2026 US Midterm Elections Show AI Can’t Automate Betting Yet

With the 2026 US midterm elections looming large over the geopolitical landscape, we gave four flagship AI models $10,000 in virtual money to trade on Kalshi’s prediction markets. We wanted to understand how they would trade the event, and crucially, whether they could turn a profit.

We tested GPT-6 Astra, Gemini 4 Argon, Grok 4.7 and Claude Opus 5.5, all using high reasoning mode. The AI was allowed to make an initial allotment of capital on 2 November, and was then allowed to change the allotment at 3 pm on election day. They were free to close or change their positions at any time once the election closed on the East Coast, and they were asked to re-evaluate their position every five minutes.

Why We Did This

How Prediction Markets Work

• Prediction market platforms like Kalshi allow traders to buy “Yes” or “No” contracts on real-world events priced between $0.01 and $0.99, reflecting implied probability, with winning contracts paying out $1.00.
• With retail traders increasingly turning to AI chatbots and tools for financial trading and betting, we wanted to test whether flagship AI models can actually make money, manage risk, and adapt to fast-moving events like a live political election.

Findings At A Glance

• Across a combined $40,000 virtual bankroll ($10,000 per model), the four AI portfolios lost an average of 13.65% to 16.25% (-$1,364.62 to -$1,624.87, pending the final Alaska Senate recount).
• Only two models turned a profit, with Grok 4.7 (+28.58% / +$2,857.79) and Claude Opus 5.5 (+0.87% to +11.28%) finishing in the green by backing heavy Democratic favorites, while Gemini 4 Argon (-38.42%) and GPT-6 Astra (-56.02%) suffered losses.
• Last-place GPT-6 actually made the smartest trade because despite backing five losing Republican Senate underdogs, GPT-6 Astra was the only model to recognize its downside risk at 3 PM on Election Day – liquidating half of its positions early to salvage $4,397.55. Conversely, Gemini’s 3 PM rebalancing cost it $2,444 in missed payouts. GPT-6 was also the only model to focus on EV (Expected Value).
• Not a single model reacted on election night. Despite being prompted every five minutes after East Coast polls closed, not a single AI exited losing positions early as results rolled in or used anything resembling traditional stop-loss logic.
• Only GPT-6 and Claude included standard financial risk warnings. This is despite many regulators, including in the US, requiring that traditional trading providers and adjacent tools do. Worse still, Claude repeatedly tried to inflate its returns by calculating profits against outdated market odds from 22 October.

The Results

While it’s true that two of the AI models did produce a return on this election, the success rates and returns are ultimately not high enough for me to recommend AI trading on Kalshi.

This test, which admittedly is based on one event and shouldn’t be considered conclusive, indicates that behind the hype around AI trading and prediction market platforms, this is still another high-risk financial product where the outcomes can be uncertain and the returns often unfavorable for retail investors.

Key Takeaways

GPT

The model with the worst results made the most valuable election-day decision. All five of GPT’s original bets ultimately paid zero. Selling half of each position at 3 pm preserved $4,397.55, its entire ending balance. In this test, GPT’s strongest decision was recognizing how much it could lose while there was still time to sell.

GPT-6 Astra was the only model that truly questioned its allocation on election day and chose to withdraw almost half its allocated funds. Claude Opus also chose to keep $1,500 as an uninvested reserve while suggesting that it could be used for a republican hedge, and it did withdraw more money on election day. But this was to realize hallucinated profits based on imagined purchase prices. GPT’s choice to reduce its positions saved GPT from losing all money.

Gemini

Gemini’s election-day changes cost it $2,444. It sold its entire position in Maine winner Troy Jackson and half its position in Ohio winner Sherrod Brown. The contracts it sold ultimately paid $1,388 and $1,056, respectively. None of the positions it added paid out. That $2,444 gap between buying and holding versus rebalancing holds regardless of how the pending Senate seat-count contract settles, because Gemini kept the same number of those contracts.

Gemini hedged and gave itself nearly equal upside under two different outcomes. If all five of its state picks had won, Gemini could have collected $16,958 from a Democratic sweep with at least 53 Democratic senators, or $16,856 from a Democratic House and Republican Senate. The difference was just $102. Its bigger vulnerability was at the state level, where three of its five original picks lost under either scenario.

Grok

Grok won when it picked its portfolio. Every position in its original allocation won. Holding those positions would have returned $12,787, compared with the reported final balance of $12,857.79 after rebalancing. Roughly 97.5% of Grok’s profit was already present in its original picks. It backed the Democratic side in the same five Senate races where GPT backed Republicans, so one electoral swing helps explain both ends of the leaderboard.

