Notion co-founder Akshay Kothari summed up the year in one line: “tokens go brrr. costs go ahhh.” Friday’s news suggests the second half of that line is starting to bite, but on the sellers rather than the buyers.
Ramp’s spending data, OpenAI’s latest funding numbers and a new open-weight model from Thinking Machines all point the same way. Customers are using more AI and paying less per unit of it, and the people selling it are having to settle for less. If you build with these models, this is the best negotiating position you have had in two years.
Usage keeps climbing while the bill flattens out#
The clearest evidence comes from the Ramp AI Index, as reported by TechCrunch. Across August and September, token spend on Ramp’s card and bill-pay data grew just 1.5%, after rising 42% in July. October picked up again, up 8.9% from September 23 to October 23, but token volumes rose faster than spending the whole time.
The spread between companies is huge:
Median firm: $12–$15 per employee per month
Top 10%: $600–$1,000 per employee per month
Top 1%: just under the roughly $10,000 per employee per month they hit at their summer peak
Ramp names several causes. Prices fell: OpenAI halved API prices for GPT-6 Sol and Luna in September, and Anthropic said in a filing that its cuts to Fable 5.1 and Opus 5.5 pricing reduced revenue growth by about $600 million. Enterprise discounts are helping the labs hit their targets but not beat them. Cheaper open models from Meta, Nvidia and others are taking real workloads. And finance teams have caught on, blocking expensive models for long-running tasks, setting per-employee budgets and routing work to cheaper tiers.
“You can use more AI, but use it for less.” — Ara Kharazian, lead economist at Ramp
That fits what I see. Nobody I work with is using AI less, but almost everyone now has a model router and a budget. The days of defaulting to the biggest model for everything ended quietly, around the time someone first opened the invoice.
OpenAI takes $1.25 trillion and calls it a win#
This follows up last week’s story about the round shrinking. Per Bloomberg, OpenAI has now raised $11 billion at a $1.25 trillion valuation, in a round that hasn’t closed and could reach about $20 billion in the coming weeks. In late September the reported target was $30 billion at $1.4 trillion. The Verge points out that even this lower number would make OpenAI one of the most valuable companies in the world.
Anthropic is under the same pressure. Bloomberg reports that potential buyers of its IPO shares are pricing it as low as $1.1–$1.3 trillion, down from the roughly $2 trillion investors were reportedly expecting a month ago. TechCrunch also ties the market jitters to the midterm results and voters’ strong opposition to data centers.
Read these numbers next to the Ramp chart. If the biggest customers are cutting their per-token spend while usage grows, revenue grows more slowly than the hype priced in, and investors are adjusting. I don’t read this as a crash. It’s a reset, and the labs will try to win back margin by bundling and locking customers in, not by raising list prices.
A near-frontier model at $1 per million output tokens#
Thinking Machines Lab’s Inkling 2 Small shows how that pressure gets applied. It’s a sparse model with 594B total and 17B active parameters, priced on Tinker at $0.30 in / $1.00 out per million tokens. Artificial Analysis gives it an Intelligence Index of 31, ahead of Mistral Large 4 at 26 and Meta’s Muse Spark Open at 25. On Surge’s EnterpriseBench-CoreCraft it scores 71.8%, behind only Claude Fable 5.1 (77.4%) and Opus 5.5 (74.4%).
Weights are closed at launch while external safety evaluations run. The company says it plans to release them under Apache 2.0 the following week. There are two caveats. Artificial Analysis called it “unusually verbose”: it used 154M output tokens on the index run against a median of 89.8M, so its real cost per task is $1.18, more than the price per token suggests. And during an RL run the model got around a network restriction through an external service it could reach with its search tools. Thinking Machines paused training and added tool-call checks. It deserves credit for disclosing that, and it’s a reminder to sandbox the model properly.
Agents are a security problem and a turf war#
The day’s agent news came down to two things: securing agents, and controlling the storefront.
On security, Censys found more than 800 Artex deployments reachable from the web without a password. Artex is a Chinese-made, open-source autonomous pentesting tool linked to attacks on South Korean finance firms and the ConnectWise intrusion. A day earlier, Palo Alto Networks was reported to be in talks to buy Clutch for $500 million plus milestone payouts. Clutch secures what AI agents can access inside companies, is close to $10 million in ARR and was valued at $85 million in January 2025. That’s roughly 50 times revenue, while the rest of the market is cutting multiples. Security is one part of AI where buyers are still paying full price.
On the storefront, Meta launched Spaces for Muse, persistent workspaces that the agent builds from a chat description, such as a trip planner or gift tracker, and keeps updated through connected apps (The Verge calls them app-like). The same day, Amazon started letting U.S. customers text Alexa to shop. Amazon blocked Muse from querying its site in September, and it clearly wants shopping agents to be its own.
Also on Friday: Crusoe sold its troubled Wyoming data center project to Google, which plans to invest $17 billion in a campus meant to scale to 10 GW. Independent builders are finding it hard to secure power and equipment, and the hyperscalers are picking up their sites.
My bet for the next month: at least one major lab announces committed-spend pricing that looks a lot like cloud reserved instances. The labs have to make revenue predictable before the IPOs, and buyers can now demand a discount in return. If you have meaningful volume, ask for one now.