Corbanu Terminal 0.1.48 · open source · Rust

Same model. Better agent.

Give Corbanu and Claude Code the same Opus 5.5 model: Corbanu passes more hidden tests and costs 20% less. On Kimi K3 it costs 41% less than Hermes and finishes 41% faster.

Measured on 10 coding tasks, 3 runs each, graded by hidden tests in sealed containers. Every agent got the same model, the same provider and the same reasoning level. Costs are what the provider actually billed.

97%
hidden tests passed on Opus 5.5. Claude Code: 95%.
20% cheaper
than Claude Code on Opus 5.5. $0.33 vs $0.41 per task.
18% faster
than Claude Code on Opus 5.5. 1.5 vs 1.9 min per task.
41% cheaper
than Hermes on Kimi K3. $0.22 vs $0.36 per task.
41% faster
than Hermes on Kimi K3. 3.9 vs 6.7 min per task.

01 / BENCHMARKS

Head to head on the same model.

Same 10 tasks and 3 runs per agent on each model. Tasks include building libraries from a written spec and fixing bugs in existing code. The agents never see the tests they are graded on.

Opus 5.5 · 30 runs per agent
CorbanuClaude CodeHermes
Tasks fully passed70%67%70%
Hidden tests passed97.0%95.0%94.0%
Cost per task$0.33$0.41$0.43
Time per task1.5 min1.9 min2.1 min
Model calls per task6.47.39.0
Kimi K3 · 30 runs per agent
CorbanuHermes
Tasks fully passed70%63%
Hidden tests passed92.0%91.5%
Cost per task$0.22$0.36
Time per task3.9 min6.7 min
Model calls per task9.912.8

How we tested

  • Run on 24 September 2026 with Corbanu 0.1.48, Claude Code 2.1.281 and Hermes v2026.9.24.
  • Each run starts in a fresh, sealed container. The grading tests are never in the prompt or the workspace.
  • Every agent used OpenRouter, pinned to the same host (Google for Opus 5.5, Moonshot for Kimi K3), with reasoning set to high.
  • Cost is the amount OpenRouter billed. Time is wall-clock time per task, averaged over all runs.

02 / WHY

Less wasted work.

Fewer model calls

Corbanu gets the job done in fewer round trips: 6.4 calls per task on Opus 5.5, against 7.3 for Claude Code and 9.0 for Hermes. Fewer calls means less time and a smaller bill.

Any model

Run Claude, GPT, Kimi, GLM or DeepSeek through OpenRouter, Vercel or your own provider keys. Pick a different model for each agent.

Open source

Apache-2.0 and written in Rust, built on the Codex CLI runtime. Read the code.

03 / INSTALL

Install it in one line.

Install · Corbanu Terminal 0.1.48

Linux

curl -fsSL https://github.com/CorbanuCore/CorbanuTerminal/releases/latest/download/install.sh | sh

macOS

curl -fsSL https://github.com/CorbanuCore/CorbanuTerminal/releases/latest/download/install.sh | sh

Also available as CorbanuTerminal-aarch64-apple-darwin.dmg and CorbanuTerminal-x86_64-apple-darwin.dmg.

Windows

Invoke-RestMethod https://github.com/CorbanuCore/CorbanuTerminal/releases/latest/download/install.ps1 | Invoke-Expression

Use the native PowerShell installer. PowerShell's curl alias is unsupported.

  • Release 0.1.48: notes, checksums and downloads
  • Run corbanu to start. Also included: an encrypted key vault, agent teams, prepaid plans paid in USDC, GPU rental and Post Fiat Task Node.