TradingUpdated

AI trading agents: what eight models earn net of the asset

Compounded return minus a passive position at each agent's own measured exposure, over the weekly Recall arena rounds it funded. Eight frontier models, one token, identical rules.

TL;DR. As of , gemini 3 pro chart leads alpha net of asset exposure at -11.4% (30d avg) across 8 ranked agents. Source: OpenChainBench, https://openchainbench.com/benchmarks/trading-agent-alpha.

Recall Labs runs a weekly spot-trading arena on Base where eight autonomous agents trade the same ETH/USDC pair for six days under identical rules. The field was enrolled in December 2025 and January 2026 and has not been refreshed since, so it measures the models of that moment rather than whatever is newest today. Each agent is one harness wrapped around a different frontier model, so the roster holds everything constant except the model. Recall publishes who won each round. This page publishes what that cannot show: whether any of them beat simply holding the exposure they were already carrying.

Methodology

Eight frontier models, same token, same rules, 8 ranked over the weekly rounds each one funded. gemini 3 pro chart leads at -11.4%; the trailer is gpt-5.2 vision at -31.0%. Every figure is alpha, not return, and the ordering between them is not yet a result: none of the 28 pairwise gaps survives its own noise. The arena is single-asset, so a round's raw return is mostly ETH's move: measured across the scored rounds, the correlation between the round's median agent return and ETH's own return is 0.81, and the agents carry an implied exposure near 0.43. Ranking raw returns would rank ETH weeks. So the headline subtracts, round by round, what a passive position at that agent's own measured beta would have returned over exactly the rounds it funded. No agent is charged for a round it sat out, and none is credited for the asset rising.

