Cognition’s Plan to Limit Tokenmaxxing to the Most Efficient Engineers

To rein in AI spending, large companies are capping employees’ usage, turning to cheaper and smaller models and using routers to send only the hardest requests to the most expensive models.

Cognition, the startup behind the AI coding agent Devin, is adding a new trick to the book: measuring how efficiently different engineers use AI so companies can allocate more resources toward their most efficient employees, Cognition told The Information exclusively.

“The limits that a user should have should not be static. They should be dynamic based on the efficiency of your usage and the value that you're getting out of it,” said Calvin Giroud, a vice president of customer engineering at Cognition.

By looking at an engineer’s coding session using Devin, Cognition can estimate the equivalent number of hours it would have taken a human to do the same amount of work. Dividing by the costs of that session, Cognition gets a measure of each engineer’s efficiency.

Then, when engineers run into usage limits, their limits can automatically increase if they have a “healthy” or “satisfactory” grade.

“This is how credit cards work, for example,” said Giroud. “You have a set credit limit, and as you prove that you are a good steward of the credit card and you're paying back things on time, you get higher and higher and higher limits.”

Helping customers rein in token use could seem paradoxical for Cognition given it charges based on usage. But the company argues that in this hyper-competitive market for AI coding tools, it is incentivized to build trust with customers that will deliver long term loyalty rather than maximizing spending now.

Unrestrained tokenmaxxing is “not a sustainable long-term approach, and many of our customers are large public companies that have to answer to their shareholders and public markets around the efficiency of the tools that they're using,” he said.

Cognition is also rolling out Devin Coach, an AI system that gives feedback on engineers’ requests to encourage them to work more efficiently. The coach is prompted to spot the top 20 to 30 most common mistakes, such as an ambiguous request or a user selecting an unnecessarily expensive model for their task.

For example, if a user tries to combine multiple distinct tasks in the same request, the coach might notice that those tasks could more quickly be completed in parallel and suggest splitting them into separate sessions.

The coach will also notice if employees are using Devin for personal tasks, such as planning a vacation. That was motivated by worries from a customer of Cognition’s that found some employees using Devin for non-work tasks, though it was a “relatively trivial percentage” of usage, according to Giroud.

Here’s what else is going on…

Big Number

Anthropic’s revenue in the third quarter hit $16.8 billion, the company on Wednesday told prospective investors in its initial public offering, The Information reported. That marked a jump of about 45% from the second quarter, and was more than 3.6 times the company’s revenue for all of 2025. But it also represents a slower growth rate than Anthropic’s second quarter, when fevered demand for its Claude Code product drove revenue up by more than 140% from the first three months of the year.

Anduril Industries is in talks to allow employees to that would value the autonomous munitions maker at about $90 billion, according to Bloomberg. Employees would be able to sell shares at $94.80 each in the tender, which would be capped at roughly $1.3 billion and is set to close in early November, the report said.

Overheard

OpenAI has accused a former employee of sharing the company’s unreleased device plans with senior leadership at Apple—a charge that echoes those Apple made against OpenAI in a lawsuit the iPhone maker filed in July.

Morgan Stanley CEO Ted Pick said on Wednesday that he expects the AI buildout will encounter bottlenecks and periods of overbuilding, but investment will continue at a massive scale over the next several years.

Chinese AI agents took exposed credentials from a company’s source code and bypassed safeguards in an attempt to find information on mapping service Amap, according to a report on Wednesday from a group of researchers known as “swarmchasers.”

People on the Move

Sachin Katti, vice president of compute strategy at OpenAI, said Wednesday he was leaving the role on November 15 after about a year at OpenAI. Katti, who joined OpenAI after four years at Intel, announced that he would be “moving on to build something new” in a post on X.

Deals and Debuts

See The Information’s Generative AI Database for an exclusive list of private companies and their investors.

Modal, which runs AI workloads on cloud infrastructure, raised $750 million in a Series D funding round led by Accel at a $15.75 billion valuation.

Berkshire Hathaway has agreed to invest $330 million in Italian digital conglomerate Bending Spoons in a private placement of stock, providing a boost to investor confidence in Bending Spoons. Shares of the Italian company leapt 40% to $52.88 on Wednesday.

Nvidia announced a five-year, $250 million compute commitment to fund agentic-safety and security research in the UK, alongside new planned investments in five UK AI and quantum startups: Boltz, Foresight, Hiverge, Prima Mente and Riverlane.

Salience Labs, which makes optical networking chips for AI datacenters, raised $125 million in a Series B funding round led by Bullhound Capital.

Labelbox, which labels data to train AI models, raised more than $100 million in a Series E funding round led by Infinitum Partners.

Booking Holdings launched Vayo, a Seattle-based AI travel-agent startup led by Mike Fridgen, with a planned $100 million investment contingent on the product's progress.

Mueon, a Portland-based startup that’s developing a new way to arrange AI server rack components, raised $65 million in funding led by Maverick Silicon, The Information reported.

Mollie committed €50 million (about $54 million) to a ring-fenced internal venture to build a dedicated banking and financial layer for autonomous AI agents, letting its payments customers run their own AI agents directly against Mollie's banking rails.

Harmonic Security, which monitors how employees use AI tools, raised $50 million in a Series B funding round led by Crosspoint Capital Partners.

Zenzap, which makes an AI-powered chat app for workplaces, raised $33 million in a Series A funding round led by Bessemer Venture Partners and Lightspeed Venture Partners.

Tamr, which uses AI to clean up enterprise data, raised $21 million in a funding round led by Volition Capital, with participation from existing investors.

Sonilo, which generates AI soundtracks for video, raised $11 million in a funding round led by B Capital, with participation from Redpoint.

Evolutionary Operating System, which builds AI agents for business operations, raised $9 million in a Pre-Seed funding round led by Uncork Capital and Field Ventures.

Othello, which coaches sales reps using AI, raised $8 million in a seed funding round from backers including a16z speedrun.

Parallel Loop, which helps enterprises build AI workflows, raised $6.6 million in a seed funding round led by 3one4 Capital.

Marklo, which builds AI marketing tools for Shopify brands, raised $5 million in a seed funding round led by Forerunner.

Shamrock AI, which helps enterprises test AI agent deployments, raised $5 million in a seed funding round led by Endiya Partners.

CodeWall, which uses AI agents for penetration testing, raised $4 million in a Pre-Seed funding round led by Osney Capital.

Data center operator AirTrunk is postponing a planned initial public offering of a real estate investment trust in Singapore to next year as higher interest rates and volatile markets dampen investor demand, Bloomberg reported.

Apple is preparing to release its Apple Intelligence on-device AI service in mainland China in December, ending a lengthy wait caused by regulatory hurdles, Bloomberg reported Thursday.

Nvidia and ChipAgents, a startup that helps customers automate chip design, on Wednesday released a benchmark that measures how different AI models perform at the task.

METR, the AI evaluation organization, released a proposal outlining how embedded evaluators could publish information about whether frontier AI companies could come close to losing control of their models.

Taiwan Semiconductor Manufacturing Co. said Thursday it will build another 3-nanometer chip factory in the Taiwanese city of Tainan, with production set for the first half of 2028. The move signals confidence that strong AI chip demand will last for years.

Thank you for reading the AI Agenda Newsletter! I’d love your feedback, ideas and tips: [email protected].

If you think someone else might enjoy this newsletter, please pass it forward or they can sign up here.