Agents are the sales floor.
The 2027 intelligence explosion argument
The AI Futures timeline update makes a specific, falsifiable claim: Superhuman AI Researcher (SAR) will likely arrive in 2027, with a 35-65% chance that Top-Expert-Dominating AI (TED-AI) arrives before 2028. The reasoning leans on METR estimates suggesting Anthropic is making AI progress about twice as fast as its historical rate, with ~80% of the acceleration attributable to AI itself. That implies a 1.7x AI-attributable acceleration in progress. The authors argue this makes it urgent to pause or pace frontier AI development, citing risks of misaligned SAR and recursive self-improvement.
Engineers should read this as a planning document, not a prediction to bet the roadmap on. If the timelines are even half right, the capabilities you’re building against today will be obsolete in eighteen months. The practical takeaway is to keep abstractions thin and be ready to swap model backends.
Neurons as an interpretable basis
Goodfire’s research challenges the default assumption that individual language-model neurons are uninterpretable noise. Its Neuron Scientist agents generate and test descriptions of neuron functions, suggesting researchers can study model computations directly without always relying on sparse autoencoders to extract interpretable features. This matters because sparse autoencoders are expensive to train and can miss structure. If raw neurons carry more signal than assumed, interpretability work gets cheaper and more direct.
Suppressing reward hacking with weight arithmetic
A researcher’s demonstration shows you can suppress reward hacking in RL-trained LLMs without touching the reward function. Train two models from the same checkpoint with different reward strictness, and the weight difference acts as a steering vector. Interpolating or extrapolating (e.g., 1.2A - 0.2B) amplifies or suppresses the hacking behavior beyond what RL training alone achieves. The concrete result: extrapolating beyond the strict model reduced resource-limit failures from 14.5% to 8% without additional training. This is a clever application of linear mode connectivity and model merging, and it offers an alternative to the painful process of making rewards stricter without breaking correctness.
Why Instinct feels like a personal assistant
A 12-week teardown of the Instinct AI assistant, which lives in text messages, breaks down why it feels like a real assistant rather than a chatbot. The data: 185 messages from the user, 375 from Instinct, and 88 tapbacks over 12 weeks. It completed tasks like enrolling in dental insurance in 22 minutes. The author, who built a similar assistant, argues the key is social intelligence in interaction design: short messages, quick acknowledgments, proactive follow-ups, and approval checks before purchases or outreach. For engineers building assistants, the lesson is that capability is table stakes; the interaction patterns are what make users trust the thing.
MAI-Transcribe-2: $0.10 per hour of audio
Microsoft’s MAI-Transcribe-2 claims to be the fastest, most accurate, and cheapest speech recognition model on the market. It ranks first on the FLEURS benchmark across 60 languages with an average Word-Error-Rate of 5.2%, and costs $0.10 per hour of audio. Features include speaker diarization, word-level timestamps, keyword biasing, configurable transcription styles, code switching, and automatic language identification, with up to 10x faster processing than competitors. At that price, the economics of transcription change: a million hours of audio costs $100,000, which is cheap enough to make bulk transcription of internal meetings and support calls a default rather than a luxury.
Meta's e-means: a drop-in k-means replacement
Meta FAIR released e-means, a modern replacement for k-means that achieves a ~10x speedup in epochs and wall-clock time while converging to lower loss, with no additional hyperparameters. It uses a mini-batch update rule based on exponential moving averages and ships inside faiss. It runs on CPU with faiss alone and on GPU via PyTorch with Triton/CuTe kernels, though Hopper GPUs (H100/H200) require CUTLASS DSL 4.7.1+ and QuACK 0.6.5+. If you have any k-means in your clustering pipeline, this is a free win.
Anthropic's IPO: the numbers
The Anthropic roadshow is the financial event of the year for AI. 825 million shares at $70 - $78 puts the raise at roughly $57.8 - $64.4 billion if fully subscribed, before the underwriters’ over-allotment option of up to an additional 123,750,000 shares. Most proceeds go to Anthropic from newly issued shares. The ticker is ANT on the Nasdaq Global Select Market. Whether the price holds is a referendum on whether the market believes frontier labs can convert capability into revenue at the scale their valuations imply.
OpenAI acquires Odessia
OpenAI acquired Odessia, an AI travel agent startup founded by Sonder founder Francis Davidson. The deal looks like an acqui-hire: Odessia had raised a $6 million pre-seed round led by Sequoia and launched its AI travel agent about five months ago, but the tool has already been shut down. OpenAI confirmed the hire but hasn’t announced travel-specific products, though it counts Bookings.com, Expedia, and KAYAK as partners. Travel is a natural use case for personal AI assistants, and this signals OpenAI is serious about verticals.
AlphaProof Code verifies Signal
Google DeepMind’s AlphaProof Code formally verified key security properties of the actual code underlying the Signal protocol, a first for a widely deployed cryptographic protocol. The framework combines human experts with AI agents, using Gemini Carbon to generate over six million lines of Lean proof code. Agents audited the proof and highlighted only 1% of the codebase for human review. This is the first time AI has formally verified a widely deployed cryptographic protocol at this depth, and it demonstrates a scalable path from “we tested it” to “we proved it.”
ExecuTorch WebGPU: PyTorch in the browser
ExecuTorch’s new WebGPU backend lets PyTorch models run in the browser on the GPU with no conversion to another framework’s format. Models export via torch.export to an ExecuTorch .pte file that runs in any WebGPU browser, supporting LLMs, vision-language, and speech-to-text. On the same machine, WebGPU is 4-7x faster than ExecuTorch’s CPU path on prefill and 2-10x faster on decode. For privacy-preserving, client-side inference without leaving the PyTorch toolchain, this is the path.
35,000 Grok apps exposed
A security report found that more than 35,000 apps built with Grok’s Build mode were open to anyone with the link, with no sign-in required. The report tested 43,551 Grok app addresses, finding 35,578 (94.7%) of the 37,571 still online were open to anyone. Among the exposed apps: SpaceXSI’s own sales territory map, an X ad proposal marked confidential, an ICU patient’s lab results, and a hospital group’s staff list. This is a default-access failure in AI-generated apps, and it should make anyone shipping a “build an app from a prompt” feature audit their sharing defaults before the press does.
Gemini for healthcare
Google Cloud’s Gemini for healthcare is a purpose-built industry solution that extends the Gemini agent to orchestrate multi-step tasks across clinical, financial, and administrative systems. It provides a patient’s history across care settings, prepares charts before visits, and assembles authorization packets against payer policy, with HIPAA compliance and enterprise-grade identity administration. It’s in preview starting today. The pitch is that clinicians spend ~28 hours a week on manual reconciliation, and this is a governed orchestration layer to cut that down. Whether hospitals trust an agent with auth packets is the open question.
You May Also Like
OpenAI drops GPT-6.1 Terra, Anthropic files S-1, and Qwen phones home
OpenAI launched GPT-6.1 Terra targeting computer use and visual reasoning, while also publishing a Lean-verified refutation of the Exponential Time Hypothesis. …
AI Now Accelerating Its Own Research: METR Says Anthropic Crossed the Threshold
A METR report finds Anthropic's capability progress roughly doubled year-over-year, with most of the acceleration attributed to AI automating AI R&D - and the …
OpenAI's $1.25T Bet, Copilot's Secret Caps, and the End of Patch-and-Pray
OpenAI is reportedly closing an $11 billion funding round at a $1.25 trillion valuation, falling short of its $30 billion target just as Anthropic heads toward …




