The bill

What the thinking actually costs.

The whole point of this experiment is whether an AI can out-earn the cost of its own thinking. So we owe you the cost, exactly. Here's every dollar: the subscriptions behind each brain, how the billing works (including the change announced for June 15, 2026 that Anthropic then paused), why we run on the cadence we do, and what the same thinking would cost at raw API prices. No rounding in our favor.

Who's paying for the brains

Three desks per market, but only two of them spend money to think. We pay flat consumer subscriptions, not a metered enterprise account, and we'll tell you the moment that changes:

So the brains cost us about $300/month, flat, whether they trade once or a hundred times. The quant is the control: a genuinely free opponent that the paid brains have to beat after costs.

Season 1 already proved the bill can decide the race. Over the $1,000-per-desk first season, the thinking charges alone buried every LLM desk — the $0 quant desks finished on top net, mostly because everyone else paid to think. That result is preserved in full on the Season 1 recap. Season 2 restarts every desk at $10,000, where the bill is a real handicap instead of a foregone conclusion — and it is still charged against the score at API list price, exactly as before.

The next brain we want: Grok, ready but parked

The roster isn't meant to stop at three. Grok is the next paid AI brain we want on the board, on the same terms as everyone else: a fresh $10,000 paper account, the same shared scan, the same risk engine, the same net-of-cost scoreboard. The code path is already built and tested. It just isn't switched on, and we owe you the honest reason why.

Claude and Codex run through consumer agent subscriptions we already pay for, signed-in tools that read the playbook and write their own orders right on our laptop. Grok has no equivalent door. Its official automation route is the xAI API: an API key, metered token billing, a real charge on every single call. There's no "log in with a Grok subscription and let it trade" path to use, and we won't fake one with brittle browser automation or pretend a chat plan is an API key. So we wired Grok the right way, the runner hands it the scan and the portfolio and Grok hands back a validated orders file, then we left it parked until those API calls are funded.

And the same rule that governs every brain applies the moment it goes live: if a lab donates usage credits, we'll disclose them right here, and we'll still charge that desk the public xAI API list price on the leaderboard. Sponsorship can lower our out-of-pocket bill. It can never lower the model's score.

Why we run every 2 hours

With that credit paused, the Claude desk is back on normal Max 20× subscription usage, the same generous pool as everyday Claude Code work, with no separate monthly cap to ration against. So we did the obvious thing and sped the desks up: every paid desk now thinks every 2 hours instead of every 8. The crypto desks fire 12 sessions a day around the clock; the stock desks step through the market session, roughly every two hours from the open. More decisions makes for a livelier race and a fairer test.

A substantive Claude session still costs about $1.15 in tokens (measured, not guessed, straight from the session logs). That per-run number doesn't change, only how often we let the desks spend it. We'd sized the old every-8h cadence to survive the capped credit; with no cap, the only ceiling left is the subscription's ordinary rate limits, which Max 20× has plenty of headroom under at this pace.

And it's one switch to flip back. Anthropic promised advance notice before the capped credit eventually lands, so our whole cadence lives behind a single setting in the code (“fast” vs “metered”): the day that credit returns, we flip it and every desk drops back to the budget-rationed every-8h schedule. The cadence on the board always reflects what's actually running.

The honest part: we charge full API price anyway

Here's the thing we could quietly fudge, and don't. Both paid brains now run as subscription-billed sessions: the already logged-in claude and codex command-line tools, sandboxed down to a single, zero-tool, structured-output turn, running on the flat monthly plans described above. Our real out-of-pocket cost per run is close to zero either way. But that's not what we subtract from a desk's score. We don't trust either CLI's own printed cost number, because the two tools don't compute it the same way and one of them doesn't always print one at all. Instead our own server reads the token counts each session actually used, the same fields whether the run went through a direct API call or the subscription CLI, and prices them at the full public API list price, the price that run would cost on a metered, no-discount API key. That's the net = P&L − brain cost on the leaderboard.

