Mubit Trains Small Models on Your Agents’ OpenAI Calls

Mubit, an AI start-up that Anil Peri says began in Aberdeen, has launched a service that records the calls a company’s AI agents send to model providers, groups the ones that repeat into tasks, and trains a small model for each task, according to its website and documentation on 23 October 2026. It charges $5 per 1,000 runs, its pricing page says, and the calls themselves stay on the customer’s own Anthropic, OpenAI or Gemini key.

Alongside the launch it has published Banyan Sapling, an open-weight model for checking claims against a policy, as a research preview on Hugging Face.

“We started Mubit in Aberdeen. No network, no lab, nobody knew us,” Mubit’s Anil Peri wrote on X on 23 October. “And the thing we kept seeing: you and your competitor run the same AI. It should be yours.” Mubit’s homepage puts the pitch this way: “Your frontier models stay. You own what’s built.”

Business Pill · DISTILLATION EXPLAINED

A short Business Engineer explainer of distillation, how a small model learns from a large one. It teaches the general idea only and says nothing about Mubit.

What Mubit Says It Does

The documentation describes the flow in four steps. “Mubit passes your app’s calls on to your model provider and records them. It groups similar calls into tasks and trains adapters, small models each made for one task,” it says. “Each adapter version is tested against your provider model’s answers.”

The homepage describes the same loop as runtime calibration: “It calibrates every call at runtime, then builds and refreshes a dedicated model for each task your agents repeat, switching over only when it does better.” The site also lists a launch week from 26 to 30 October 2026 and a waitlist.

Source: mubit.ai homepage, mubit.ai/pricing and docs.mubit.ai, read on 23 October 2026. These are the company’s own descriptions; we did not test the service.

What Was Posted on X

Mubit’s Anil Peri wrote on X that the company started in Aberdeen with no network and no lab, and that ‘you and your competitor run the same AI. It should be yours.’ The framing is his own.

The Open Model: Banyan Sapling

Banyan Sapling is “the first model in the Banyan family” and focuses on verification, the homepage says. Given a policy, evidence and a claim, it returns one of four labels: supported, violation, insufficient information or inconsistent context, with calibrated uncertainty, according to the model card. The card lists Qwen3-30B-A3B-Instruct-2507 as the base model, 23,042,164,740 stored parameters and an Apache 2.0 licence.

The card says the final run trained one epoch over 50,000 examples: 42,809 controlled synthetic cases and 7,191 ContractNLI training claims. It reports zero false ‘supported’ decisions in the calibrated output and 14 in raw predictions. It is also direct about the limits: “the current shared calibration leaves most contract cases unresolved.”

Source: MubitLabs, Banyan Sapling model card on Hugging Face, read on 23 October 2026. An answered case is one where the calibrated result is not ‘unsure’. The card says real-world accuracy and nominal conformal coverage “are not established by these results.”

What Was Posted on X

Wall St Engine summarised the launch on X: companies “can connect the OpenAI or Anthropic calls their agents already make, or import activity logs from tools such as Braintrust and Langfuse.” AI commentator Rohan Paul wrote that two companies on Mubit “end up with different small models, because each is shaped by its own agents’ work.” Both posts are the posters’ own summaries.

Connecting the Coverage

In our 27B agent-interface piece, we reported a Simular paper in which a 27-billion-parameter open model fully solved 21.3% of OSWorld V2 desktop tasks at an estimated $1.10 per task against $72.40 for the reported Claude Opus 4.8 result. In our harness scaling piece, a paper found that training an agent across 18 harness configurations beat training through one. In our Steinberger token-spend piece, OpenAI’s Peter Steinberger estimated spending around $10 million in tokens last month.

And in our Kimi K3.1 Pro piece, Moonshot listed $2.00 per million input tokens and $10.00 per million output tokens. What’s new: Mubit sells the step between those stories, turning a company’s repeated calls to a rented model into small models it can download.

The Structural Read

As we read it, the unit of value is the repeated task: the docs define a task as calls with matching model, instructions, tools, tool choice, and output format, and say a task first trains by default at 25 kept examples.

As we read it, the frontier model stays in the loop as the yardstick: the docs say each adapter version is tested against the provider model’s answers, and the pricing page says models serve only after matching or beating the frontier model in evals.

As we read it, the open model is a component, not a chatbot: the Banyan Sapling card describes it as a verification component that classifies supplied text, and says retrieval and final business decisions belong to the surrounding workflow.

Three Implications

FOR TEAMS RUNNING AGENTS The docs say the connection is a base URL and a Mubit key, and that model IDs stay the same.

FOR FRONTIER MODEL PROVIDERS The pricing page says Mubit never marks up a token and that calls stay on the customer’s own provider key.

FOR BUYERS OF VERIFICATION The Banyan Sapling card reports zero calibrated false ‘supported’ decisions against 14 in raw predictions, and says real-world accuracy is not established by its results.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The framework says “what changes between paradigms is where intelligence accumulates and who can capture it.”

As we read it, Mubit is a bet that some of that accumulation can sit with the company running the agents, in small models trained on its own repeated calls. The framework was published in July 2026 and does not discuss Mubit.

What Is Not Established

Mubit’s site, pricing page, documentation and launch posts that we read do not state funding, revenue or named customers. Hoxton Ventures lists Mubit in its portfolio, with Anil Peri as founder, founded in 2025, headquarters in Aberdeen, fund Hoxton IV and invested in 2026; the page does not give an amount. Of customers, Peri wrote that the company works “with some of the best companies and engineers in the world”. The service’s quality claims are the company’s own, and we did not test it. The Banyan Sapling figures come from Mubit’s own evaluation, which its card calls “a regression and development characterization rather than an independent estimate”.

We read the homepage, pricing page, documentation and Hugging Face card on 23 October 2026 and did not contact Mubit. Nothing here is investment advice.

The Bottom Line

Mubit says it will record the calls a company’s agents already make, train a small model for each task they repeat, switch to it only when it matches or beats the frontier model in evals, and let the customer download it, for $5 per 1,000 runs. Its first open model, Banyan Sapling, answered 45.00% of its own combined test cases, all matching their reference labels, and left the rest unresolved.