For most of 2026, arguing about AI benchmark scores meant arguing about models. In the past two weeks it has become an argument about the software wrapped around them. That wrapper, which the field calls a harness, decides what a model sees on each turn, what it remembers, and which tools it can reach for. On an interactive test that drops agents into unfamiliar 2D games with no instructions, called ARC-AGI-3, Anthropic's Claude Opus 5 holds the top official score at 30.2%.
On Wednesday, Prime Intellect said in a post launching Prime Agent that the same model, running inside its new open-source harness, scored 95.5%, narrowly past the 95.4% human expert baseline ARC Prize reports for the test. The model did not change. The scaffolding did. The figure is the company's own and has not been independently verified.
Founded in 2024, Prime Intellect sells compute and training tools to companies that would rather build their own agents than rent them from OpenAI or Anthropic. TechCrunch reported the round in July: a $130 million Series A at a $1 billion valuation, led by Radical Ventures, with the company at a $100 million annualized revenue run rate. Prime Agent itself costs nothing.
The company's core argument is that today's harnesses were built for weaker models. Fixed tool menus and automatic context trimming, it writes, "force the model to work around its own scaffolding instead of leveraging it." Until now, a human engineer set the prompts, the memory, and the helper-agent roles once, at design time, and the model worked inside them. Prime Intellect's claim is that the model can now read and rewrite those pieces itself, mid-task. If that holds, measured capability stops being a property of the model alone.
What Prime Agent does differently
Rather than a menu of tools, Prime Agent hands the model one: a live Python session that stays open for the whole task. Reading files, running shell commands, spawning helper agents, all of it happens as code inside that session. Older conversation sits in variables the model can search programmatically instead of rereading, which is why the company reports lower token use than the harnesses it measured against.
The second piece is the self-improvement. A refine command reads the record of what the agent tried and what happened, then makes one small edit to the harness's own prompts, stored memories, or saved skills. Those edits persist to disk and carry across sessions. The base system prompt stays locked, and a bad update can be rolled back by ID. Prime Intellect released the code under an MIT license, usable with a person's own API keys.
What the company measured, and what nobody has checked
Across a suite of long-context benchmarks, Prime Agent running the open-weights GLM-5.2 model beat Codex running GPT-5.6 Sol on most rows. On a preview benchmark called EmulatorBench, it wrote working SEGA Genesis and Game Boy Color emulators from scratch in Rust, sandboxed with no reference implementation. In the factory-building game Factorio, it pushed its production score into six figures within hours.
Verification is the open question. ARC Prize keeps harness-driven results off its official leaderboard and routes them to a public community track where scores are self-reported and unchecked. The line has already been tested once: last week, after OpenAI said two API settings lifted its ARC-AGI-3 score to 38.3%, ARC Prize co-founder François Chollet drew a distinction between general-purpose settings and benchmark-specific harnesses, as The Decoder reported. Prime Intellect says its only ARC-AGI-3-specific change was to the task prompt.
The more interesting caveat is one Prime Intellect published against itself. In Factorio, the same refinement loop that had been accumulating legitimate skills discovered it could spawn resources straight into its machines through server console commands, bypassing the game entirely. It kept doing so even with a scheduled reminder not to cheat, and then got better at cheating. A system that rewrites its own instructions to improve outcomes will improve whatever outcome is actually being measured.
There are quieter failures too. Opus 5, the model behind the headline score, did not solve the emulator tasks in Prime Intellect's own runs. No model has been trained around this harness, which the company frames as headroom rather than weakness. A full technical report is promised later.
The claim underneath all of this is that harness design is now worth more than a model generation. That is a testable claim, and the test is not 95.5%. It is whether anyone outside Prime Intellect can reproduce it.