Circuit & Chisel’s Pareto targets cheaper, reliable AI agents

By Olivia Hart

Circuit & Chisel’s Pareto processed more than 2 trillion tokens in 48 hours, founder Louis Amira said. Launch-day demand reached 5 billion to 10 billion tokens per minute, versus roughly 1.6 billion served.

What did the launch demand reveal?

Demand ran several times above Circuit & Chisel and its partners’ expectations, Amira said. Serving all demand free would have cost many millions of dollars per day.

The service slowed, and some users left, according to Amira. Circuit & Chisel added capacity overnight and moved to paid access two days after launch.

Circuit & Chisel launched Union Alpha with partners as another model. Some users felt misled when they learned several models produced the output; the company said users should not care whether a model or router supplied it. The interview did not explain Union Alpha’s relationship to Pareto.

Takeaway: The launch showed strong demand for cheaper inference, but exposed capacity and disclosure weaknesses.

Why did Circuit & Chisel start with payments?

Amira co-founded a venture inside Google’s Area 120, worked at Chainlink, and led crypto and AI partnerships at Stripe. He said early choices determine later options, and he saw technology stall outside demos because real-world plumbing was missing.

According to Amira, he and David Noël-Romas founded Circuit & Chisel after agents repeatedly failed to prove their identities or make payments. ATXP was their first protocol, designed for autonomous payments between AI agents.

ATXP now provides identity and payment capability, while Unbiased provides platform access to Pareto. Pareto combines frontier and open-weight models to address inference quality and cost.

Amira said ATXP worked: agents could hold an identity and spend money, yet trusted output mattered more to customers. The company accelerated Pareto work earlier in the year; Pareto now decides how an agent thinks, while ATXP lets it act and pay. He called that sequence backward.

Takeaway: ATXP gives agents identity and a way to pay; Pareto must first earn their operators’ trust.

How does Pareto combine models?

Amira described Pareto as a system, not one model. It assigns work across models from several laboratories, including frontier and open-weight models, then combines their results; one model can also review another’s work.

Routing considers capability, cost, speed and reliability for each task, plus how much checking it warrants, Amira said. Pareto avoids averaging answers because averaging can obscure errors; competing answers expose mistakes.

When the model mix changes, Circuit & Chisel re-tests Pareto and ships a numbered version, according to Amira. Customers can pin a version when they need repeatable behavior.

Takeaway: Pareto treats disagreement and versioning as safeguards, not model averaging.

Can model routing actually cut costs?

Routine requests often go to cheaper models because they answer those tasks as often as the costliest model, Circuit & Chisel says. Pareto reserves frontier models, or groups of models, for harder requests.

On company-published tests, Pareto scored within a few points of GPT-6 Astra and Claude Opus 5.5 on most major benchmarks, at 30% to 70% lower cost per task. That is roughly $30 to $70 less per $100 of baseline cost, under the company’s test assumptions.

Amira said routing still fails sometimes despite his claim that trade-offs are insignificant. The interview disclosed no model prices, latency measurements or independent audit. Circuit & Chisel publishes evaluation code, routes every model through one harness and publishes benchmark losses and routing errors.

Amira said: “A scoreboard where the house always wins is just an ad.” The company launched Pareto through multiple AI gateways and gave previews to companies of different sizes, using near-real-time feedback to improve performance.

Takeaway: The claimed savings are meaningful on paper, but remain tied to company testing and routing accuracy.

How does zero-data retention work?

Amira said Circuit & Chisel’s zero-data-retention plans store neither prompts nor outputs, and require the same from every provider receiving those requests. Providers that cannot meet zero-retention terms receive no ZDR traffic.

Enterprise customers receive a data processing agreement and a ZDR addendum in writing, Amira said. The company also labels plans that do not qualify as ZDR, reducing the chance customers mistake them for protected offerings.

For local control, Circuit & Chisel is developing an on-premises version for customer hardware. Amira gave no release date, and the product remains unfinished.

Takeaway: Zero-data retention is contractual and provider-dependent; the strongest local-control option remains in development.

Can multiple models make agents more reliable?

Amira said agents often fail outside demonstrations because a mistake in step two becomes a large error by step ten. Pareto can have models solve the same problem, check one another or confirm a result before an agent proceeds.

The system aims to reduce hallucinations and conflicting outputs, but its benefit depends on model and check selection. Circuit & Chisel says it measures satisfaction and performance on real tasks, publishes failures and fixes issues found.

The interview offered no independent reliability study. Pinning a numbered Pareto version limits behavioral changes, according to Amira, but consistency still depends on underlying providers.

Takeaway: Multi-model review can catch errors, but it does not remove them or guarantee consistent agent behavior.

Can agents select models, tools and payments?

Today, a person chooses the model, tools and payment method. Amira expects agents eventually to make those choices, select the model for each step, find required tools, pay each provider and check results before continuing. He said Pareto supplies intelligence, while ATXP supplies identity and payment.

He said limits and liability rules remain prerequisites for trusting agents with money, which is why Circuit & Chisel started with judgment. Payments should become more useful as confidence in an agent’s decisions increases, according to Amira.

Amira expects labs to improve their models and said Pareto will benefit without outspending them on training. Circuit & Chisel will build on and distribute that work. Some labs may build routers with a home-team bias, he said, but Circuit & Chisel owns no model and depends on delivering the best answer for the customer’s money.

Amira said Pareto should deliver the right model for the right task at the right return, sparing users from testing every release. He called that burden model fatigue and said more models would increase the value of a neutral selection layer.

Takeaway: Circuit & Chisel’s bet is that model proliferation, not a single breakthrough, will make orchestration and autonomous payments valuable.