The Same 87.5% on ARC-AGI-1 Cost $4,560 in 2024 and 30 Cents in 2026. ARC-AGI-4 Lands in 2027.

Abstract

ARC-AGI is the benchmark François Chollet built to measure skill-acquisition efficiency rather than accumulated knowledge. Since 2019 ARC Prize has rebuilt it three times specifically to keep it unsolved, and now says it intends to ship a new version every year, each unsaturated on release. ARC-AGI-4 is scheduled for early 2027.

Two things have happened at once, and they are usually reported separately.

The test got harder on purpose. ARC-AGI-1 (2019) is a static grid puzzle. ARC-AGI-2 (March 2025) rebuilt the task set to defeat brute-force search. ARC-AGI-3 (March 2026) abandoned static puzzles for interactive games with no instructions and no stated goal, scored on how many actions an agent needs relative to a human. Humans solve all three at essentially 100%. Frontier systems scored below 1% on ARC-AGI-3 at launch.

The price of passing collapsed, and one pair of runs makes it exact. In December 2024 ARC Prize published o3 at 87.5% on the ARC-AGI-1 semi-private set for about $4,560 a task. In August 2026 it published Grok 4.6 at 87.5% on the same set for $0.30 a task — identical score, identical benchmark, identical split, twenty months apart. That is 15,200×, or roughly 320× a year, with no scoring caveat attached to it. Gemini 3.7 Flash now beats both at 95.5% for $0.12.

This piece argues that the second fact is the more interesting one, that the benchmark treadmill is what keeps it measurable, and that cost per task is now the load-bearing number — with two large caveats. Costs across different splits and harnesses are not comparable, and the harness has become so powerful that ARC Prize's own report shows the same model scoring 0.0% and 97.1% on the same environment depending on the wrapper around it.

We then decompose the collapse — and find that at most 35× of a 15,200× fall is hardware, on NVIDIA's own best claim. The rest is software: mostly the decision to stop sampling a thousand times. And we look at where open-weight models sit, which is level on the saturated version and a third of the way behind on the unsaturated one.

Table of contents

1. What ARC-AGI actually measures
2. What "Verified" means, and why the split matters
3. The three versions, and the fourth
4. Seven years of scores: who took it, and how
5. The cost curve: what a solved task has cost
6. The current frontier, model by model
7. Open weights against closed: level on v1, a third behind on v2
8. The prize that pays for efficiency, not for score
9. ARC-AGI-3: the harness became half the benchmark
10. GPT-6 Astra: 62.7% and 99.9% from the same model
11. How much of the gain is hardware, and how much is software
12. What it means for the people selling the compute
13. What the cost collapse does and does not mean
14. What would change the conclusion
15. Primary sources worth reading

1. What ARC-AGI actually measures

The definition is unusual and everything else follows from it. ARC Prize's framing, from Chollet's 2019 paper On the Measure of Intelligence:

The intelligence of a system is a measure of its skill-acquisition efficiency over a scope of tasks.

Not how many things can it do. How cheaply does it learn to do a new one. That is why cost sits on the leaderboard axis instead of in a footnote: a system needing a thousand attempts to solve a puzzle a person solves in thirty seconds has demonstrated search, not skill acquisition.

Two design choices protect that:

• Core Knowledge priors only. Tasks use objectness, counting, basic geometry and topology — things humans have early and without instruction. Language, cultural knowledge and trivia are deliberately excluded, so a large training corpus is not an advantage.
• Novelty on first contact. Every task is meant to be one the test-taker has never seen and the designers never anticipated. That is what makes memorisation useless, and what makes each version expire once models catch up.

The result is a benchmark that is easy for humans and hard for machines, and one that has to be replaced as soon as it stops being hard.

2. What "Verified" means, and why the split matters

Almost every number in this piece is an ARC Prize Verified result, and the word is load-bearing.

Verified means ARC Prize Foundation staff ran the model themselves, on a hidden test set, under an academic oversight panel drawn from NYU, UCLA, Columbia and the Santa Fe Institute. In ARC Prize's words, "only scores evaluated on our hidden test set through our official verification process will be recognized as verified performance scores." Self-reported figures, it notes, vary in dataset curation and prompting method in ways that prevent apples-to-apples comparison.

