Bullet Is Fast, but It's Built on Rented Land

Every few weeks another coding agent launches claiming a benchmark win, and mid-2026 has trained most of us to scroll past. Bullet, which launched out of YC's Summer 2026 batch this week, deserves a closer look anyway — not because its headline number means anything (it doesn't), but because its actual pitch, raw speed, is the right fight to pick. The problem is where the company is standing while it picks it.

The benchmark that stopped measuring anything

Bullet's founders, Adi Kulkarni and Alex Sima, lead with 479/500 — 95.8% — on SWE-bench Verified, at an average of 119 seconds per task, plus a marketing claim of 30–60% faster task completion than Claude Code and Codex.

Hacker News did the arithmetic within hours: frontier models now score 95–96.2% on SWE-bench Verified with ordinary harnesses. Claude Opus 5 posts 96% on its own. Bullet's 95.8% doesn't demonstrate that the harness adds correctness; it demonstrates that the harness doesn't subtract it. To their credit, the founders conceded the point in the thread and promised runs on harder suites like Terminal-Bench. They also admitted the eval excluded their model router — reasonable for an apples-to-apples comparison, but it means the one architectural piece they actually own wasn't in the test.

This is where agent evaluation is in 2026: SWE-bench Verified is a floor check, not a differentiator. When the top of the leaderboard is compressed into a single point of spread, a startup citing it is telling you about its marketing instincts, not its engineering.

The speed claim is the real product

Strip the benchmark theater and Bullet's thesis is sound: coding agents are now bottlenecked on wall-clock time, not intelligence. A model that's 1% smarter but takes eight minutes per task loses to one that closes the loop in two, because the expensive resource in the loop is you — waiting, context-switching, coming back to review.

Bullet attacks latency five ways: routing easy tasks to fast models and escalating hard ones, grep-style targeted code search instead of whole-repo embedding, aggressive context hygiene (bounded tool output, evicting stale results), batching independent tool calls in parallel, and generally shaving loop overhead. They self-report 16% fewer round trips and 27% lower cost against their baseline.

Here's the uncomfortable part, which an HN commenter also landed on: almost none of this is proprietary. Claude Code already parallelizes independent tool calls and spawns cheap subagents for search. Grep-first retrieval over embedding indexes is now the consensus position for agentic coding, not a contrarian bet. Context hygiene is prompt-and-harness discipline anyone can replicate. The genuinely defensible piece is the router — automatic model-and-effort selection per step — and that's precisely the part they didn't benchmark.

So the honest read on the 119s/task figure: plausible, unverified, and directionally important even if the exact number is soft. If it holds up, the interesting consequence isn't Bullet itself — it's that time-to-resolution and cost-per-resolved-task become the numbers vendors compete on. Nobody has standardized that measurement yet. Whoever does will define the next leaderboard, and it'll be a more useful one than SWE-bench ever was.

Built on land the landlord already cleared once

Bullet is free, the business model is "TBD," and the onboarding hook is that you bring your existing Claude Code or Codex subscription — Sima's words: "you pretty much just have to sign in to your account and it will connect."

If that sentence gives you déjà vu, it should. In January 2026, Anthropic deployed server-side checks that rejected subscription OAuth tokens from anything that wasn't the genuine Claude Code client, cutting off OpenCode and every similar harness overnight. By April, the policy was formal: Claude Pro, Max, and Team subscriptions don't cover third-party tools, full stop. The subscription-arbitrage era — running thousands of dollars of API-equivalent load through a $200/month plan via someone else's harness — is over, and it ended with an error message, not a deprecation window.

Bullet supports API keys and local models too, so the product survives enforcement. But the frictionless "sign in and go" funnel — the thing that makes a free tool spread — sits directly on a distribution channel that Anthropic and OpenAI control and that Anthropic has already shut off once this year. Add the other launch-week trust dings HN surfaced — closed source, a .deb installer that wants root, chat sharing enabled by default (since fixed) — and you have a tool asking for a lot of faith on a foundation it doesn't own. The founders say open-sourcing is coming. Until it ships, that's a promise, not a property.

What to actually do with this

If you're an individual developer: trying Bullet costs nothing (npm install -g @trybullet/cli, macOS/Linux, Node 18+), and doing it with an API key on a non-sensitive repo is a fine weekend experiment. Time your ten most common task shapes against your current agent; the 30–60% claim is testable in an afternoon, and self-reported vendor numbers are exactly the kind you should re-derive locally.

If you're deciding team tooling: don't move onto a closed-source, pre-revenue harness whose smoothest auth path may be one policy email from breaking. The switching cost isn't the install — it's the workflow habits and CI hooks you'd build around it.

Either way, steal the ideas now, because they work in the tools you already run. Route mechanical edits to a cheaper, faster model tier and save the frontier model for design decisions. Cap tool output sizes so one noisy log dump doesn't poison the context window. Push Claude Code or Codex to parallelize independent reads and searches instead of serializing them. Bullet's contribution this week isn't a moat — it's a well-argued reminder that most agent latency is harness sloppiness, not model inference.

The diagnosis is right: 2026's coding agents lose more time to loop overhead than to IQ. The company making that argument is closed-source, free-with-no-model, and parked on a rival's revoked driveway. Watch the Terminal-Bench numbers and the open-source release. Skip the benchmark press release.