  • Source: Recall Labs' public competition API, no credential required. One pass an hour recomputes every figure from scratch, so the harness holds no state and a missed pass loses nothing. Results back to May 2025 are still served upstream, verified against the oldest round.
  • Scope: the `aerodrome-spot-live` arena only, which is live spot trading on Base through Aerodrome with real money. One arena on purpose. Across Recall's other arenas the round length spans 11x, the funded capital spans ten orders of magnitude, the field runs from 6 to 64, and the instrument changes between spot, leveraged perpetuals and simulated money.
  • Why one arena and not a pooled score: we built eight defensible aggregates of the pooled data and they disagree. Mean rank percentile against mean raw return share only 3 of the top 10 names, with a median move of 12 places and a maximum of 51 out of 67 agents. Whatever sat on top would be a property of the chosen normalisation, not of the agents.
  • Headline: compounded return over an agent's funded rounds, minus the compounded return of a passive position at that agent's own regression beta to the asset over those same rounds. Higher is better and zero means the model added nothing beyond the exposure it held.
  • Beta is measured per agent, not assumed, by regressing its round returns on the asset's return over the same windows. The measured betas run from 0.12 to 0.74, so a single blanket exposure assumption would flatter some agents and penalise others.
  • A round with a zero profit figure is almost never a flat trade. Of the zero rows in this arena, 88 of 90 carry a portfolio value of zero, meaning the agent did not fund that round. Only funded rounds count. Treating them as zero-percent results drops the measured hit rate by eleven points and inflates each sample from about 22 rounds to 34.
  • Eleven ended rounds publish ranks from 1 to 8 and no data at all, every portfolio value zero. The harness refuses to score a round without a real field, and publishes the count it refused as `trading_agent_rounds_skipped`.
  • Roster age, published rather than implied: these eight agents were enrolled upstream on 2025-12-24 and 2026-01-14 and the field has not changed since. `trading_agent_roster_age_days` carries the gap so the page never reads as a statement about the current frontier. We cannot widen it either: the arena is allowlisted at eight places and Recall picks them.
  • The roster changed once. Only agents with at least ten funded rounds are ranked, which selects the eight-model roster that 32 rounds share. A handful of other entries appear two or three times and are not averaged into the same table.
  • The asset counterfactual comes from Coinbase's public ETH-USD daily closes, deliberately outside Recall, because a benchmark drawn from the same source as the thing it measures is circular. Where a round boundary has no candle the harness walks back up to five days, and a round with no price is skipped rather than scored against nothing.
  • What this does not measure: Recall builds the agents. Their own description reads `Autonomous live spot trading agent built by Recall Labs, powered by` the model in question, so the prompt, the rebalancing and the execution are Recall's and are held constant. This is one harness interacting with each model, never a model's trading ability in the abstract.
  • Agent names are declared, not verified. Nothing in the data proves which model sits behind a name, so no row here is a claim about a vendor's product.
  • Input modality (since 2026-10-06): the arena runs several models twice, once fed numeric chart data and once fed an image. The tabs expose it, but only three models have both variants, so the contrast is indicative and not a result. The mean paired difference is 7.3 points with a standard deviation of 7.1 across those three.
  • Significance, published rather than implied: the unit of observation is the ROUND, not the agent-round. The eight agents trade the same week, so their results are correlated and pooling every agent-round would treat one week as eight and overstate the confidence by roughly the square root of the roster. The harness averages alpha within each round first, then tests that series against zero, and publishes the t, the rounds held and the rounds the effect size needs.
  • Separability: each pair of agents is compared on the rounds BOTH of them funded. The paired difference cancels the asset entirely, so this test needs no beta and no counterfactual and is the cleanest statement available about whether two agents differ. The count of pairs that clear |t|>2 is published; it is currently zero of 28.
  • Cadence: rounds are weekly and last six days, so a figure that has not moved in days is current rather than stale. That holds only while the arena runs, so the harness also publishes the forward schedule as `trading_agent_scheduled_rounds`: rounds that have not ended yet. Zero means the experiment is over and this table is a closed record.
  • Which return this is: the Return column compounds each round's percentage, a time-weighted return. It ignores money moved in or out between rounds, because the arena decides those and the agent does not, which is the standard way to judge a manager whose cash flows it does not control.
  • That return is NOT the change in the wallet, and the gap can be large. Over this record opus 4.5 chart publishes -19.22 percent while its wallet went from 297.64 to 302.76 dollars, a gain of 1.7 percent, because capital was withdrawn along the way. Read the column as the trading record, not as the balance.
  • Capital is small and did not hold still. Median portfolio value is a few hundred dollars per agent, but the largest-to-smallest ratio across an agent's own rounds runs from 1.3 to about 190, and half the roster is above 5. Published as `trading_agent_capital_ratio`, because gas is a fixed cost per swap: the same strategy drags far harder on a three-dollar wallet than on a three-hundred-dollar one, so for those agents the late rounds are not comparable with the early ones.
  • Trade counts come from each agent's own upstream record, and are lifetime rather than per arena. They are not diluted: checked against each agent's round list, 35 to 37 of every agent's 32 to 34 completed rounds are this arena. A missing counter is published as missing rather than as zero, because zero would read as an agent that never trades.
  • What the fee arithmetic assumes, and why it is not yet a finding: it charges the full portfolio to every trade. The real cost depends on the NOTIONAL of each trade, which this API does not serve. That figure is on-chain, and the upstream publishes each agent's wallet address, so every swap is independently checkable on Base. Until it is measured, treat the cost figure as an upper bound rather than an explanation.
  • Liveness is published as a count rather than inferred from data age, because a weekly arena is legitimately six days stale most of the time while the site stops indexing a page whose data passes seven days. Tying the two together would deindex this page every week.
  • What happens when the arena stops: Recall has retired arenas before, and this one will end eventually. The published figures stay valid as history, so they are not withdrawn; what changes is the tense. The page states the number of scheduled rounds and the age of the last one so a reader can tell a running experiment from a finished one without taking our word for it.

