The Leaderboard Is Dead. Here's What I Actually Reach For.

A developer retired their benchmark leaderboard of 300+ models and now selects AI tools by job rather than score. As of mid-July 2026, they recommend Google Antigravity 2 CLI for urgent debugging, Codex for legacy code, Claude for greenfield projects, and Pi for hands-on use, while avoiding terra and luna for agent work. The developer emphasizes that these field notes are a snapshot that will be outdated by August.

Let It Break — part 2 Tags: ai agents devtools productivity Last post I killed my benchmark — 300+ models tested, leaderboard retired. The fair question: fine, no rankings — then how do you pick? Like this. By job, not by score. These are my field notes as of mid-July 2026, and half of them will be wrong by August. That's not a weakness. A snapshot that admits it's a snapshot is more honest than a leaderboard pretending to be permanent. Production bug, needed fixing yesterday: Google Antigravity 2 CLI "agy" — catchy, I know . It will burn tokens like there's no tomorrow. It also delivers — this is the one tool I throw at a problem when I need it debugged in five minutes, not fifty. The only failure mode is when it gets loopy, and you'll know within a minute. You're not optimizing cost during a fire. You're optimizing time-to-out. Also agy, and this is the one area Google simply has it. Throw it a screenshot, a link, whatever — it just makes it. Everyone else gets you close. Agy gets you exact . Codex. Old code, big code, code with history — Codex is your friend. gpt-5.6-sol is probably the pick; I used 5.5 and it was fine for genuinely complex work. Two models I'd currently avoid for agent work running under Hermes or Claw : terra and luna. Either they're not suitable for agentic loops, or it's early days and it'll get silently fixed. Test again next month; that's the whole methodology now. Claude. Great at spinning up new things, great at branding work — not pixel-perfect, but good enough, and good enough ships. The honest criticism: it's slow, and I can't tell you why. The strange part is that Anthropic's models are first-class citizens inside Antigravity and feel like second-class citizens inside Claude's own product — capped, throttled, something. The web UI is great. The CLI in yolo mode is palatable. Just. Of the major products, it currently feels the least polished. This is a July statement; it might change in days. But it's true today. Pi. When I'm driving the harness myself rather than delegating, Pi is my favourite flavour — transparent about what it's doing, adapts to my tastes instead of fighting them. It's also happy being driven by a local LLM, which matters more than people admit. Codex takes local models well too. Hermes agent, running DeepSeek Flash and Pro. Usually it drives Pi or Codex, sometimes Claude. It doesn't excel at anything. It does the job. Always. It's the most neutral thing in my stack — no surprises in either direction — and that's precisely why it's the one I hand entire ticket queues to. Excellence is for the specialists above. Reliability is for the thing that runs unattended. Notice what replaced the leaderboard: not better rankings — jobs . Fire, pixels, legacy, greenfield, hands-on, unattended. Each job has a current answer and every answer has an expiry date. The old me would have benchmarked all six of these tools against each other and published the scores. The current me writes down what I reached for this week and why, and lets it go stale in public. Field notes over leaderboards. Snapshots that know they're snapshots. See you in August, when half of this is wrong.