Reason Machines' Reason Harness: 20x Cheaper, or 22%?

§ 01What the launch post says#

• The claim. “Frontier coding performance at up to 20x lower cost.”
• The product. “Introducing the world’s most cost-efficient harness: Reason”
• The benchmarks. “Reason outperforms top harnesses on Terminal-Bench and DeepSWE while dropping cost per task to a fraction of the frontier baseline.”
• Who posted it. Adi Singh, whose X profile says he is a co-founder of Reason Machines. The post went up at 15:34 UTC. The company’s account quoted it 36 minutes later.
• Its size. It had about 1,810,000 views when we read it at 02:22 UTC on November 14.
• What it leaves out. The post’s text does not say what the 20x is measured against.

§ 02What the chart in the video shows#

The post carries a 90-second video. We read two of its charts from the video’s frames. The second chart has a banner that reads 22× lower cost, with this line under it: “at a higher pass rate (72.3% vs 68.4%)”

• What the 22× compares. The banner’s line runs from the first bar to the last: Claude Code on Claude Opus 5.5 at $13.04, and Reason on GPT-6.1 Sol at $0.60. By our arithmetic that is 21.7 times.
• The two bars use different models. One is Claude Opus 5.5 at max effort. The other is GPT-6.1 Sol at high effort.
• Same model, same chart. On GPT-6 Astra the chart gives Reason $3.96 and Codex $4.87 per task. By our arithmetic that is 19% lower per task, and 22% lower per solved task once the pass rates of 72.3% and 69.0% are counted.
• The model change inside Reason. Moving from GPT-6 Astra to GPT-6.1 Sol takes the chart’s cost for Reason from $3.96 to $0.60 at the same 72.3% pass rate. By our arithmetic that is a factor of 6.6.
• Where the costs come from. The chart’s footnote names three sources. The first: “Reason, Pi and Codex measured on Harbor, 113 tasks × 3 attempts.”
• The other two. The footnote says Devin Fusion’s figures are reported by Cognition, and that the two Claude Code costs are Artificial Analysis’s published cost per task, averaged over its three index benchmarks.
• One thing to notice. The chart’s subtitle says DeepSWE v1.1. By its own footnote, the two Claude Code costs are an average over three benchmarks.

§ 03What Reason Machines’ own posts say#

• October 10. The post that introduces Reason 1.1 reports 65.1 against 62.5 for Codex on the Coding Agent Index v1.5, both on GPT-6 Astra. It gives Reason the lowest cost per solved task: 22% below Codex and 10% below Pi.
• Its own caveat. “They are not official Artificial Analysis results, and Artificial Analysis has not reviewed or endorsed them.”
• The index has moved. The same post says Artificial Analysis moved to version 2.0 of the index on October 22, 2026, so these scores are not directly comparable with its current leaderboard.
• What it is built on. The post says: “Reason 1.1 is built on Pi” Pi is one of the other harnesses on the chart above.
• Why it says it is cheaper. The post credits two choices: tool calls that run in parallel inside one program, and a small fixed set of tools. It adds: “we have not yet tested how much of the saving each accounts for”
• November 9. The company’s launch note says teams get “up to 30% lower cost per task than Codex, Cursor or Devin”
• The company’s X account, also on November 9: “On a harness 22% cheaper per solved task than Codex.”

§ 04What Reason is#

• In its own words. “Reason is a coding agent that runs your repositories, tools and computers.”
• Who makes it. Reason Machines, which says it is backed by Y Combinator in the Fall 2026 batch. Its November 9 note names two co-founders, Adi and Sven, and says: “We’re raising our seed round”
• Not new this week. The company’s blog introduced Reason 1.0 on September 14 and Reason 1.1 on October 10, 2026.
• What it runs on. Its guides say it takes ChatGPT, Claude, SuperGrok or X Premium+ and GitHub Copilot sign-ins, or your own API keys. Work on those is billed by the provider.
• Price. Pro at $50 and Max at $200 a month, per its guides, with a 3-day free trial for eligible new customers.
• Open source. The company’s pages say two things. Its file for AI agents says Reason has a free, open-source engine. One of its guides says of Reason: “with no free plan, and it isn’t open source” We did not look for the code.

§ 05What we did not check#

• We have not used Reason, and we did not run any benchmark.
• We did not open the public benchmark jobs the company links, or its benchmark repository.
• We did not check any cost or pass rate against Artificial Analysis or Cognition.
• We read the charts from frames of the video. We did not transcribe what is said in it.
• Both charts in the video give Reason on GPT-6 Astra a cost of $3.96 per task. The first chart plots the three-benchmark index, and the second is labelled DeepSWE v1.1. We do not know which of the two that cost belongs to.
• We do not know what the 20x in the post’s text is measured against. The nearest figure we found is the 22× on the chart.

§ 06Where CellCog sits#

CellCog is our product, an AI employee platform, so read this page as coming from one side. Reason Machines also markets Reason as AI employees, and its own comparison of nineteen such products lists CellCog. We have not used Reason, and the company has not seen this page. CellCog is not a coding harness and is on none of the charts above. CellCog AI employees run on Anthropic’s Claude models. OpenAI makes the image model behind this page’s pictures.

§ 07Sources#

• Adi Singh, co-founder of Reason Machines: post on X with a 90-second video, November 13, 2026, 15:34 UTC.
• Reason Machines: post on X quoting the launch post, November 13, 2026, 16:10 UTC.
• Reason Machines: Introducing Reason 1.1, October 10, 2026.
• Reason Machines: Reason: the most efficient agent harness to maximize your AI usage, November 9, 2026.
• Reason Machines: post on X, November 9, 2026.
• Reason Machines: its guides and its site text for AI agents. All of the company’s pages were read at 02:07 UTC on November 14, 2026.