Worth a trial, but read the benchmark like Kahneman: Bullet, the free YC S26 coding agent launched August 13, 2026, resolved 95.8% of SWE-bench Verified at 119 seconds and $0.73 per task — genuinely fast, yet below Claude Opus 5’s 97.0% on the same benchmark, and its “35–67% faster” claim compares against a minimal research harness, not Claude Code itself.
Bullet launched on Hacker News on August 13, 2026 with the most focused pitch in the coding-agent category: same models, tighter loop, less waiting. Two founders out of AppLovin and Citadel, YC S26, and an origin story about being so tired of watching Claude Code grind that they built a replacement. I have burned enough hours watching agent spinners that I installed it the same week. But the benchmark deserves to be read slowly, because it is a small masterclass in how numbers persuade.
What Bullet actually is: a free coding agent — currently v1.4.15, installed with npm or a desktop app, no API key required to start — built around three mechanisms. It routes each prompt to the cheapest model that can handle it instead of sending everything to the frontier. It runs targeted code search rather than embedding your whole repository. And it executes independent tool calls in parallel while intercepting stuck loops. The founders’ internal measurement claims 16% fewer round trips and 27% lower cost per task. Claude Code, the incumbent it names as its final boss, requires a paid Claude plan — $20 a month for Pro, $100 to $200 for the Max tiers — which makes Bullet’s price of zero a real asymmetry while it lasts.
The headline result: 479 of 500 issues resolved on SWE-bench Verified — 95.8% — in a single attempt per task, graded by the official Docker harness, averaging 119 seconds and $0.73 per task, with GPT-5.6 Sol running underneath. Two-thirds of tasks finished inside two minutes. That is a genuinely strong engineering result, and publishing per-repository breakdowns and patch data is more transparency than most vendors offer.
Now the Kahneman reading. Thinking, Fast and Slow gives the exact vocabulary for what a well-built benchmark page does to you: System 1 constructs a coherent story from what is shown — what you see is all there is — and coherence, not evidence quality, is what confidence feels like. Three framing details change the story. First, on the very comparison table Bullet cites, Vals reports Claude Opus 5 at 97.0% and GPT-5.6 Sol at 96.2% on the same benchmark — Bullet’s accuracy is slightly below the frontier, so the honest claim is “nearly as accurate, much faster and cheaper,” not “better.” Second, the 35–67% speed advantage is measured against mini-SWE-agent, a deliberately minimal research harness — not against Claude Code as you actually run it. Third, this is a vendor-run evaluation of a three-week-old product. None of that is scandalous; it is ordinary marketing physics. But as Kahneman put it, subjective confidence is a feeling, not a judgment.
| Kahneman criterion | Bullet /5 | What I found |
|---|---|---|
| Outside view (does the claim survive independent framing?) | 3 | Official harness and published patches, yes — but no independent replication yet, and the speed comparison is against a research scaffold. |
| Base rates (is the test your reality?) | 2 | SWE-bench is 500 curated Python issues; your repo is not. Perfect scikit-learn scores say little about your gnarly TypeScript monorepo. |
| Reversibility (cost of being wrong) | 5 | Free, installs in a minute, changes nothing about your repo. Trying it risks almost nothing. |
| System 2 load (does speed help you verify?) | 2 | A faster agent produces more diffs per hour than you can deliberately review — speed moves the bottleneck to you. |
That last row is the one I care about as a coach, not an engineer. The constraint in agent-assisted development stopped being generation speed months ago; it is review capacity. When the loop tightens, the temptation is to let a lazy System 2 endorse whatever the agent proposes, because everything feels fluent and fluency reads as correctness. Bullet’s model routing adds a second, subtler delegation: the judgment about which model is good enough for this task is now made by the tool, invisibly. Kahneman’s premortem is the right ritual before switching — imagine it is six weeks from now and a subtle bug shipped to production; write the story of how. In my version of that exercise, the cause is never agent latency. It is review debt.
My verdict after a week of real use on well-scoped tasks: the speed is not imaginary, and for bounded fixes — a failing test, a clean refactor, a data-pipeline tweak — the tighter loop is a genuine pleasure. I have kept Claude Code for long-context work where I want the strongest model reasoning across many files without a router deciding otherwise. The decision rule I would give any founder: run both on twenty of your own tasks and count your corrections, because your codebase is the only benchmark with valid base rates. Free-and-fast is a fine reason to try a tool. It is not yet a reason to trust one.
Sources: Bullet’s Launch HN thread, codewithbullet.com, Bullet SWE-bench Verified methodology post.
Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · AI Coach App — Building It in 8 Hours