Reports on October 15, 2026 say ARC Prize has measured Google's Gemini 4 Argon at 97.5% on ARC-AGI-3, which would make Google DeepMind the second frontier lab above 95% on the interactive reasoning benchmark. The catch sits in the same sentence of the reports: that score uses the provider adapter harness. Under the standard harness, Argon reportedly scores 49.8%.
That is a gap of nearly 48 points for the same model, and understanding it matters more than the headline. This post explains the two harnesses, lays out what is claimed and what we could verify, and gives context from earlier ARC-AGI-3 results.
TL;DR: the reported numbers
| Benchmark / harness | Reported score | Reported cost |
|---|---|---|
| ARC-AGI-3, provider adapter harness | 97.5% | about $4.8K |
| ARC-AGI-3, standard harness | 49.8% | about $5.8K |
| ARC-AGI-2 | 91.3% | $0.32 per task |
| ARC-AGI-1 | 98.5% | $0.16 per task |
Source: an aggregator summary of posts from the ARC Prize account and François Chollet. We were unable to load the underlying X posts or the leaderboard's data table, so treat these as reported, not confirmed. The most useful check is the ARC Prize leaderboard.
What is ARC-AGI-3?
ARC Prize describes ARC-AGI-3 as "the first interactive reasoning benchmark designed to measure human-like intelligence in AI agents." Instead of static puzzles, agents explore novel environments, acquire goals on the fly, build world models and learn continuously. ARC Prize says a 100% score "means AI agents can beat every game as efficiently as humans," and that the environments are 100% human-solvable.
That interactive design is why the harness matters. In a static puzzle, the model sees the problem once. In an interactive game that lasts many steps, the model must remember what it tried, and how the harness handles memory changes the result.
The two harnesses
The leaderboard's own explainer describes them: ARC-AGI-3 results use either the Standard harness, which "carries forward model-selected notes throughout the environment," or the Provider Adapter harness, which "preserves reasoning state and compacts longer conversations so the model can reuse prior work."
In plain terms, the standard harness gives every model the same simple memory: whatever notes it chooses to keep. The provider adapter lets the model vendor plug in its own state handling, which can carry reasoning forward and compress history. A lab that has trained its model to work with persistent reasoning state benefits a great deal from the adapter. That is the same story we covered when OpenAI reported tripled scores from retained reasoning and compaction.
The leaderboard also notes that only systems costing less than $10,000 to run are shown, that models unable to produce full outputs have the remaining tasks marked incorrect, and that results marked "preview" are unofficial and may rest on incomplete testing.
What the numbers say, and what they don't
What they plausibly say: Argon is a very strong ARC-AGI-1 and ARC-AGI-2 model (98.5% and 91.3% as reported), and with a harness that preserves its reasoning state it can solve almost all ARC-AGI-3 environments at a cost near $4.8K for the run.
What they do not say: that the model is 97.5% as capable as a human learner in general. The 49.8% under the standard harness shows that half the score depends on scaffolding. It also does not say the adapter run is a fair apples-to-apples against rivals unless they were tested with the same adapter rules.
Commentary quoted in the reports is split. Chollet is quoted saying Argon has a similar ARC-AGI-3 performance profile to Astra and that Google is "back at the frontier." Another observer, who posts as scaling01, is quoted putting an 85% probability on reinforcement learning on games being the cause of the jump and 15% on looping. Both are opinions, not findings; we link no evidence either way.
How does this compare with earlier ARC-AGI-3 results?
- July 2026: Claude Opus 5 was verified at 30.2% on ARC-AGI-3, about four times the previous OpenAI score.
- July 2026: OpenAI reported that retained reasoning plus compaction lifted a public-set score from 13.3% to 38.3%.
- August 2026: NVIDIA's AVO agent claimed 100% on the public set in a long-horizon agent setup, with caveats about the public set and the method.
- September 2026: GPT-6 Astra launched with a 99.9% ARC-AGI-3 score under OpenAI's provider adapter harness, a detail OpenAI disclosed.
- October 15, 2026: Argon reportedly reaches 97.5% under the adapter and 49.8% under the standard harness.
The trend is clear: scores that were near zero for frontier models a few months ago now sit near the ceiling when the harness lets a model keep its own reasoning state. That says as much about harness design as about raw model ability.