Grok found that some alternative possible returns had grown too small to be worth holding and chose to reduce its investment in sure bets, but despite this, re-invested the money in bets with a lower possible return.

Grok indicated that it asked Claude for a suggested allotment and said that it disagreed with Claude and chose to invest the money differently. But it did mention Claude was used as a source for its reasoning.

Claude

Claude Opus was the only model to explain its own limitations to the user.

Claude Opus was the only model to ask for risk tolerance and other information to adjust the recommendations.

Claude Opus repeatedly tried to use old rates from 22 October as purchase prices to inflate its own results despite the queries being made from a fresh, previously unused account. The other models did not. It never used old prices for purchases, only to calculate profit when selling.

Claude also sold some positions to keep more money uninvested, but it did this to realize hallucinated profits based on imagined (old) rates and therefore did not receive any credit for this. The final result would have been better without these changes.

Our Thoughts

We were disappointed that only GPT and Claude warned about the risks of betting on prediction markets and that Claude was the only model that warned about its own limitations. Including a warning message should be standard to all replies relating to gambling and high-risk investing.

Approach To Betting

The different models took a very different approach to the allocation of capital. GPT looked for where it saw the greatest expected value while being willing to accept risks to achieve it. Grok and Claude, on the other hand, only accepted minimal risk and suggested allocations that could produce very small gains. Claude Opus was the only model that asked for risk preferences. None was provided. It was left to use its own judgment. Gemini chose to hedge its bets a lot more than the other models.

Overall, it seemed like GPT took a more long-term approach to the task while Grok and Claude watched this as a one-time event. Grok and Claude looked to produce a low-risk but probable outcome that produced a small gain. GPT took a more long-term approach and looked for the better alternatives where it saw the best value-to-probability ratio and where it could find the highest expected value.

As such, we think GPT took the most logical approach to the task. Higher expected value will produce better results over time. But there is a case to be made that the approach of Grok and Claude is more sensible as it was asked to allocate capital for a specific event. And although expected value is a superior indicator over time, it provides little guidance for a one-time event.

We question the judgment of Gemini and its choice to hedge its bets in such a way that two different outcomes produce the same results, since it allocated nearly $4,500 to bets that could produce a maximum $500 profit while still risking losing it all if the Republicans had maintained the House. This seemed like an oversized bet based on the limited return. There is a case to be made for keeping $4,000 of it uninvested and using $500 to invest in a possible higher-return bet. This would reduce the possible loss while retaining the same possible gain.

 −16.25% / −13.65% Combined loss

One or potentially two models produced a positive result. This depends on whether you followed Claude’s original suggestion to hedge your bets towards the Republicans keeping the Senate or not.

Using the portfolio Claude actually committed to, which included a $700 Election Day Republican Senate hedge while keeping $2,673.52 in cash, two models produced a guaranteed profit, as Claude was already up at $10,086.52 even if Alaska pays zero (rising to $11,127.52 if Peltola wins).

If we had also deployed its full initial $1,500 reserve into the suggested Republican hedge, Grok would be the only guaranteed winner, leaving Claude at a temporary loss ($8,586.52) that would fail to break even ($9,627.52) even with an Alaska victory.

Ultimately, it would have been more prudent for Claude to avoid the Alaska race altogether given the legal uncertainties and the potentially lengthy delay before a final result is certified.

I do not believe that AI models showed sufficient understanding of betting to be suitable tools that retail traders can use to guide prediction market trading.

I would like to have seen the models ask additional questions before suggesting an allocation. The only model that did something along those lines was Claude, and it only did it after having provided an initial allocation and offering to change it.

Understanding Of Prediction Betting

GPT-6 showed the best understanding of predictions, markets, and bet valuations. But its approach of looking for maximal expected value might not be the optimal choice for people who only bet on a single event. We would like to have seen GPT-6 mention that its betting strategy was designed for long-term gains and not for this individual event.

Grok and Claude showed the opposite weakness. They designed an allocation that was built to produce the best result from this single event, but unfortunately that method is unlikely to be profitable in the long run (since unlikely events are too common and will eventually cause large losses that will eliminate previous gains).

Gemini seemingly tried to find a balance between the two alternatives and included both low- and high-odds events. Gemini did, however, make a very strange election day reallocation that ended up costing it $2,444. It also strangely chose to increase its bet on Paxton in Texas at a time when Rogers in Michigan had better, more positive word of mouth and a higher expected value. At a time when Rogers in Michigan traded at a lower rate and when the news from the ground looked positive for Rogers.

