Hacker News treated Dirac's TerminalBench result as more than another benchmark screenshot. The curiosity came from the method. Dirac says it reached 65.2% on Terminal-Bench-2 with gemini-3-flash-preview, edging past Google's published baseline and slightly clearing Junie CLI, while cutting API cost by roughly two-thirds. That combination, better score with lower spend, is exactly the kind of claim HN likes to pull apart.

The project README makes the pitch in concrete terms. Dirac tries to keep context small, uses hash-anchored parallel edits when touching files, and leans on AST-aware retrieval to decide what code to pull into the prompt. In plain English, it is trying to stop the usual coding-agent failure mode where the model drowns in its own context window and starts paying for irrelevant tokens. That framing landed in the thread. Community discussion kept circling back to whether the real advantage was not the model at all, but the surrounding machinery that chooses what the model sees.

That also explains the first wave of skepticism. One of the main HN questions was whether Dirac should be understood as a harness, a fine-tuned system, or both. Another thread asked how portable the gains are: could the same editing strategy help when the underlying model is Qwen or another open model, or is the current result tightly coupled to Gemini flash? Those are fair questions, and the project page does not pretend they are settled. What it does provide is a clearer technical story than most benchmark brag posts: anchored diffs, structured code lookup, and an explicit argument that token efficiency is a performance feature.

The more interesting read from HN is that coding-agent evaluation is moving away from raw frontier-model worship. Readers were less impressed by “we topped a leaderboard” than by the idea that better file targeting and smaller prompts can change outcomes. That is a useful shift. If the next round of agent competition is decided by who can search less junk, edit more precisely, and stay coherent over large repos, then the benchmark conversation gets a little more serious.

For now, Dirac looks like a strong example of the new open-agent playbook: do less prompt stuffing, expose the mechanisms, and show the cost table next to the score. HN's reaction suggests that builders are ready to reward that level of engineering detail, but only if the claims stay reproducible and model-agnostic enough to survive outside a single leaderboard run.

#terminalbench

HN notices what made Dirac top TerminalBench: fewer tokens, sharper edits