VISTA is a visual harness. Running Claude Opus 5, it completes all 25 public ARC-AGI-3 games — a 100% win rate, a perfect 100 Relative Human Action Efficiency score, and 56.0% fewer actions than first-time human players.
The interesting part is not the number. It's that Schema reached ~99% on the same 25 games a few weeks earlier by building a harness on the opposite premise, and that a third result got a third of the way there by flipping two API switches. Three teams, one benchmark, three incompatible theories of what a model needs — and all three theories work.
What VISTA actually does
Three design choices, all of them subtractive:
Raw pixels, no schema. The agent observes a 512×512 PNG — the official 64×64 frame upscaled 8× by nearest neighbour, with one-pixel grid lines between cells. It is never told the world is a 64×64 grid. No object list, no state vector, no parsed representation: the harness hands over the image and lets the model's visual priors do the parsing. The authors note the agent forms and discovers visual concepts on its own as a result, which is the whole point of not pre-digesting the input.
Free-form language, no constraints. The agent decides what to reason about, in what language, for how long. It keeps two notes files — GUIDE.md for what survives across levels, WORKING.md as a scratchpad for the current one — with the instruction to build a compact, revisable model of the game. "Compact" pushes toward higher-level abstraction in the spirit of Occam's razor; "revisable" licenses the agent to be wrong early. Before each action it has to state the visual result it expects, which turns every move into a cheap falsifiable prediction rather than a guess.
Lossless visual memory. This is the real mechanism. Every frame the environment returns — including intermediate animation frames — is stored with its turn and frame index, uncompressed. An inspect tool brings any past state, animation frame, or enlarged region back into view through the same visual channel; read_pixels returns exact colour samples for details too small to eyeball. Several views can be requested together for read-only comparison.
That last one is worth naming precisely, because the authors do: it is an explicit attention mechanism, operating at frame, region, and pixel granularity, steered by the model rather than learned. A standard VLM's memory is the KV cache — implicit, lossy, compressed, bounded. VISTA leaves the pixel history intact outside the model and lets the model decide what to pull back in. When it approaches its context limit it writes a continuation state and resumes in a fresh context, with notes, visual memory and action history all still addressable.
The disagreement
Schema, introduced in July, makes the model think like a physicist: it writes each game's mechanics as an executable program, backtests that program against recorded history, and then plans by searching inside it at zero action cost. State grounding and mechanism discovery, solved jointly, in one editable symbolic world model. Its controlled comparison lifted a Claude Code baseline from 42.83% to 98.98% RHAE on these same 25 games.
VISTA does none of that. No program synthesis, no search, no symbolic world model, no verification loop against recorded history. It reasons "in a fuzzy but flexible manner" — the authors' phrase, and clearly a deliberate one — and its notes are prose, not code.
| Harness | Base model(s) | Public-set result | Core mechanism |
|---|---|---|---|
| ARC Prize official | Claude Opus 5 | 30.2% (semi-private, verified) | none by design |
| OpenAI, two settings | GPT-5.6 Sol | 13.3% → 38.3% | retained reasoning + compaction |
| Schema | Opus 4.8 + Fable 5 | ~99% (self-reported) | executable symbolic world model |
| VISTA | Claude Opus 5 | 100% (self-reported) | lossless pixel memory + prose notes |
Two harnesses built on contradictory epistemologies — compile the world into code, versus refuse to compile it at all — land within a point of each other at the top of the same 25 games. That tells you something, and it isn't that one of them found the right theory. It's that at ~99% the benchmark has stopped discriminating between them, and that most of the credit belongs to the base models both are wrapped around. This blog already argued that the v3 leaderboard ranks model-harness pairs rather than models. Two mutually exclusive harnesses saturating it is the strongest version of that argument yet.
The thing all of them have in common
Strip the architectures and every harness result of the past two months is the same move: stop throwing state away.
OpenAI's two settings — retained reasoning and compaction — tripled GPT-5.6 Sol's public-set score from 13.3% to 38.3% and cut output tokens 6×, purely by not discarding the model's private reasoning after every action. Schema refuses to discard inferred mechanics, freezing them into a program that can be backtested instead of re-derived. VISTA refuses to discard pixels. None of them touched a weight.
And it's not confined to ARC. Meta's Muse Code appends every model call, tool run, approval and edit to a local event log so a crashed session resumes exactly where it stopped rather than re-deriving state. Cursor's training megakernel is built for bitwise determinism so a run can be reproduced rather than re-guessed. Liquid's harness-proxy trajectory capture records what actually happened inside real harnesses instead of simulating it. Y Combinator's QM gives every agent a durable sandbox and a scoped identity that survives the session.
Different layers, one instinct. Transformers forget by construction: the KV cache is bounded, lossy, and not addressable by the model as memory. The harness is where you put the thing that doesn't forget — and 2026's harness advances are, almost without exception, memory-architecture advances wearing different costumes.
Which reframes the question the user of any of these systems should be asking. Not how good is the model, but what does this harness let the model stop re-deriving. That is now a bigger lever than a model generation.
What the 100 does not mean
VISTA's authors are unusually direct about the limits, and they deserve credit for it: the models were released after the public ARC-AGI-3 games shipped, so training exposure cannot be ruled out, and they say plainly that the private set remains the real test of generalisation.
Three more things belong on the caveat list:
The 30.2% and the 100% are not the same eval. ARC Prize's verified 30.2% for Opus 5 is on the semi-private set under a deliberate no-harness methodology; VISTA's 100% is on the 25 public games. Reading it as a single model going from 30 to 100 is the wrong arrow. What it does show is how much headroom a harness has on games where the model's raw ceiling was assumed to be the binding constraint.
Self-reported means the community leaderboard. ARC Prize runs two boards precisely for this: the official one, no harness, verified, same observations and action limits for every provider; and a community board that permits harnesses, accepts self-reported scores, and is explicitly not verified. ARC Prize calls harness research economically valuable and cautions in the same breath against reading its scores as AGI progress. Both halves of that are correct.
The human comparison is not apples to apples. RHAE scores an agent's action count against a first-exposure human baseline — people getting one pass at an unfamiliar game. VISTA's agent gets lossless rewind to any earlier frame, arbitrary zoom, exact pixel sampling, and persistent notes across a context reset. Using 56% fewer actions than a human is a real result about the harness; it is not evidence the model explores more cleverly than a person would with the same affordances. Worth remembering, too, that a perfect 100 RHAE means matching or beating the human action baseline on every level — per-level credit is capped — not that it is a hundred times more efficient.
What to expect next
- The private set is the only test that settles this. Two harnesses at ~99–100% on 25 known games is a claim about 25 known games. Whichever of these designs survives on environments its authors have never seen is the one that was actually about memory rather than about familiarity.
- The 3D extension is the more interesting claim. VISTA notes its framework isn't restricted to 2D and demonstrates a 3D observation case. Lossless sensory memory with model-steered re-inspection is a far more consequential idea in an embodied setting than in a grid game, and that's where it should be judged.
- Expect harness ablations to become standard release material. Schema published a controlled baseline comparison; OpenAI published a settings ablation. Once "which harness" moves this much score, publishing a model number without the harness it ran in stops meaning anything.
References: VISTA — A Visual Harness · Schema harness · OpenAI — How enabling two settings tripled our scores on the ARC-AGI-3 benchmark · ARC Prize — Community Leaderboard · ARC Prize — Verified Testing Policy · follow-up: Prime Agent and the third theory of memory · related coverage: ARC-AGI, Explained · Meta's Muse Code · Open Kernels You Can't Actually Run · YC's QM harness