An F1 game generated entirely by Opus 5.
It built its own tooling to grade its work. 3 permanent tools, 136 single purpose, and over 2,000 screenshots.
We let it run for 19 hours. 137 agents, 1.7 billion tokens, $1,200 in usage. See the results
Better Stack
@BetterStackHQ
Tracing reimagined with eBPF and OpenTelemetry.
Instrument your clusters without any code changes.
Get an AI SRE and a robust MCP server.
Joined January 2021
- A semi-playable F1 game generated entirely by Opus 5. @mattshumer_ gauntlet Loop turns a few sentences into a full game. We let it run for 19 hours. 137 agents, 1.7 billion tokens, $1,200 in usage. Here is what came out ⬇️
- MCP was wrong from the start but they finally fixed it. The 2026-07-28 spec removes initialize and Mcp-Session-Id entirely. There is no connection to hold open anymore. No more sessions. We explain the changes
- TanStack Charts is my new favorite charting library. It brings real grammar of graphics to TypeScript, fully typed, framework-agnostic and tree-shakeable. Compose marks, scales & interactions to get nearly any chart you want. @tan_stack does it again 💪
- The model so cheap, it may as well be free: DeepSeek V4 Flash. It scores the same as Gemini 3.6 Flash but costs 18× less. I compared 8 models in this video. DeepSeek's a competent coder, but it falls short on UI design. Check out the results.