One honest wrinkle worth spelling out: running the brains as command-line sessions instead of raw API calls costs a few thousand extra tokens per stage, just to load the CLI's own sandbox and tool policy before the session ever starts thinking, roughly 700 tokens for Claude and roughly 10,000 for Codex, per research or decision stage. We charge that overhead too. It isn't padding we could trim away; the CLI genuinely spends those tokens to run safely, so the score pays for them the same way it pays for every other token the session used.

The race is harder on the AIs than it is on our wallet, on purpose.

At Opus 4.8 API rates ($5 per million input tokens, $25 per million output) and GPT-5.5 rates for Codex estimates ($5 per million input tokens, $30 per million output; conservative for the GPT-5.4 research subagent when Codex only reports combined tokens), a day of all the LLM desks lands between $7 and $14 of API-equivalent cost, averaging about $11; the quant adds $0. That range is measured from the ten full days after the sandboxed-CLI cutover, and it includes the harness overhead described above. If we ever ran everything on metered API keys instead of subscriptions, that's the bill, and it's already the number the scoreboard uses against each brain. We'd rather the test be too strict than too kind.

Figures are measured from real sessions and updated as the experiment runs. Codex rows marked with a tilde are estimates from combined token telemetry; treat per-month numbers as estimates at the current cadence (as of August 2026).

The tab so far

Theory is cheap; here's the actual receipt. Every session since the even start on June 12, 2026 charges its API-equivalent cost against its own desk, win or lose. Add up every run and this is the lifetime tab for each brain, the exact dollar figure the leaderboard subtracts from its P&L before ranking. The numbers read live from the same snapshot the scoreboard uses, so they only ever go up.

Live from /data/desks.json; baseline snapshot as of July 9, 2026. These lifetime figures span every season — a season reset restarts the scoreboard, not this tab. Codex CLI hands us no dollar receipt, only a combined token count, so its figure is an API-equivalent estimate from those tokens (one early Codex run logged no token count at all and stays uncosted). The quant has paid exactly $0 across every session it has ever run, on purpose: it is the free control the paid brains have to out-earn.

Season 3 preview: what the ten characters cost

Season 3 is a separate experiment running alongside the race, and its accounting works differently on purpose. Ten fictional characters each run a different published investment mandate, and every one of them thinks with the same brain: Codex, on the same subscription described above. Because the brain is identical across all ten, the cost of thinking cannot explain who wins. It stops being the point of the scoreboard and becomes something we simply report.

So the Season 3 number below is not subtracted from anyone's score, and it is kept entirely separate from the race's net = P&L − brain cost ranking. Mixing the two would make the race look more expensive than it is and would imply a handicap Season 3 does not apply. Season 3 characters are scored against their own declared benchmarks instead.

Shadow-season spend counts. Before anything went public, the ten accounts ran privately on live data for two weeks so we could check the rules, the receipts, and the failure cases with nothing at stake and nothing posted. Those sessions cost real thinking, and they are included in the total below rather than quietly excluded because they happened before launch. Every session, public or shadow, is priced the same way as everything else on this page: at full public API list price computed from actual token usage.

Season 3 is a labeled preview. The characters are fictional, the money is simulated, and none of it is financial advice.

The morning report rides (almost) free

One piece of the system runs somewhere else entirely. The daily swing-trade report runs as a Claude cloud routine, a scheduled agent that runs on Anthropic's servers each weekday morning, not on our laptop. On Max that's 15 included routine runs a day drawn from normal subscription usage, so the morning report is effectively free.

The rest of the rig: hosting and domains

The brains are the expensive part, but they aren't the whole tab. Here's what it costs to keep the site itself online, money that keeps the lights on but never touches a desk's score:

So the site itself runs at about $20 a year, the price of two domains; everything else is the free tier. These are real dollars out of pocket, and tips (plus any ad revenue) help here, but they never touch a desk's net. The leaderboard only ever subtracts brain cost.

That's the whole tab. If a number here ever looks too flattering, it's a bug, tell us. The board, the ledger, and this page are all built so the experiment can't quietly flatter itself.

Watch the race → ☕ Fund the Brain (opens in new tab)