There are four splits and they are not interchangeable:

• Public eval — downloadable. 400 tasks on ARC-AGI-1, 120 on ARC-AGI-2. Contaminated by definition; useful for development, not for claims.
• Semi-private — 100–120 held-back tasks used for verified API runs. Most figures below.
• Private — never exposed to an API. Used only for the Kaggle competition, inside a sealed notebook. This is the number Chollet designed.
• Community leaderboard — self-reported, unverified, harness-driven. ARC Prize explicitly cautions against reading it as evidence of progress.

When a score moves by 60 points between two of these, that is usually the split talking, not the model.

3. The three versions, and the fourth

ARC-AGI-2 exists because o3 broke ARC-AGI-1 with test-time search. ARC Prize stripped out tasks susceptible to brute force, ran controlled testing on over 400 human participants to calibrate difficulty, and added tasks targeting symbolic interpretation, compositional reasoning and contextual rule application. Every task in it has been solved by at least two humans within two attempts. The average individual human scores 60% on public eval — down from 64.2% on ARC-AGI-1 — while the panel collectively reaches 100%.

ARC-AGI-3 is not a harder puzzle set; it is a different genre. An agent is dropped into a 64×64, 16-colour game with, in ARC Prize's words, "no instructions, no rules, and no stated goals." It must explore, infer the mechanics, work out what winning even means, and carry what it learns into later levels. Scoring is RHAE — Relative Human Action Efficiency — comparing the agent's action count on each level against the upper-median best human action count, level by level. A score can exceed 100% (capped at 115%) if the agent is more efficient than the human baseline. Beating ARC-AGI-3 means matching human action efficiency, not merely finishing.

ARC-AGI-4 is scheduled for early 2027. Chollet has said ARC Prize is moving to a yearly release cadence, each benchmark designed to be fully unsaturated on release — which requires guessing where capability will be a year out. That treadmill is what this piece is about.

4. Seven years of scores: who took it, and how

Four eras, and they are distinct:

• 2019–2021 — brute-force program search. The first Kaggle competition (2020, $20,000 pool, 913 teams) was won by icecuber at about 20% on the private set, searching over hand-written primitives with no deep learning. That was taken as evidence Chollet was right that scale alone would not do it.
• 2022–2023 — the DSL grind. Lab42's two ARCathon competitions, $100,000 in prizes each, pushed the private-set state of the art to roughly 33% by the end of 2023.
• 2024 — the jump. ARC Prize 2024 moved the private record from 33% to 55.5%, on two techniques: deep-learning-guided program synthesis, and test-time training — fine-tuning the model on the task in front of it. MindsAI posted the 55.5% but chose not to open-source it, so it was ineligible for prize money.
• December 2024 — o3. ARC Prize verified o3 at 75.7% on the 100-task semi-private set in its high-efficiency configuration and 87.5% in its low-efficiency configuration. The second used 1,024 samples per task and 5.7 billion tokens.
• 2025–2026 — saturation. GPT-6 Astra is verified at 98.5%, Claude Fable 5.1 and Claude Opus 5 at 97.5%, Gemini 3.7 Flash at 95.5%, Grok 4.6 at 87.5%. ARC-AGI-1 is now a regression test.

The pattern that matters is the one the 2026 survey The ARC of Progress towards AGI found across 82 approaches: performance falls 2–3× from ARC-AGI-1 to ARC-AGI-2 for every paradigm — program synthesis, neuro-symbolic and neural alike. No architecture escapes the cliff. Humans stay at 100% on all three versions.

5. The cost curve: what a solved task has cost

The summary table at the top is the argument in one card. December 2024's o3 run is the canonical "we bought the score" moment, and ARC Prize published the whole grid:

Note the labels. ARC Prize calls the cheap run high-efficiency. The 87.5% headline everyone quoted was the low-efficiency configuration, burning 172× the compute of the one beside it.

The clean comparison. In August 2026, Grok 4.6 scored 87.5% on ARC-AGI-1 semi-private at $0.30 a task at medium reasoning effort. Same score, same benchmark, same split, twenty months later, for 1/15,200th of the price — about 320× a year. There is no score adjustment or split adjustment to argue about. It is the same number on the same test.