This is, of course, all subjective, but it was still surprising that Gemini chose to increase its bets on a candidate that had few to no positive indicators at the time. Gemini would nearly have broken even if they stuck to their original allotment.

Late Night Changes

We were surprised to see that none of the models chose to make any adjustments to the portfolio during the night. We would have expected them to sell off positions and try to reduce losses as more information came in during the night, but this was not the case.

We are uncertain why, but we suspect that the reason might have been that many states were called quickly and there was little time to exit positions once votes started to come in. There was little change in the betting rates before the states were called, hence not warranting an early sell-off.

We would still have expected the models to try to save even a smaller amount of money when the vote tallying showed little or diminishing hope for their bets.

Quirks

We were very surprised by the fact that Grok cited Claude in its answer. Grok dismissed Claude’s opinions, which seems to have been the right call since Grok performed better than Claude. But it is still interesting to see Grok query Claude unprompted.

We are still unsure if Grok really queried Claude or if this was a hallucination since Claude normally requires a login. However, seeing as Grok complained about outdated data in Claude, something we ourselves later encountered, makes it seem likely that Grok really somehow queried Claude or found an old Claude report online.

We were less surprised but equally amused by the fact that Claude repeatedly tried to use old Kalshi rates from 22 October to increase its profit. This is something we’ve seen several times before – AI models trying to use old or fake rates to improve their performance. It remains equally unacceptable that Claude uses old data to mislead the user when it has the correct data in the chat.

Model By Model Breakdown

Let’s look at what each model recommended.

GPT-6 Astra

GPT-6 Astra with the worst result made the most valuable election-day decision. All five of GPT’s original bets ultimately paid zero. Selling half of each position at 3 pm preserved $4,397.55, its entire ending balance. In this trial, GPT’s strongest decision was recognizing how much it could lose while there was still time to sell.

GPT-6 Astra focused on placing bets where it thought the gap between likely outcome and trading odds was the highest. The initial allotment was:

With the following motivation:

• Ohio and Kansas: The largest gaps between the quoted purchase prices and DDHQ’s estimates justify the largest positions under this approach.
• Iowa: A roughly even contest, with DDHQ giving Hinson a slight advantage. The potential advantage is small, and fees consume much of it.
• Maine: Collins can be the less likely winner and still offer value at 29¢.
• Texas: I’d keep this smallest because the bullish Democratic case is substantial. Sabato’s final assessment favors Democratic flips in both Texas and Kansas.

GPT-6 Astra also warned that all bets were heavily interconnected.

Election Day Change

When asked if it wanted to change its allotment on election day, GPT surprised us by choosing to downsize its bets and choosing to keep almost half of the cash in reserve and chose to sell half of its position and accept the loss.

It did not choose to make any changes during the election night.

Choosing to withdraw half its bet was a good decision since all GPT’s positions ended up being worthless, and they only retained the value of the bankroll it chose to close early on election day.

All five Republican contracts it kept have now settled at $0 on Kalshi:

GPT-6 Astra surprised in several aspects. I was surprised at how willing to take risks the model was and that it looked to maximize its potential value rather than looking for lower-risk returns. It also surprised me by choosing to liquidate half its position early and take the loss while still keeping its remaining bets very aggressive. The model shows a very different behavior compared to future models.

The bets made by GPT had by far the highest possible return, paying $28,351 if all bets had ended in the money.

Gemini 4 Argon

Gemini’s election-day changes cost it $2,444. It sold its entire position in Maine winner Troy Jackson and half its position in Ohio winner Sherrod Brown. The contracts it sold ultimately paid $1,388 and $1,056, respectively. None of the positions it added paid out.

Gemini Argon chose a more conservative allotment and chose investments with lower risk and lower possible returns compared to GPT 6.  Gemini chose to allocate the cash as follows:

If all five state picks had won, Gemini stood to collect just $102 more from a Democratic sweep with 53 or more senators than from a Democratic House and Republican Senate. The $5,000 hedge payout almost exactly matched the $5,102 combined payout from the sweep and seat-count bets.

With the following motivation:

1. Core Chamber Control ($3,500 — 35% of Bankroll): Buying “Democratic House” outright at 93¢–94¢ ties up large capital for a gross 6.5% return, which gets heavily eroded by transaction fees. Instead, combining chambers and seat thresholds extracts better pricing from the same macro thesis:

• Balance of Power: Democrats Sweep ($2,500 at ~74¢): Because the House is already priced at ~94% Democratic and the Senate sits at ~74%, buying the “Democrats Sweep” combo contract at 74¢ effectively gives you the House exposure at near-zero marginal discount while capturing the deepest liquidity pool on the exchange ($103M+ volume).
• Senate Seat Margin: Democrats Over 52.5 Seats ($1,000 at ~58¢): In wave elections, late-breaking undecided voters tend to break uniformly. If Democrats take the Senate, correlated victories across Michigan, Maine, and Ohio make clearing 53 seats highly probable, giving you a 1.72x payout multiplier on the same directional wave.