A worked example: why memory handling moves the score
Imagine an agent dropped into a small puzzle game with no instructions. The first twenty steps are exploration: it presses buttons, watches what changes and forms a guess about the goal. Around step fifty it realises a pattern it saw at step five was a clue. An agent that can still "see" its early reasoning can act on that realisation. An agent that kept only a short note such as "red tile moves left" may have lost the detail that mattered.
The standard harness leans on notes the model chooses to write. The provider adapter lets the vendor keep reasoning state and compress older turns instead of dropping them. For long games with sparse feedback, which is exactly what ARC-AGI-3 uses, the second approach preserves far more of what the agent learned. That is plausibly why the same model lands near 50% on one harness and near 98% on the other.
This does not make the adapter result illegitimate. ARC Prize publishes both, and real products can implement the same kind of state handling. It does mean that a lab with strong reasoning-state tooling has a structural advantage on this benchmark, and that comparing a standard-harness number from one lab with an adapter number from another would mislead.
What do the ARC-AGI-1 and ARC-AGI-2 numbers add?
The reported 98.5% on ARC-AGI-1 and 91.3% on ARC-AGI-2 describe the static-puzzle generations of the benchmark, where the model sees a grid transformation task and must infer the rule. Those benchmarks are close to saturated at the top end, which is why attention has shifted to the interactive version. The costs reported, $0.16 and $0.32 per task, are low enough that they say more about efficiency than about difficulty.
The important contrast is between static and interactive reasoning. A model can be nearly perfect at inferring a rule from a few examples and still need extra memory machinery to finish a long game. Argon reportedly shows both profiles at once: top-tier static scores, and an ARC-AGI-3 result that swings with the harness.
Why does Google's position matter?
If the reports hold, Google sits alongside OpenAI at the top of this benchmark. Chollet's quoted remark that Argon has a performance profile similar to Astra's, and that Google is back at the frontier, is the main evidence, and it is a view about the profile rather than a head-to-head. For developers choosing between providers, it suggests the top three or four labs are now close on this benchmark, and that differences will show up in cost, latency and tooling rather than in raw puzzle ability.
How to read ARC-AGI-3 scores from now on
- Always ask which harness. A score without one is incomplete. Compare standard to standard and adapter to adapter.
- Look at cost. Argon's reported $4.8K for the adapter run and $5.8K for the standard run are totals for the test set, not per-task numbers like ARC-AGI-1 and 2.
- Separate public and private sets. Public-set results can be gamed more easily than private-set results. The leaderboard's verification policy says what is shown.
- Do not treat 100% as AGI. ARC Prize itself frames a 100% score as matching human efficiency on these games, not as general intelligence.
- Wait for the official entry. If the leaderboard lists Argon, note the date, the harness, the "preview" flag if any, and the cost.
For model-by-model context on Google's flagship, see our Gemini 4 Argon vs Opus 5.5 vs Grok 4.7 vs GPT-6 Astra comparison and the top use cases people built with Argon. For how benchmarks work in general, our AI benchmarks complete guide is the place to start.
What this means for what you build or pay
For builders, the lesson is practical. The harness around a model, meaning how it stores notes, compacts history and carries reasoning forward, can move results as much as the model itself. If you build agents that run for many steps, test your own memory and compaction strategy, because it may matter more than switching models. For buyers, a headline benchmark number is a reason to run your own tasks, not a reason to switch.
What we could not confirm
- The official ARC Prize entry for Argon and its verification status.
- The exact harness configuration used for the adapter run.
- Whether the "second lab" ranking counts every lab or only those the leaderboard lists.
- Whether Google published its own statement about the ARC-AGI-3 result.
We will update this post if the leaderboard or Google publishes more.
Related reading
- Opus 5 hits 30.2% on ARC-AGI-3
- OpenAI ARC-AGI-3: retained reasoning and compaction
- GPT-6 Astra launch, benchmarks and pricing
- NVIDIA AVO and ARC-AGI-3
- Gemini 4 Argon vs Opus 5.5 vs Grok 4.7 vs GPT-6 Astra
- AI benchmarks complete guide
Scores are as reported on October 15, 2026 and may change as ARC Prize updates its leaderboard.