The axis above is logarithmic — every gridline is ten times the one below it — because on a linear axis four of those five bars have no visible height at all. That is the honest shape of the thing: a decline this steep can only be drawn as a decline in orders of magnitude.

Loosen it and the range holds. Gemini 3.7 Flash gets a higher score, 95.5%, for $0.12 — 38,000× against o3's low-efficiency run, or 217× against its $26 high-efficiency one. So the annualised decline sits somewhere between roughly 25× and 560× a year depending on which December 2024 configuration you take as the baseline. All of those land inside the 9-fold-to-900-fold annual range Epoch AI found for the cost of reaching a fixed benchmark score, which our own token-economics piece cited three weeks ago. The index and the benchmark agree, which is worth more than either alone.

ARC-AGI-2 is running the same collapse one version behind:

• March 2025, launch day: o3-preview-low scored 4% at $200 a task; o1-pro about 1% at a similar price; the 2024 Kaggle winner 3% at $0.25; GPT-4.5 scored 0%.
• Late 2025: Claude Opus 4.5 thinking reached 37.6% at $2.20; Gemini 3 Pro with a third-party refinement layer hit 54% at $30.
• July–August 2026: Grok 4.5 52.6%, Grok 4.6 67.1% at $0.76, Gemini 3.7 Flash 84.6% at $0.25.
• September 2026: Claude Fable 5.1 90.0% at $4.49; GPT-6 Astra 95.0% at max reasoning.

Grok 4.6 alone tells the ARC-AGI-2 story: 16.8× the launch score for 1/263rd of the launch price, in seventeen months.

And note what the frontier is not. Fable 5.1 pays $4.49 a task for 90.0% where Gemini 3.7 Flash pays $0.25 for 84.6% — 18× the cost for 5.4 points. That is what a Pareto frontier looks like, and it is why ARC Prize plots cost on an axis rather than ranking a single column.

6. The current frontier, model by model

Every row below is an ARC Prize Verified semi-private result at the model's best published setting, as of early September 2026. ARC-AGI-3 figures are standard-harness.

Three things are visible here that a single-column ranking hides.

The cliff is universal and it steepens with each version. Astra loses 3.5 points from ARC-AGI-1 to ARC-AGI-2 and then 32 more to ARC-AGI-3. Grok 4.6 loses 20 points, then 65. Opus 5 loses 7, then 60. Nobody walks across all three.

More reasoning effort is not monotonically better. Grok 4.6 scores 87.5% on ARC-AGI-1 at medium and 87.0% at both high and extra-high. Grok 4.5 scores 87.2% at medium against 85.7% at high, and ARC Prize notes on that run that raising effort from medium to high "did not result in increased performance or cost." If you are buying reasoning by the dial, the top setting is not automatically the one to buy — a finding you can only see because cost and score are published together.

Where costs are published, they cluster oddly. Gemini 3.7 Flash and Grok 4.6 sit at $0.12–$0.76 a task. Anthropic's Fable 5.1 sits at $1.40–$4.49 for a few more points. ARC Prize publishes no per-task cost on the Claude Opus 5 or GPT-6 Astra ARC-AGI-1 and ARC-AGI-2 pages, so the two highest-scoring systems on this table have no price attached to those scores at all.

7. Open weights against closed: level on v1, a third behind on v2

Two open-weight models carry ARC Prize Verified scores, and they change the shape of the cost argument entirely.

• Kimi K3 (Moonshot AI, verified 31 July 2026) — 94.5% on ARC-AGI-1 at $0.77 a task and 60.4% on ARC-AGI-2 at $1.59. ARC Prize called it a new high score for open-weight models on both benchmarks.
• DeepSeek V4 Flash 0731 (verified 7 August 2026) — 89.0% on ARC-AGI-1 at $0.02 a task and 61.4% on ARC-AGI-2 at $0.04. It took the open-weight ARC-AGI-2 record a week later, by one point, at 1/40th of Kimi's price.