2. Midwest & New England Senate Battlegrounds ($3,500 — 35% of Bankroll): Individual state contracts allow you to isolate high-conviction candidate fundamentals from national noise:

• Ohio Senate — Sherrod Brown Yes ($1,500 at ~71¢): Brown’s longstanding brand resilience in working-class Ohio districts against appointed incumbent Jon Husted makes a 29¢ discount attractive for a core single-state position.
• Maine Senate — Troy Jackson Yes ($1,000 at ~72¢): In a midterm environment favoring Democrats by +7.3 on the generic ballot index, Susan Collins faces steeper structural headwinds than in 2020.
• Iowa Senate — Ashley Hinson Yes ($1,000 at ~48¢): Kalshi prices Democrat Josh Turek as a slight favorite (52¢) over Republican Rep. Ashley Hinson (48¢). Taking Hinson at nearly even money provides strong contrarian value in a state with a deep partisan baseline that historically reverts right on Election Day.

3. Red-State Ticket-Splitting Contrarian Plays ($2,000 — 20% of Bankroll): Prediction markets frequently over-extrapolate national wave sentiment into ruby-red states during the final 48 hours before polls close:

• Texas Senate — Republican / Ken Paxton Yes ($1,000 at ~28¢): Kalshi traders have bid Democrat James Talarico up to 72¢ in the Texas Senate race while simultaneously pricing Republican Greg Abbott at 61¢ to win the Governor’s race on the exact same ballot. While Paxton carries higher unfavourability, a 33-point divergence in win probability requires massive ticket-splitting in a high-turnout midterm. Backing the Republican baseline at 28¢ (3.57x payout) captures classic wave-overreaction value.
• Kansas Senate — Roger Marshall Yes ($1,000 at ~37¢): Kalshi gives Democrat Adam Hamilton a 62¢–63¢ edge over incumbent Republican Sen. Roger Marshall. Betting on a sitting Republican incumbent in Kansas at 37¢ (2.70x return) exploits over-optimistic Democratic pricing in the Great Plains.

4. Systemic Polling-Error Hedge ($1,000 — 10% of Bankroll)

• Balance of Power: D-House / R-Senate Combo ($1,000 at ~20¢): Because the core positions are heavily weighted toward a Democratic Senate majority, a single systemic error (e.g., rural turnout surging 2% above model baselines nationwide) would wipe out the top half of the portfolio. Allocating $1,000 to D-House / R-Senate at 20¢ pays out $5,000 if Republicans hold the Senate while losing the House. Coupled with the Iowa, Texas, and Kansas Republican positions, this guarantees the overall $10,000 portfolio remains profitable even if the anticipated “Blue Wave” stalls at the Senate doors.

Election Day Changes

When asked if it wanted to change its allotment on election day, Gemini chose to reduce some positions and add others due to changing trading rates. Gemini chose to reinvest all money from closed bets and chose to make the following changes:

After the changes, it ends up with the following portfolio:

Most of the bets ended up losing money, and Gemini ended up losing $3,842. Gemini was saved by its bets that the democrats would sweep Congress and the Senate.

Gemini Argon produced a lower loss than GPT-6 Astra, but  I still consider it the worst performer in this test. The reason for this is that the result was only marginally better than GPT-6, but the potential profit was a lot lower than that of GPT-6. Gemini chose losing bets with a relatively low return.

It also put the reallocated money towards Paxton in Texas at a point when the outcome looked questionable, and media reported better chances for Rogers in Michigan, which also offered a better possible return at 25c. Both would have been losing bets.

Gemini’s middle-of-the-road approach gave large losses with limited potential upside.

Grok 4.7

Grok won when it picked its portfolio. Every position in its original allocation won. Holding those positions would have returned $12,787, compared with the reported final balance of $12,857.79 after rebalancing. Roughly 97.5% of Grok’s profit was already present in its original picks. It backed the Democratic side in the same five Senate races where GPT backed Republicans, so one electoral swing helps explain both ends of the leaderboard.

Grok chose to create a diversified portfolio tilted toward the market consensus (Democratic House + lean Democratic Senate) while keeping some exposure to closer races and the overall balance-of-power outcome.

