Reason Coding Harness Promises 20x Lower Costs. Its Own Benchmarks Show 22%

Every team running coding agents knows the feeling: the agent fixed the bug, and the invoice fixed your mood. Reason Machines, a Y Combinator startup, is selling the cure. On November 13, 2026, cofounder Adi Singh introduced Reason on X as “the world’s most cost-efficient harness,” promising “frontier coding performance at up to 20x lower cost.” The post took off. The company’s own published benchmarks, however, tell a more modest and more interesting story: a 22% lower cost per solved task than OpenAI’s Codex on the same model.

What happened

Reason Machines was founded by Adi and Sven, who studied EECS at UC Berkeley and, in their words, “dropped out to build what a self-driving company runs on,” according to the company’s launch blog post dated November 9. Singh’s X bio lists the company as part of YC F26. The blog post claims “20,000 users, 14,000 X followers, and 3M+ impressions” in the four weeks since launch, and says the company is raising its seed round.

The November 13 video post pushed the product into the trending list. In it, Reason patches a security bug, merges the fix and repairs a customer’s broken export straight from Slack. TestingCatalog summarized that Reason “solved more tasks than Codex on the same model while spending 22% less” and that users can give it standing jobs, such as reviewing every new pull request.

The numbers, side by side

Here is where the claims diverge. The launch post on X says “up to 20x.” The November 9 blog post says “up to 30% lower cost per task than Codex, Cursor or Devin.” The detailed Reason 1.1 benchmark write-up from October 10 reports:

The 1.1 write-up also says Reason “costs 15–23% less per attempt than Codex on every benchmark” and uses 23% fewer input tokens per attempt (1.92 million against 2.48 million) and 15% fewer output tokens on GPT-6 Astra. Crucially, all results are on Artificial Analysis’s Coding Agent Index v1.5 (DeepSWE v1.1, Terminal-Bench 4.0 and SWE-Atlas QnA). Artificial Analysis moved to version 2.0 on October 22, so the figures are not directly comparable with its current leaderboard. In the materials i10x reviewed, Reason did not publish a same-model comparison showing a 20x saving. A gap that large likely depends on comparing a cheap model inside Reason with an expensive frontier setup, but the company has not said so explicitly.

How it works

The core idea fits in one sentence from the company: “The model sets the pass rate; the harness sets the bill.” Reason says the same model can cost up to five times more in one harness than in another. Two design choices drive its savings:

• Cache-friendly prompts: “The prompt and tools never change between turns and each request only appends to the last, so every turn is fully cache-eligible up to its new content.” As AI researcher Rohan Paul explained on X, providers bill repeated prompt prefixes at a steep discount, which compounds over long agent sessions.
• Codemode: A tool that runs JavaScript calling the other tools, so “a search, five file reads and a test run can take one model call.” Fewer round trips mean fewer tokens.

Everything else (subagents, cloud computers, preview URLs, pull requests with screenshots) sits behind “deferred tools” the model finds on demand, keeping the base prompt small. Reason also publishes its benchmark runner on GitHub, so anyone can reproduce a run with the Harbor evaluation framework:

The repository’s quick check, as a script we call run-deepswe.sh:

uv tool install "harbor[e2b,modal]==0.22.0" --with "reason-machines-harbor>=0.3.44"
export REASON_API_KEY=<YOUR-REASON-KEY>
 
# One DeepSWE task on local Docker, to check your setup.
harbor run --dataset datacurve/deep-swe-1-1 --n-tasks 1 \
   --agent reasonmachines_harbor:ReasonAgent \
   --model openai/gpt-6-luna \
   --ak reasoning_effort=xhigh \
   --allow-agent-host api.reasonmachines.com \
   --allow-agent-host relay.reasonmachines.com

Sessions run from a web app, CLI, macOS or Linux desktop app, Slack or an HTTP API, on cloud Linux, a cloud Mac, the user’s own Mac or an SSH box. Users can bring API keys, open-weight models and subscriptions from Grok, Copilot or Kimi with automatic failover, according to the launch post.

Pros and cons

For users and solo developers

• Pro: If the caching savings hold, a flat $50 per month with bring-your-own subscriptions can stretch existing AI plans further.
• Con: Another subscription on top of the ones you already pay for, and savings depend heavily on which model you plug in.

For developers and engineering teams

• Pro: Reproducible benchmark code, Slack-native standing jobs (like reviewing every pull request) and support for many models instead of a single vendor.
• Con: The headline numbers are self-reported, and the benchmark version they use has already been replaced by Artificial Analysis.

For companies

• Pro: Agent spend is becoming a real budget line. Even a 20% to 30% cut on large agent workloads adds up.
• Con: A seed-stage vendor in the middle of code, credentials and cloud machines is a security and continuity risk that procurement teams will want to examine carefully.

What X is saying

The launch post itself:

Frontier coding performance at up to 20x lower cost.

Introducing the world’s most cost-efficient harness: Reason

Stop burning thousands on runaway agent bills.

Reason outperforms top harnesses on Terminal-Bench and DeepSWE while dropping cost per task to a fraction of the frontier baseline.

Comment “Reason” and we’ll send you early access + free compute credits.

— Adi Singh (@xadisingh) November 13, 2026 · View on X

TestingCatalog focused on the measured numbers:

TestingCatalog (@testingcatalog)

— 🚨 AI News Reason Machines has launched Reason, its AI coding agent designed to maximize work completed per dollar.\n\nDuring testing, Reason solved more tasks than Codex on the same model while spending 22% less.\n\nUsers can give it standing jobs to run autonomously, such as reviewing every new pull request, and run it using their own ChatGPT, Claude, GitHub Copilot, or Grok membership.|November 13, 2026 · View on X

Rohan Paul explained the caching trick:

Reason Machines just launched Reason, a coding harness that wraps a single model like GPT-6 Astra and solved tasks 22% cheaper than OpenAI’s Codex.

How: A traditional coding agent resends its context to the model on every step, so Reason trims that load:

Reason holds the prompt prefix identical between turns, by putting its fixed instructions and small tool list at the top of every prompt and only adding new code and results at the end, never rewriting or reshuffling the earlier text.

Since every turn starts with the exact same text, OpenAI recognizes it from the previous turn and bills those repeated tokens 90% cheaper, at $1 instead of $10 per million.

— Rohan Paul (@rohanpaul_ai) November 13, 2026 · View on X

The launch post asked readers to comment “Reason” for early access and free credits, a format that inflates reply counts and says little about real-world use. i10x did not find a substantive Grok assessment of Reason on X; Grok’s replies only restated the launch claims.

Critical analysis

Reason is tackling a real problem. Agent sessions are long, and providers discount cached tokens heavily, so a harness that keeps its prompt stable can save meaningful money. The design choices are sound and, unusually for a startup, documented with reproducible code.

The marketing is another matter. “Up to 20x” and “22% on the same model” are not the same claim, and only the second is backed by a published, same-model comparison. Readers should also note that the comparison is Reason’s own reproduction of a third-party benchmark, not an official Artificial Analysis result, and that the frontier labs keep tuning their own harnesses (OpenAI’s Codex, Anthropic’s Claude Code) for caching too. A 22% edge today can shrink with one vendor update.

What’s still open

• Which comparison produces the “up to 20x” figure. Reason has not published it in the materials i10x reviewed.
• Independent results on Artificial Analysis’s current Coding Agent Index v2.0.
• How Reason secures customer code and connected subscription credentials across cloud machines.
• Funding: the founders say they are raising a seed round; no amount or investors have been announced.