One phone prompt rewrote opencode2, dax says it cost $500
The same week, Prime Intellect says a swarm of over 2,000 agents spent two weeks rewriting its own agent in Rust.
dax says Ryan rewrote all of opencode2's server and TUI from a single prompt sent from the opencode iOS app. According to dax, the run was Opus 5.5 orchestrating a bunch of subagents, and it cost $500.
it was just opus 5.5 orchestrating a bunch of subagents it cost $500
That is the whole report. There is no breakdown of how long it ran, how much of the output survived review, or what the before and after look like. One prompt, one orchestrator model, one number on the bill.
The other scale of the same idea
Prime Intellect published a much larger version of the same experiment in the same week. Over two weeks, it says Prime Agent orchestrated a swarm of over 2,000 agents to rewrite itself in Rust, using more than 10,000 sandboxes, more than 200B GLM-5.3 tokens and 16,000 agent-to-agent messages. The company calls it its largest test yet of Prime Agent's multi-agent capabilities and of the infrastructure running the swarms.
Prime Agent orchestrated a swarm of over 2,000 agents to rewrite itself in Rust.
Prime Intellect does give results for the rewrite. It says Prime Agent now reaches usable input about 13 times faster and uses 83% less startup memory than before. Those are its own numbers on its own product, not an independent benchmark.
What is actually comparable here
Not much, which is the interesting part. Both are self-rewrites of agent tooling driven by agents, and both were announced as achievements rather than as writeups. One is a single operator firing a prompt from a phone with a single frontier model fanning out to subagents. The other is a two week industrial run across thousands of sandboxes with an open weights model doing the token volume.
For working engineers, the thing neither post answers is quality. A rewrite that compiles and starts faster is a measurable claim. A rewrite that a maintainer still wants to live in six months from now is not something either party has tested yet, and nobody has looked at both codebases side by side. The cost gap is the only clean comparison available, and $500 against 200B tokens is a wide range for what is nominally the same move.