Pareto Hit Two Trillion Tokens in 48 Hours. Its Founder Says Price Was the Blocker

Pareto, a multi-model inference system from Circuit & Chisel, processed more than two trillion tokens in its first 48 hours — and its founder says the surge proved that price, not capability, is what holds back most AI adoption. Demand on launch day exceeded 5 to 10 billion tokens per minute, while the system could serve only about 1.6 billion. Serving all of that demand for free, Louis Amira said, would have cost many millions of dollars per day.

Amira, who co-founded Circuit & Chisel with David Noël-Romas, laid out the numbers in an interview with Unite.AI. His résumé includes co-founding a venture inside Google's Area 120 incubator, working at Chainlink, and leading crypto and AI partnerships at Stripe. The thread connecting those roles, he argues, is infrastructure: new technology that works in demos and stalls in production because the real-world plumbing does not exist.

Circuit & Chisel currently ships three products. ATXP provides identity and payment capabilities for AI agents. Pareto combines frontier and open-weight models (models whose weights anyone can download and run) to cut the cost and raise the quality of inference. Unbiased gives customers platform access to Pareto.

What did the founders build first, and why did the order turn out wrong?

The company started with payments. "An agent couldn't prove who it was, and it couldn't pay for anything," Amira said. "So we built ATXP to give agents an identity and a way to pay."

Customers then forced a reversal. Agents could hold an identity and spend money, but businesses refused to hand budgets to systems they did not trust. "Nobody hands a budget to an agent that picks the wrong model for the job or gets the answer wrong," Amira said. Work on quality and cost intensified earlier this year, and that became Pareto.

He is blunt about the sequencing: "You need both, and you need them in that order, because no one gives an agent a wallet before they trust its judgment. We just built them in the wrong order."

How does a multi-model system beat a single model?

Pareto is not a model. It sends each task to models from several labs — both frontier and open-weight — and assigns each one a role, sometimes generating and sometimes reviewing. Routing decisions weigh each model's capability, cost, speed, and reliability for the task at hand, plus how much verification the task deserves.

The design principle is disagreement, not consensus. "We don't average answers. Averaging blurs the errors together," Amira said. "We want the models to disagree, because disagreement is where mistakes show up." When the pool of available models changes, the team re-tests and ships a new numbered version of Pareto.

What are the actual cost numbers?

In Circuit & Chisel's published tests, the mix scored within a few points of GPT-6 Astra and Claude Opus 5.5 on most major benchmarks at 30-70% lower cost per task. The mechanism is mundane: most requests are routine, and a cheap model answers them as correctly as an expensive one. Pareto reserves frontier models, or several models working together, for the requests that need them.

Amira claims no significant trade-offs, with one caveat. "Pareto won't route every request perfectly, but we measure how often it gets the routing wrong and improve on (and publish!) the results," he said.

On evaluation credibility, the company publishes its evaluation code so anyone can rerun the numbers, runs every model through the same harness including its own, and displays losses at the same size as wins. "A scoreboard where the house always wins is just an ad," Amira said. Pareto also launched in multiple AI gateways, and preview access to new models went to companies of all sizes, with feedback incorporated in near real time.

What did two trillion tokens reveal?

The launch numbers carry the market signal. Demand ran several times above what Circuit & Chisel and its partners expected, and the company ran out of compute. On day one the model slowed enough that some users abandoned it. The team added capacity overnight and moved to paid access two days after launch, which Amira called "the honest way to ration."

The lesson he draws is economic: "Price is holding back a lot of AI use. There are builders whose product works at one price and doesn't work at five times that price. Bring the price down and they show up."

The launch also surfaced a disclosure problem. Some users felt misled when they learned that Union Alpha, a model launched with partners as "just another model," was in fact several models working together. Amira defends the framing — "there isn't a practical difference in how users should experience the output of a router" — but the episode shows how sensitive the market remains to how model systems are labeled.

How does zero data retention work across multiple providers?

For enterprises worried about sensitive prompts passing through third-party providers, Pareto offers zero-data-retention plans: no stored prompts or outputs, either at Circuit & Chisel or at any provider in the chain. Because a ZDR promise is "only as good as the weakest link," ZDR requests route exclusively to models running under zero-retention terms; providers that cannot meet those terms get no ZDR traffic.

Enterprise customers receive a data processing agreement and a ZDR addendum in writing. Plans that are not ZDR are labeled as such. A version customers can run on their own hardware is in progress.

Can multi-model systems fix agentic reliability?

Agents fail for compound reasons, Amira argues: "One small mistake in step two becomes a big one by step ten." Multiple models attack the hard steps directly — they can work a problem separately, check each other, or confirm a result before the agent proceeds. "A single model has no second opinion. Pareto does," he said.

The team measures user satisfaction and performance on real tasks, publishes the results including failures, and fixes what it finds. Customers who need deterministic behavior can pin a specific version of Pareto so it does not change underneath them.

Do agents get wallets next?

Amira's answer is yes, and soon. Today a person picks the model, calls the tools, and pays; increasingly the agent will do all three — selecting the right model for each step, finding tools, paying each provider, and checking results before moving on. That convergence requires exactly the two pieces Circuit & Chisel built: Pareto for task-appropriate intelligence, ATXP for identity and payment.

The gating factor is trust. Many companies are working on limits and liability before agents touch money, which is why Circuit & Chisel "started with judgment." In Amira's framing, the payment piece becomes more useful as trust in the agent grows.

Can a router company compete with the labs?

On the training-versus-orchestration question, Amira splits the difference: the labs will keep producing better models, and Pareto gets better every time one ships. "We'll never outspend the labs on training, we build on their work and help distribute them," he said. His pitch is relief from "model fatigue" — delivering the right model, for the right task, at the right ROI.

The competitive argument is structural. The labs could build routers themselves, and some will, but Amira notes: "A lab's router has a home team. We don't own a model, so the only way we win is if the customer gets the best answer for the money." Every new model that enters the market strengthens that position, which means the handful of frontier labs racing each other on benchmarks are, whether they intend it or not, feeding the router layer too.