DeepSeek V4 Flash is the cheapest verified result on the board by a wide margin, and the comparison that makes it land is the one this whole piece is built on: it scores 89.0% on ARC-AGI-1 — higher than o3's famous 87.5% — for $0.02 a task against $4,560. That is 228,000×, on downloadable weights, twenty months later.

The gap to closed models is real, and it widens exactly with how unsaturated the version is:

Four points on ARC-AGI-1 (98.5% for GPT-6 Astra against 94.5% for Kimi K3). 33.6 points on ARC-AGI-2 (95.0% against DeepSeek V4 Flash's 61.4%). And on ARC-AGI-3, nothing at all.

That is the honest summary of the open-versus-closed race as ARC-AGI sees it. On the benchmark that is finished, open weights have essentially caught up — four points, at a fraction of the price. On the benchmark that is live, they are a third of the way behind. And on the agentic benchmark, no open-weight model has a verified score at all.

The price story runs the other way. DeepSeek V4 Flash gets 61.4% on ARC-AGI-2 for $0.04 where Gemini 3.7 Flash gets 84.6% for $0.25 — 6× cheaper for 23 points less, and both are far below the $4.49 Fable 5.1 charges for 90.0%. If you are buying fluid reasoning by the task rather than by the headline, the open tier is now a serious line on the shopping list, and it was not one a year ago.

8. The prize that pays for efficiency, not for score

ARC Prize put its money where the definition is. The rule that matters is not the score bar alone — it is the score bar inside a compute envelope.

The 2025 competition asked for 85% on the private ARC-AGI-2 set within Kaggle's efficiency limits: a notebook with no internet access, fixed hardware, and roughly $50 of compute per submission across 120 tasks — about $0.42 a task. It drew 1,455 teams and 15,154 submissions. The winner, NVIDIA's NVARC, scored 24.03% on the private set at $0.20 a task using synthetic data generation and test-time training on a 4-billion-parameter model, and took the $25,000 top-score prize. The grand prize went unclaimed.

Hold two numbers next to each other, because this is the trap:

Gemini 3.7 Flash is verified at 84.6% on ARC-AGI-2 at $0.25 a task. The prize asked for 85% at about $0.42. Four tenths of a point below the score bar, and 40% below the cost bar. It did not win, and could not have.

They are not measurements of the same thing. The verified run is on the semi-private set with a frontier datacentre behind it and a bill settled afterwards. The prize is scored on a private set nobody outside ARC Prize has seen, inside a sealed notebook with no internet. The best score anyone has achieved under those constraints is 24.03% — and nothing in the published figures lets you decompose the gap to 84.6% into how much is the constraint and how much is a better model.

For 2026 the structure changed. The ARC-AGI-2 track carries a $700,000 pool split into Progress Prizes ($275,000), a Grand Prize for the best solution writeup ($275,000), and a $150,000 Bonus Prize for the first solution scoring at least 85% on the private set — which ARC Prize says it intends to roll forward to 2027 if unmet. The ARC-AGI-3 track carries $850,000, including a $700,000 grand prize for the first agent to score 100%, with $37,500 milestone prizes at each of two 2026 deadlines. Everything must be open-sourced to be eligible. Total across ARC Prize 2026: over $2 million.

9. ARC-AGI-3: the harness became half the benchmark

ARC-AGI-3 is where the cost story stops being cheerful.

At release in March 2026 the official semi-private leaderboard read: Claude Opus 4.6 (Max) 0.50%, Gemini 3.1 Pro Preview 0.40%, GPT-5.4 (High) 0.20%, Grok-4.20 0.10%. Humans: 100%.

Running these evaluations is expensive enough that ARC Prize wrote the cost into the rules. It imposes an action budget of five times the human baseline explicitly because evaluating high-reasoning frontier models across the full set runs to "tens of thousands of dollars in API costs," and the leaderboard hides any system that required more than $10,000 to run. The published figures bear that out: Grok 4.6 scored 2.11% on ARC-AGI-3 for $5,612 in total run cost. Astra's standard-harness run cost $26,098 — well over the display filter.

Then there is the harness problem, which ARC Prize's own technical report states more sharply than any commentary has. Testing general harnesses built against three public environments, then running them on the full public set:

In a variant of environment TR87, Opus 4.6 scores 0.0% with no harness and 97.1% with the Duke harness, yet in environment BP35, Opus 4.6 scores 0.0% under both configurations.

