# Prime Intellect's Open-Source Harness Lifts Claude Opus 5 to 95.5% on ARC-AGI-3 Without Touching the Model

The gain came from a persistent code runtime and versioned agent memory, not new weights, and the score is self-reported rather than verified by the benchmark's authors.

- Published: 2026-08-31T06:23:05.714Z
- Canonical: https://polylog.news/ai/2026-08-31/prime-intellect-s-open-source-harness-lifts-claude-opus-5-to
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Anthropic News](https://www.anthropic.com/news/claude-opus-5)

Prime Intellect has published the [technical report](https://arxiv.org/abs/2608.23552) behind Prime Agent, the open-source agent harness it released in early August, and the results argue that a large share of current agent capability resides outside the model itself. Running [Anthropic's Claude Opus 5](https://www.anthropic.com/news/claude-opus-5) without any additional training, the harness [raised ARC-AGI-3 performance from about 30 percent to 95.5 percent](https://www.primeintellect.ai/blog/prime-agent) on the benchmark's Best@1 measure, slightly above the 95.4 percent human expert baseline the benchmark reports.

The metric matters as much as the number. ARC-AGI-3 scores agents on Relative Human Action Efficiency, which compares the number of actions an agent takes in an interactive environment against a human baseline and squares the ratio, so an agent that needs twice as many moves keeps only a quarter of the credit. [The benchmark](https://arcprize.org/arc-agi/3) gives no natural-language instructions and requires the agent to infer goals from interaction. Frontier systems scored below one percent on it earlier this year.

Prime Agent's two ideas are a Recursive Language Model, in which the model works inside a persistent IPython kernel and calls sub-agents as ordinary function calls rather than packing everything into one context window, and a Continual Harness that stores prompts, memories, executable skills and sub-agent specifications as typed, versioned state carried across runs. The report also describes an 85.5-hour autonomous nanoGPT training run with 19 validated records, and says the harness matches or beats the vendors' own scaffolds, including Claude Code and Codex, on long-context coding, GPU-kernel generation and emulator construction.

Two caveats belong next to that figure. The runs are [self-reported and have not been independently reproduced](https://www.marktechpost.com/2026/08/06/prime-intellect-releases-prime-agent/) by the ARC Prize organizers, and Prime Intellect, which sells decentralized training and inference capacity, benefits directly if the industry concludes that the harness rather than the checkpoint sets the ceiling. The Russian-language channel [AI ML Big Data](https://t.me/ai_machinelearning_big_data/10805) summarized the report as an operating system for long-lived agents, which is closer to the claim than "smarter model" is.

## What this means

If a free, open-source runtime can move a frontier model from 30 percent to human-level on a benchmark it previously failed, the defensible product is the model plus tokens, not the proprietary scaffold. That erodes the pricing power of vendor-owned agent shells such as Claude Code and Codex, and it shifts spend toward raw inference, because a harness that plans, retries and rewrites its own skills consumes far more tokens per completed task. Anthropic and the GPU capacity providers gain volume from this pattern. Anyone selling scaffolding as the product loses differentiation.

## What to watch

- Whether ARC Prize or an independent group reproduces the 95.5 percent figure on held-out environments, which would separate a genuine reasoning gain from harness overfitting to the action-efficiency metric.
- Whether Anthropic, OpenAI or Google fold continual memory and versioned skills into their own agent products, which would signal they see the harness as the contested layer.
- Token consumption per completed agent task at the major inference vendors, since a harness that runs for tens of hours changes the unit economics of agent products more than model pricing does.