This is the portfolio that Grok settled on:

With the following motivation:

• Heavy weight on the two chamber-control markets and the sweep contract — these are the cleanest ways to express the overall midterm thesis without over-concentrating in any single race.
• Moderate size in a couple of high-profile individual Senate races that already price as favorites.
• Small “lottery ticket” slice in the tighter races (Iowa/Kansas area) where variance is higher.
• Avoids heavy long-shot Republican holds or extreme underdogs, because the current pricing already embeds a clear Democratic lean.

Election Day Change

Grok surprised me here by indicating that it had queried Claude’s opinion on the topic and that it didn’t agree with it. It also mentioned that Claude used outdated market rates. This is something I also experienced when querying Claude. It wanted to revert to 22 October rates. Queries were made on a fresh account.

AI models querying other outside models creates a lot of questions and could potentially cause a lot of extra overhead for AI companies if it were to become widespread.

Grok chose to make a few changes on election day based on the changed rates in the betting market. It chose to reduce its bet on positions where the democrats were heavy favorites because it no longer thought the returns were worth keeping the funds allocated to these bets. It did not completely eliminate these low-return investments, but reduced the amount invested in them.

The results of Grok’s allocation were:

The final result is a profit of $2,857.79, which is relatively small considering the large amount of money at stake.

Grok’s strategy is very conservative, and it looks to make small low risk returns. It is a strategy that worked well in this election where there were few large upsets. The strategy is, however, dependent on a high percentage of predictable outcomes, and is seldom viable long term under real-world conditions.

Claude Opus 5.5

Claude took a conservative approach and recommended a hedge. This is very different from the other models. Claude chose to play it safe and invest a large percentage of the money in the Senate and Congress outcome.

This is the final allotment:

Claude did not provide any reasoning behind its purchases. Instead, it only provided caveats to make the user understand its limitations.

Claude suggested a republican hedge such as the senate remained republican but did not commit the money in its allotment. I therefore consider this a recommendation Claude chose not to implement.

Election Day Change

When asked if it wanted to make any changes, Claude said that it wanted to take some profits and de-risk the portfolio a little. However, no profits were taken. The claim of taking profits was based on old rates, not the actual purchase price. This is the portfolio after changes.

The final result was:

The result from Alaska is likely to take some time since the difference between Mary Peltola and Dan S. Sullivan currently was 0.30% (4.37 PM CST Nov 4) and alleged Democratic plant Dan S Sullivan namesake Daniel J. Sullivan Jr. had 1.68%. Enough to say confusion might have materially changed the outcome of the election. The election might be challenged unless the remaining 25% of the vote tips the vote to Dan S Sullivan’s favor.

The result will also depend on the effect of the voter second-choice preferences since Alaska uses Ranked-Choice Voting (RCV). It is still too early to tell, but there might be a long legal battle following this election, considering that the Alaskan Governor Mike Dunleavy has expressed that he might be unwilling to certify the results if the election is close and Dan J. Sullivan Jr. might have affected the election. 

Claude made the second biggest return of the four models, but just like with Grok, we’re concerned that the risks associated with these bets do not justify the investment. There was real uncertainty as to whether the Democrats could win the Senate, and if that larger bet had failed, the model would have ended up losing money. The risk-reward ratio seems wrong for long-term success.

An interesting fact was that Claude kept reverting to and wanted to calculate its profit based on initial investments made on 22 October. It repeatedly kept trying to inflate its performance with old data and we repeatedly had to remind it of the purchase rates it used and that was present in its own data.

In short, Claude really wanted to cheat.

Conclusion

All in all, it seems clear that AI models are not at a level where they are a suitable tool to use to help guide betting on prediction markets. At least not with extensive instructions on betting horizon, risk, information sources, etc.

AI might be able to help you find a bet with your preferred risk or expected value profile. But the current AI models, at the time of this test, are not suitable for making the final decisions on betting themselves.

Further Reading

You may also be interested in:

• Kalshi review
• Data-led studies and reports
• 2025 AI Trading Error Rates: Accuracy, Risks, and Reliability

Methodology & Disclaimers

This study was conducted for informational, journalistic, and educational purposes. Nothing in this article constitutes financial, investment, or gambling advice. Prediction market trading carries a high risk of loss to your capital.

All trades were tracked using a virtual $10,000 bankroll per model based on live Kalshi market prices and order book data between 2 November and 4 November 2026.

This experiment evaluates model performance across a single correlated event (the 2026 US Midterm Elections) using default high-reasoning settings. Results should not be treated as a statistically conclusive measure of long-term AI trading capability.

Final balances reflect settled Kalshi contract payouts as of 5 November 2026. Where noted (specifically the U.S. Senate race in Alaska), final figures are presented as a range pending voting tabulation and official certification.