Four throughlines cut through today’s edition. The price of intelligence keeps climbing and few switch providers over it: Anthropic’s annualized revenue hit $65 billion, Vercel found it billing 4.4 times the average per token while capturing 65 percent of gateway spend, and Groq raised $350 million at $3.5 billion, half its pre-Nvidia valuation.
The Price of Intelligence: Who Actually Pays for Compute
Money is moving through the AI stack in stranger ways than headline valuations suggest, from premium token pricing to an architecture that turns cheap shared inference into a per-user GPU bill.
- Groq Raises $350M at Half Its Pre-Nvidia Valuation. Groq closed a $350 million round at a $3.5 billion valuation, half of what Nvidia’s licensing deal implied eight months ago when Nvidia hired away Groq’s founding CEO and president.
- Test-Time Training Could Make Inference as Costly as Training. A model architecture that updates its own weights per prompt trades cheap shared inference for a GPU bill that scales with users rather than context length.
- Vercel Data: Anthropic Takes 65% of Gateway Spend on 30% of Tokens. Vercel’s July index shows Anthropic charging 4.4 times the average price per token, and developers are paying the premium without switching providers.
- A Faster Model Isn’t a Better Answer. It’s Budget to Spend. OpenAI’s Ultrafast preview on Cerebras shows raw decode speed only pays off when builders spend the saved time on verification instead of banking it as a win.
Measurement Is Breaking: When Benchmarks Stop Measuring Skill
Three evaluations this week found that impressive numbers do not always mean impressive systems, from staged agent pipelines to video models with no taste.
- MIT and Harvard Find Multi-Agent Pipelines Fake Their Gains. Role Anchor found that 86 percent of a pipeline’s reinforcement-learning gains disappeared once a module was forced to stop faking its assigned role.
- Magic Hour Tests Fable 5 and Sol 5.6 on Ad-Making Taste, Both Flop. The AI video company’s own harness scored two frontier models on four 15-second ads and found neither has taste worth trusting yet.
- A Benchmark That Grades AI on Finding Out, Not Knowing. dig.bench drops models into 70 unexplained text games and scores whether they can discover the rules themselves rather than recall a memorized answer.
Who Controls the Stack: Forges, Memory, and Open Weights
Control over the agentic stack is consolidating in places users rarely inspect: default settings, shared memory, and the incentives behind who gives models away for free.
- Cursor Ships Origin to Paid Users Hours Before GitHub Went Dark. The June announcement promised an agent-native forge, but the actual rollout shipped with an opt-out default and data terms the company has not published.
- Warp’s Shared Agent Memory Is a Governance Problem in Disguise. Agent Memory lets Claude Code, Codex, and Warp’s own agent share what they learn, raising the harder question of who can read, write, and correct that shared record.
- Nvidia and Meta Are Funding Open Models for Opposite Reasons. Nathan Lambert argues Nvidia subsidizes open models to grow chip demand while Meta gives them away to squeeze rivals’ token sales.
Build vs Rent: The Enterprise AI Calculus
Two data points from opposite ends of the market are converging on the same question: when does a company build its own model instead of renting a frontier one.
- Linear’s Own Data Shows CEOs Adopting AI Faster Than Anyone Else. At companies with more than 200 employees, CEO usage of Linear’s AI features quadrupled in six months, outpacing every other role in the data.
- Sequoia Tells Founders: Rent Frontier AI, But Own Where It Counts. Sonya Huang, Pat Grady, and Sonali Singh lay out four conditions for building proprietary models, a buildout few startups can actually staff.
Quick Hits
The rest of what moved today, in one line each.
- Anthropic’s Annualized Revenue Passes $65 Billion, Bloomberg Reports. Bloomberg reports Anthropic’s run rate has climbed more than sevenfold since the end of last year, a pace that outstrips every other frontier lab’s disclosed growth curve this cycle.
- OpenRouter Halves Price on OpenAI’s GPT-5.6 Sol Through September 18. The cut applies across batch, flex, and priority tiers, dropping flex pricing as low as $1.25 in and $7.50 out, undercutting most comparable frontier-model rates through mid-September.
- On a 24GB GPU, Qwen3.8-27B Beats Gemma 4 as Context Grows. Gemma 4 31B needs a lossy cache trick to fit 64K context, while Qwen3.8 and Qwen3.6 keep more than 4 GiB of headroom at the same context length on identical hardware.
- A Formula for How Often to Recycle Scarce Training Data. New pretraining research finds the ideal repeat count for high-quality data rises with model size and can be estimated cheaply by running the formula on small proxy models first.
- Local LLM Runtimes on Apple Silicon Still Can’t Match CUDA. A community guide maps why Mac inference software still trails Nvidia’s stack, walking through which runtime and quantization choices actually hold up under real workloads today.
- OrcaRouter Publishes Refusal-Stripped Qwen3.8-27B for Apple Silicon. A third party removed Qwen’s safety tuning entirely and quantized the 27 billion parameter model into four separate MLX tiers for local Apple Silicon deployment.
- OpenAI’s ChatGPT for Teens Hinges on Guessing Who’s a Minor. A prediction model, not a birthdate typed by the user, decides who gets routed into OpenAI’s restricted teen product, leaving the actual age-verification method largely invisible to parents.