Same model, same environment. Zero, or ninety-seven, depending on the wrapper — and the wrapper's advantage does not transfer to environments it was not built for. ARC Prize's conclusion is that specifically engineered harnesses are not a useful way to measure AGI progress, so the official leaderboard uses no harness at all: one system prompt for every model, no tools, a deliberately minimal interface.

10. GPT-6 Astra: 62.7% and 99.9% from the same model

The most recent datapoint, from the first week of September 2026, is the cleanest public illustration of everything above.

Plotted, the four rows say it in one look: three lines climb with the reasoning dial and one sits flat against the ceiling from the very first setting. The best standard-harness ARC-AGI-3 result was 62.7% at max reasoning, costing $26,098. The best provider-adapter result was 99.9% at high reasoning, costing $18,817 — higher score, and 28% cheaper. The adapter preserves opaque reasoning state between requests and compacts long conversations, so the model reuses its own prior work instead of rebuilding context every turn.

The number that travelled was 99.9%, frequently set against GPT-5.6 Sol's 7.8%. Those are not the same test: Sol's figure is standard-harness. Read on the standard harness, where every model faces identical conditions, Astra's 62.7% is still a very large record — the previous best was Claude Opus 5 at 30.16% — and it is the number a researcher should quote.

Treat provider-adapter and standard-harness numbers as different sports. They are not two observations of one quantity, and the adapter being both cheaper and better is precisely why the distinction will not resolve itself.

11. How much of the gain is hardware, and how much is software

A 15,200× fall in the price of the same score invites an obvious question almost nobody answers: what actually got cheaper? The ARC-AGI record is unusually good for this, because one of the three components is measured inside a single published table.

Step one — sampling, and it is the biggest single move. ARC Prize's December 2024 grid gives two runs of the same model, on the same day, on the same hardware, at the same token price:

170× fewer tokens, 175× less money, zero hardware involved. That is not a chip getting faster or a price list being cut. It is a decision to stop drawing a thousand candidate programs and keeping the best one. Test-time search was always a way of buying a score, and the first and largest part of the cost collapse was the industry learning it no longer had to.

Step two — everything between then and now. From o3's $26 to Grok 4.6's $0.30 is 87× over twenty months. This is where hardware enters, and its ceiling is worth being precise about.

NVIDIA's strongest public claim, from SemiAnalysis InferenceX benchmarks published in April 2026, is that a GB300 NVL72 rack serves inference at $0.123 per million tokens — 35× lower than Hopper, with 50× the tokens per watt. Grant that in full, even though it is a vendor-published peak configuration and none of these ARC runs are certified to have run on one. Hardware then accounts for at most 35× of the 87×, leaving roughly 2.5× for model efficiency, distillation, quantization, serving-stack gains and competitive pricing.

Put the chain on a log scale, the only fair way to apportion a multiplicative decline:

Fewer samples contributes 175×, or about 54% of the total. Hardware at NVIDIA's own ceiling contributes at most 35×, about 37%. Model efficiency, serving and pricing account for the remaining ~2.5×, about 9%.

Roughly two thirds of the collapse is software, and that is the generous reading for hardware — it assumes the entire Hopper-to-Blackwell-Ultra gain reached retail API prices within twenty months, which it plainly did not. NVIDIA's own figures make the point from the other side: on identical GB300 silicon, throughput on DeepSeek-R1 rose 2.7× in six months from serving software updates alone.

One caveat is load-bearing. This decomposes retail price per solved task, not physical cost. Every dollar here is what a buyer paid, which bundles margin, competitive positioning and possibly loss-leading. Nobody outside the providers can separate those, and this is the point where most analysis quietly assumes price equals cost. We are not going to.

12. What it means for the people selling the compute

If the price of a solved task falls 15,200× and most of that is software, the natural conclusion is that hardware demand is in trouble. The ARC-AGI record says the opposite, and shows the mechanism unusually clearly.

