I want to be upfront about something: I didn’t build this. Our engineering team did. But I get to write about it, and I’ve been waiting a while to write this one.
This week we submitted the Backboard CLI to the official Terminal-Bench 2.1 leaderboard. The score: 85.4% ± 0.8%, with a pass@5 of 0.888, running Claude Opus 4.8 via Bedrock.
For context, the top published entries on the leaderboard right now are Claude Code with Fable 5 at 83.8% and Codex with GPT-5.5 at 83.1%. Those are Anthropic’s and OpenAI’s own coding agents. Built by the labs that built the models.
Our submission is above every published result. It’s pending review on the leaderboard now, and you can go look at every trial yourself.
Who we are
Backboard is a fledgling startup out of Nepean, Ontario. Not San Francisco. Not a lab with a frontier model and a billion dollars of compute. A Canadian team building AI infrastructure, and a CLI coding agent as one piece of it.
Terminal-Bench is one of the harder agentic benchmarks out there: 89 real terminal tasks, the kind of messy multi-step work that actually looks like a developer’s day. The leaderboard is dominated by the biggest names in AI running their own models through their own agents.
So when our team’s harness put up a higher score on a model we don’t even make, that told me something I’ve believed for a while: the harness matters as much as the model. How an agent plans, delegates, and manages context decides how much capability you actually get out of any model you point it at.
How the run went down
The team didn’t cherry-pick. Five sequential jobs, one identical config, all 89 tasks, 5 attempts each. That’s 445 trials, every one of them included. Errored trials counted as zero. Pinned dataset, default settings, no overrides, all logs public.
Total cost of the entire run: $280.72.
I love that number. Frontier-level agentic coding performance, fully benchmarked, for less than a conference ticket.
We’re open-sourcing it
This is the part I’m most excited about. We’re going to open source the CLI. Soon.
A fledgling startup doesn’t win by hoarding. We win when developers actually use the thing, kick the tires, break it, and tell us what’s wrong with it. The benchmark score gets people to look. The tool has to earn the rest.
Try it
The CLI is available today at backboard.io/cli. It’s model-agnostic, so you’re not locked into any one provider.
If you try it, I’d love to hear what you think, good or bad. We’re small enough that your feedback actually reaches the people who built it. Usually the same day.
And to the team: you know who you are. This one’s yours.
Receipts: PR #200 on the Terminal-Bench 2.1 leaderboard repo | Terminal-Bench leaderboard
답글 남기기