Fewer model calls: the technical choices behind Reason

Persistent sessions and a choice of models: Reason relies on a minimal execution environment to reduce the cost of running agents.

A single session can run commands on a local Mac and a cloud computer, then be resumed from another interface. That is how Reason Machines describes its agent environment: the model reasons in the cloud, while commands execute on the selected machines. The product combines that continuity with a promise of savings based on how model calls are organized.

The software layer involved is the harness: it prepares context, exposes tools, and manages task execution. In its introduction to Reason 1.1, the team reports a Coding Agent Index score of 65.1, compared with 62.5 for Codex and 61.4 for Pi, using GPT‑6 Astra. It claims a 22% lower cost per solved task than Codex. That metric accounts for unsuccessful attempts by dividing the average cost per attempt by the aggregate success rate.

The comparison uses version 1.5 of the index, comprising DeepSWE v1.1, Terminal-Bench 4.0, and SWE-Atlas QnA, totaling 303 tasks. The protocol calls for three attempts per task. The index and its component benchmarks have since changed: scores published in that post are not directly comparable with the current Artificial Analysis index. The claim of cost reductions by a factor of up to twenty in the social media post is not the result detailed in this same-model comparison.

The commercial bill also requires distinguishing two amounts. The published methodology separates the model provider’s token prices from what Reason charges when the platform pays for those calls: the provider’s price then carries a 25% markup. With a personal API key, the provider bills model calls directly. Some older trials use the amount charged when detailed token records are unavailable. The claimed savings therefore cannot be applied uniformly across every access method.

The technical approach combines a small set of general-purpose tools with program execution. A JavaScript tool can group multiple searches and file reads, then return only the useful information to the model. More specialized functions are discovered on demand. The team also describes keeping instructions and tool definitions stable to encourage context caching. Version 1.1 is built on Pi; the developers have not experimentally isolated each design choice’s contribution to the savings.

The effect of harnesses on costs is also the subject of outside research. In HarnessTax, researchers compare twenty-one model–harness pairs, covering seven models and three tools: Claude Code, Codex CLI, and Pi. On SWE-bench Lite and Terminal-Bench 2.0, some cost differences reach a factor of five at similar success rates. The study does not test Reason; its findings apply to the tasks, configurations, and prices evaluated.

The product extends these technical choices to persistent work sessions. The launch post lists cloud Linux and Mac computers, a personal Mac, and SSH-accessible hosts. The session thread provides continuity across those access points; the documentation describes one agent working across multiple computers. The founders also advertise hundreds of parallel sessions and workflows that can run for several days.

Support for users’ own API keys and compatible subscriptions, including Grok, Copilot, and Kimi, is part of the announced offering, alongside provider failover mechanisms. The interface also presents Reason alongside Claude Code, Codex, and Cursor, allowing multiple agents to be tracked in one workspace.

In Slack, mentioning Reason in a thread can initiate a request for a pull request, a presentation, or an answer. The API supports starting a session from a script, a continuous integration pipeline, or another application, then monitoring progress and retrieving the result. Work started that way can continue in the app, command line, or another client connected to the same session.