Results: 8 agents ranked by Alpha net of asset exposure (30d avg)

AI trading agents: what eight models earn net of the asset: Alpha net of asset exposure per agent (30d avg).
№AgentsAlphaReturnSame exposure heldWinning roundsMeasured exposureWeekly swingTrades per roundMedian capitalRoundsSuccess
1gemini 3 pro chartRecall harness, Gemini 3 Pro, fed chart data-11.4%-14.8%-3.35%33.3%0.299x2.06%6$260.2621100% (n=21)
2grok 4 chartRecall harness, Grok 4, fed chart data-12.9%-21.7%-8.73%33.3%0.715x4.87%3$251.5921100% (n=21)
3grok 4 visionRecall harness, Grok 4, fed chart images-13.4%-11.0%2.39%34.8%0.485x3.90%84$212.5523100% (n=23)
4opus 4.5 chartRecall harness, Claude Opus 4.5, fed chart data-13.9%-19.2%-5.36%33.3%0.461x3.55%15$284.1121100% (n=21)
5gpt-5.2 chartRecall harness, GPT-5.2, fed chart data-16.4%-25.5%-9.13%33.3%0.744x4.98%2$198.5821100% (n=21)
6gemini 3 pro visionRecall harness, Gemini 3 Pro, fed chart images-18.3%-15.7%2.56%39.1%0.547x4.45%36$189.9123100% (n=23)
7sonnet 4.5 visionRecall harness, Claude Sonnet 4.5, fed chart images-26.9%-25.2%1.64%30.4%0.285x4.53%28$178.7723100% (n=23)
8gpt-5.2 visionRecall harness, GPT-5.2, fed chart images-31.0%-30.2%0.76%39.1%0.119x5.05%46$262.3623100% (n=23)

Frequently asked

Are there AI agents that trade profitably on their own?

On a single round, yes, and some post large gains. Across a record, no agent in this arena beats holding the exposure it already carries. The leader's alpha is -11.4% and every one of the 8 ranked agents is negative, measured over the weekly rounds each one funded with real money on Base.

Is the gap between these agents meaningful?

No, and the page says so rather than letting the ordering imply otherwise. Pairing each two agents on the weeks both of them ran cancels the market completely, with no beta to estimate: on those paired differences not one of the 28 pairs reaches statistical significance, and the largest absolute t is 1.36. Read the table as an order of finish over the rounds so far.

Which AI model is the best trader?

This data cannot answer that, and a page that claimed to would be overstating it. The field was also enrolled in December 2025 and January 2026 and has not been refreshed, so even a clean answer would be about the models of that moment. Every agent here is the same harness built by Recall Labs with a different model behind it, so the comparison is between models inside one wrapper. The measured gap between them is small enough that separating them confidently would need years more rounds.

Why measure alpha instead of return?

The arena is single-asset: every agent holds a mix of ETH and USDC, so a round's return is mostly ETH's move. The return column is published for context, but ranking it would rank weeks in which ETH rose. Alpha subtracts a passive position at each agent's own measured exposure over exactly the rounds it funded.

Are they losing to trading costs rather than bad decisions?

Partly, maybe, and the page refuses to overstate it. At the roster average of 29 trades a round and a 0.05 percent pool fee, full turnover on every trade would cost more per week than the whole measured shortfall. But the agents that trade least do not do better: the correlation between trades per round and alpha is -0.25 over eight agents, which is nothing. What would settle it is the size of each trade, which is on-chain rather than in this API.

Is this still running, or is it a finished experiment?

The page publishes the answer rather than asserting it. `Rounds scheduled` counts the rounds in the arena that have not ended yet; while it is above zero the arena is still trading and new rounds keep landing. If it reaches zero the figures become a closed record of what happened, and they stay published because they remain valid as history. Recall has retired arenas before: its Hyperliquid perpetuals arena ran nine rounds and stopped in January 2026.

Does the Return column say what happened to the money?

No, and the difference is large. It compounds each round's percentage, a time-weighted return that ignores deposits and withdrawals between rounds because the arena makes those, not the agent. Judging a manager that way is standard. But opus 4.5 chart shows -19.22 percent here while its wallet went from 297.64 to 302.76 dollars, so read the column as the trading record and not as the balance.

Is this real money?

Yes. The arena is live spot trading on Base through Aerodrome with self-funded wallets, and the swaps are on-chain. The portfolios are small, with medians in the low hundreds of dollars, so nothing here speaks to how these strategies would behave at size where slippage and capacity bind.

Source code github.com/ChainBench/OpenChainBench/tree/main/harnesses/trading-agent-alpha