Demand did not follow the cheap task down. It followed the frontier task up. As ARC-AGI-1 collapsed to cents, the interesting problem moved to ARC-AGI-2 and then to ARC-AGI-3 — where a single standard-harness evaluation of GPT-6 Astra cost $26,098, Grok 4.6's cost $5,612 for 2.11%, and ARC Prize caps agent actions at five times the human baseline specifically because full frontier runs reach "tens of thousands of dollars." The typical frontier evaluation is more expensive today than a year ago, on a benchmark family whose per-task price fell four orders of magnitude.

That is Jevons made visible on one dataset, and it is the bull case for NVIDIA in miniature. Our NVIDIA model does not price tokens at all: the data-centre compute line is a unit driver — rack systems shipped × price per rack, with price drifting down about a percent a quarter. Cheaper inference reaches that model only by changing how many racks get bought. Same shape on the other side of the trade: our Alphabet model carries Gemini inside Google Cloud on sequential revenue growth and capex at 140% of segment revenue, with no token line either, and Alphabet's shares are underwritten on Cloud converting a backlog rather than on what a token costs. Neither model has a place to put a cost-per-task figure, and that is the correct design — the quantity that matters to both is volume, not unit price.

Two structural details in the ARC-AGI data are worth watching for both companies:

• The cheapest verified frontier point is not running on merchant silicon. Gemini 3.7 Flash — 95.5% on ARC-AGI-1 for $0.12 and 84.6% on ARC-AGI-2 for $0.25 — is served by Google largely on its own TPUs. The best price-performance on the board belongs to a vertically integrated stack. That is an argument about who captures the margin, not about whether the compute gets bought.
• The cheapest point overall is open weights. DeepSeek V4 Flash at $0.02 a task is downloadable. Weights anyone can serve put a floor under price competition that no vendor controls, and they push inference toward whoever has the cheapest capacity rather than the best model.

Neither changes a fair value, and we are not going to pretend otherwise. Both change the question worth asking about the next four quarters: not "is intelligence getting cheaper" — it plainly is, by a factor nobody forecast — but "does the cheap task get replaced by a more expensive one fast enough, and who owns the silicon it runs on."

13. What the cost collapse does and does not mean

Ten readings, in descending order of confidence.

1. The efficiency definition was the right call. Every interesting fact here is visible only because ARC Prize made cost a reported metric in December 2024. A leaderboard ranked on score alone would show o3's 87.5% and Grok 4.6's 87.5% as no change at all, instead of a 15,200× repricing.
2. Test-time compute was real but shallow. Sampling 1,024 programs and picking a winner looks like abstraction and is not. ARC-AGI-2 was built to make that search combinatorially worse, and the 2–3× drop across every paradigm says it worked.
3. Falling token prices are not the same as intelligence getting cheaper. On the decomposition in section 11, roughly two thirds of the collapse is software and at most a third is hardware, even granting NVIDIA its own best claim in full. The single largest component — 175× — was a sampling decision taken on one day in December 2024. These are different economic events with different implications for who captures the value.
4. Cheap models now beat 2024's expensive ones outright. Gemini 3.7 Flash at $0.12 outscores the o3 configuration that cost $4,560. On ARC-AGI-1 the capability is commoditised.
5. The hard versions did not get cheap. ARC-AGI-2 is mid-collapse. ARC-AGI-3 has barely started: $5,612 for 2.11%, $26,098 for 62.7%, and a $10,000 line the leaderboard uses to hide runs above it.
6. Humans remain the efficiency champion where it counts. ARC Prize books its own panel at $17 a task — $115–$150 for a 90-minute session plus $5 per correct task — and reckons the true floor is nearer $2–$5. Against an agent scoring 30% on ARC-AGI-3 for tens of thousands of dollars, that gap is the entire remaining problem.
7. Saturating ARC-AGI-1 is not AGI, and ARC Prize says so every time. ARC-AGI-1 is now a regression test, ARC-AGI-2 a reasoning-efficiency test, ARC-AGI-3 an agent-learning test. Labs putting all three on model cards is real progress in reporting; treating any single one as an AGI verdict is not.
8. Open weights are level where the problem is solved and behind where it is not. Four points from the closed frontier on ARC-AGI-1, 33.6 points on ARC-AGI-2, and absent entirely on ARC-AGI-3. Openness follows saturation with a lag, and the lag is the difficulty of the version.
9. Cheap compute did not reduce compute spending. The per-task price fell four orders of magnitude while the cost of a frontier evaluation rose into five figures. The workload migrated rather than shrinking, which is the mechanism the Jevons argument actually needs.
10. The treadmill is the mechanism, not a complaint. A benchmark rebuilt yearly to be unsaturated on release is the only reason the cost curve stays measurable. If ARC-AGI-1 were still the frontier, the interesting quantity would have stopped moving in 2025.

