Prime Agent Crossed the Human Line by 0.1 Points. The Real Story Is on the X-Axis.

Prime Intellect open-sourced Prime Agent under MIT — a general-purpose coding harness, not an ARC-specific one. Running Claude Opus 5, it scores 95.5% on ARC-AGI-3, solving 179 of 183 levels, against a charted human-expert baseline of 95.4%.

Ignore the margin. A tenth of a point over a human baseline is not a crossing, it's a coincidence, and treating it as a threshold event is the kind of reporting this benchmark keeps producing. Three other things in this release are real.

Read the x-axis

Prime Intellect's headline chart plots score against output tokens per game on a log scale, which is the correct way to present a harness result and rare enough to deserve credit. Four harness runs and three reference lines:

Four readings the table alone won't give you:

The token gap is two orders of magnitude. On Prime Intellect's own axes, the official-harness Sol curve is still climbing toward its 13.3% at around 3M output tokens per game. Prime Agent running the same model passes that score somewhere under 20K. Same weights, same games, roughly 100× the token efficiency — and the OpenAI settings result already showed 6× of that was available from two API switches alone.

Sol and Opus 5 have different scaling shapes. Below roughly 100K tokens per game Sol is ahead of Opus 5 — cheaper, faster to useful. Past the crossover Opus 5 keeps climbing while Sol flattens at 78.3%. If you only ever ran a 50K-token budget you would conclude Sol is the better model here, and you would be wrong at 300K. Single-budget model comparisons are worth roughly nothing once the harness scales with compute.

The harness gain is not uniform across models. Terra scores 25.7% where Sol, its sibling in the same GPT-5.6 family, gets 78.3%. Whatever Prime Agent is exploiting, one model in a family can have it and another can lack it. "This harness lifts models by N×" is not a statement that transfers.

GLM 5.2 gets 8.6%. It's in the chart because Prime Agent runs open and closed models alike, and publishing the number where your harness barely helps is the behaviour you want from a self-reported result. It's also a real gap: the open-weights substrate that has been closing distance on coding and reasoning is nowhere near the frontier on this one.

The divergence between score and levels solved is worth a line too: Sol completes 164 of 183 levels — 89.6% — but scores 78.3%, because RHAE squares the action-efficiency ratio and caps per-level credit. Fumbling to a win costs you. Note also that the 95.4% human-expert baseline charted here is a different construction from the 100% human reference used on ARC Prize's own v3 material; the two aren't interchangeable, which is one more reason not to read 95.5% as a line being crossed.

The RLM: don't put the data in the context

The architecture is a Recursive Language Model — an idea from Zhang, Kraska and Khattab at MIT CSAIL in December, which Prime Agent is the first production harness to build on wholesale.

The premise: rather than ingesting large input data directly, the model gets a persistent Python REPL — in Prime Agent, a persistent IPython kernel as its only tool — and the input sits in that REPL as a variable. The model writes code to inspect, chunk and transform it, and calls sub-LLMs from inside the REPL via rlm(...). Sub-agent answers land in Python variables. The final answer is returned from a variable, not composed in the parent's context.

Follow what that means for the parent model's context window. It holds the system prompt, the user prompt, its own reasoning, and REPL output it explicitly chose to print. It never holds the data. The context window stops being a container and becomes a control plane — which is why the original paper reports handling inputs about two orders of magnitude beyond the model's context length, and why it frames the target as context rot rather than context limits. Not "how do we fit more in," but "why is it in there at all."

Three harnesses, three theories of memory

Yesterday's post argued that 2026's harness advances are memory-architecture advances wearing different costumes, and that VISTA and Schema saturated ARC-AGI-3 from contradictory premises. Prime Agent is the third premise, and it's the most radical:

Schema compiles the world into a program. VISTA refuses to compile and keeps raw pixels addressable. Prime Agent declines to load the world, and manipulates it by reference. Three incompatible answers, all of them clearing 78%+ on a benchmark where every frontier model scored under 1% at launch. The consistent finding across all three is negative and worth stating plainly: none of them touched a weight, and the disagreement between them tells you the benchmark has stopped discriminating between architectures at the top.

The part that isn't about ARC at all

The second abstraction is a Continual Harness: supplemental prompts, memories, skill descriptions and subagent specifications held as durable state that the agent updates through evidence-backed revisions. The /refine command lets it rewrite its own prompts, memory and skills mid-task based on what is and isn't working. Skills are importable Python packages. There's a background daemon you can detach from and reattach to, direct agent-to-agent messaging, automatic compaction, persistent goals, and an autonomous mode with configurable budgets.

That closes a loop this blog has been circling all year. If the harness is where memory lives, and a harness swap moves a score further than a model generation does, then a harness that edits its own state is doing self-improvement at the scaffolding layer — no gradients, no training run, no weights. It's the same instinct as Muse Spark 1.1 generating and grading its successor's training data, moved out of the training pipeline and into the runtime, where iteration costs an API call instead of a cluster.

It is also, and Prime Intellect says this in its own README rather than burying it, not a security sandbox. Prime Agent executes model-generated Python and project commands with your user permissions. "Programmatic everything" is precisely the property that makes the RLM work and precisely the property that makes a self-modifying agent with a live interpreter something you should run in a container you're willing to lose. Both facts come from the same design decision.

What to expect next

• The private set, again. These are self-reported public-set numbers, which by ARC Prize's own structure belong on the unverified community leaderboard — the same caveat that applies to Schema and VISTA, and the same one that will settle all three.
• The non-ARC claim is the one to test. Prime Intellect's strongest assertion is that the gain isn't benchmark-specific — it's a coding harness that happens to score well here. That's checkable on coding and long-horizon agentic benchmarks by anyone, and unlike an ARC score it would actually matter to people shipping software.
• Expect token-axis charts to become the norm. Once one team publishes score-versus-tokens on a log scale and it reveals a crossover, a plateau and a 100× efficiency gap that no single number captures, publishing the single number starts to look like a choice.

References: PrimeIntellect-ai/prime-agent (MIT) · Prime Intellect — Prime Agent · Zhang, Kraska & Khattab — Recursive Language Models · alexzhang13/rlm · ARC Prize — Community Leaderboard · related coverage: Two Harnesses Solved the Same Benchmark by Disagreeing About Memory · ARC-AGI, Explained · Meta's Muse Code · YC's QM harness