14. What would change the conclusion

• A private-set entry above 24.03% inside the ~$0.42 envelope. The number ARC Prize designed, and the only one that says a method was invented rather than bought. The $150,000 bonus for 85% rolls to 2027 if nobody reaches it.
• ARC-AGI-3 evaluation costs falling by an order of magnitude. $5,612 for 2.11% and $26,098 for 62.7% is not a cost curve; it is the position ARC-AGI-1 was in during 2024.
• Standard-harness ARC-AGI-3 scores continuing to climb. 0.50% at launch in March 2026 to 62.7% in September is the fastest movement in the benchmark's history, on the version meant to last. If it holds, ARC-AGI-4 arrives in early 2027 into a field that has already learned to play.
• Per-task costs published for the top of the table. The two highest ARC-AGI-1 and ARC-AGI-2 scores currently carry no price. Efficiency is the definition; a score without a cost is half a result.
• An open-weight model with a verified ARC-AGI-3 score. There is none today. The first one would say the agentic frontier is diffusing rather than concentrating, and would put price pressure on the only tier that currently has none.
• Independent measurement of realised inference cost per token. The hardware share in section 11 rests on a vendor's own peak-configuration benchmark. A neutral series of what providers actually pay would settle the hardware-versus-software split rather than bounding it.
• Any evidence about what inference actually costs a provider to serve. Every figure here is a retail price. The gross margin on frontier inference is the most load-bearing undisclosed number in AI economics, and no benchmark can supply it.

15. Primary sources worth reading

• ARC Prize and its leaderboard, where the $10,000 filter and harness notes are stated before any row.
• ARC Prize Verified — what verification actually involves, and the testing policy.
• On the Measure of Intelligence (Chollet, 2019) — the skill-acquisition-efficiency definition everything rests on.
• ARC-AGI-2 technical paper — human panel methodology, the anti-brute-force redesign, launch-day scores with costs.
• ARC Prize 2024 technical report — the 33% → 55.5% jump, test-time training and program synthesis.
• ARC-AGI-3 technical report — RHAE scoring, the no-harness policy, and the 0.0%-versus-97.1% result.
• The ARC of Progress towards AGI (2026) — the living survey of 82 approaches and the 2–3× cross-version cliff.
• o3 breakthrough post (December 2024) — the cost table in section 5.
• ARC Prize 2025 results and the 2026 competitions.
• NVIDIA AI inference — the GB300 NVL72 $0.123-per-million-tokens claim and the 35×-versus-Hopper figure used in section 11, sourced to SemiAnalysis InferenceX benchmarks.

Every benchmark score, cost per task and prize condition here is ARC Prize's own published figure, from its leaderboard, per-model results pages, competition rules, blog and the ARC-AGI-2 and ARC-AGI-3 technical reports; ARC Prize builds the tasks, runs the harness and computes the costs itself. Its December 2024 o3 figures were restated after publication and are approximate. Scores from different splits — public, semi-private, private — and different harnesses are not strictly comparable, which is much of the argument above; the 87.5% pair in the headline is the one comparison holding split, benchmark and score constant. The hardware figures are NVIDIA's own, sourced to SemiAnalysis InferenceX benchmarks published April 2026, and are vendor-published peak configurations rather than measurements of these runs — section 11 uses them as a ceiling for that reason. The Epoch AI 9-fold-to-900-fold range is carried from our earlier token piece. Every ratio, annualised decline and share of the decomposition is ours, computed on ARC Prize's and NVIDIA's published points. All prices quoted are retail prices paid by a buyer, not costs to serve, which nobody outside the providers can observe.