What Will Trump Say at the Correspondents’ Dinner? AI Odds

TL;DR

• Kalshi's 34 mention contracts for tonight's speech run from Fake News at 87¢ down to Epstein at 4¢ - and seven price-blind AI models scored every one.
• The models' biggest consensus call: MAGA at 31¢ versus an 81% blend - all seven models between 72% and 92%.
• Overnight, the market moved onto the panel's numbers on Iran (3+ mentions), which fell from 56¢ to 35¢ against a 30% model blend.

Tonight, President Trump steps to a podium in front of the White House press corps, and Kalshi has turned his script into 34 separate markets. Will he say "Fake News"? Almost certainly, says an 87-cent price. Will he say "Epstein"? Four cents says no chance. In between sit 32 other words and phrases, and somewhere in that list, the crowd has almost certainly mispriced a speech habit. We asked seven AI models to score every single one, without letting them see a single market price, and overnight, the market moved toward the models on several of their strongest calls. It hasn't caught up everywhere.

The Market

Venue	Kalshi — a CFTC-regulated exchange (18+; availability varies by state, as of July 2026)
The event	"What will Trump say during the White House Correspondents' Dinner?" — 34 yes/no contracts, one per word or phrase
The contract	Each market resolves YES if Trump says that word/phrase as part of the July 24, 2026 dinner, per Kalshi's settlement rules
Event date	Tonight — Friday, July 24, 2026, 7:00pm EDT
Prices below as of	the morning of July 24, 2026

The Setup: This Is Not a Normal Correspondents' Dinner

Context decides what gets said at a podium, and this event carries more of it than any correspondents' dinner in memory:

• It's a rescheduled dinner. The original April 25 event at the Washington Hilton was cut short by a shooting; the attacker was charged with attempted assassination of the president. Friday's redo moves to the Waldorf Astoria with reduced capacity and enhanced security.
• It's Trump's first correspondents' dinner as president. He skipped all of them in his first term. This time he accepted the White House Correspondents' Association's invitation to speak.
• There is no comedian. The WHCA dropped its comedian headliner in 2025, and April's entertainment was mentalist Oz Pearlman. Two straight dinners without a booked comic is a running storyline, and an open lane for the featured speaker.
• The week's news is his own election-security push. Trump's July 16 primetime address (declassified documents, claims of Chinese interference, ordered prosecutions) is the dominant political story going into the weekend.
• He handed out the World Cup trophy five days earlier. Spain beat Argentina 1-0 at MetLife on July 19, with Trump presenting.
• America's 250th birthday was three weeks ago, a celebration he has referenced all month.
• Morning-of update: NBC News reports that mentalist Oz Pearlman returns as tonight's entertainer, that Trump is expected to "briefly address" the April shooting, and that he plans to reuse jokes originally written for the April dinner. Melania Trump, JD Vance, and Marco Rubio, all present in April, are not attending this time.

Our models received this context (verified, dated facts with the sources above), minus the morning-of update, which broke after the verdicts were generated on July 22–23, and nothing else beyond each contract's terms. Crucially, they never saw the market prices, so where they land far from the market, that's an independent read, not an echo. Where the morning's news bears on a specific verdict (the entertainer confirmation most of all), we flag it below rather than silently standing on a stale premise.

A disclosure on the panel: our standard Model Verdict panel includes Grok and Gemini Pro; both were unavailable through our routing provider when these verdicts were generated, so this article runs on seven models: Claude Fable, Claude Opus, Claude Sonnet, ChatGPT, GLM, Kimi, and DeepSeek. The Claude verdicts were generated July 22, the rest July 23; prices in the table are the morning-of-event snapshot, July 24.

The Full Board: 33 Phrases, Seven Models, One Blend

Sorted by the gap between the AI blend and the market. A positive edge means the models think YES is more likely than the market price implies; negative means they think the market is too high. (The 34th contract, "Event does not qualify," is excluded; see The Traps below for the honest reason.)

Phrase	Market	Fable	Opus	Sonnet	GPT	GLM	Kimi	DeepSeek	Blend	Edge
MAGA / Make America Great Again	31¢	90	85	75	74	72	82	92	81%	+50
Hoax	16¢	90	85	5	69	58	78	35	60%	+44
World Cup	43¢	93	85	88	78	62	80	85	82%	+39
Kamala	30¢	85	83	8	69	63	73	85	67%	+37
Israel / Israeli	25¢	90	42	75	36	62	78	35	60%	+35
China	57¢	97	90	88	90	92	93	85	91%	+34
Comedy / Comedian	33¢	90	85	62	64	58	62	45	67%	+34
Afford / Affordable / Affordability	29¢	82	85	75	34	42	42	75	62%	+33
Hottest	67¢	80	15	14	32	8	80	20	36%	−31
AI / Artificial Intelligence	25¢	85	85	70	46	22	72	12	56%	+31
IQ / Genius	34¢	82	90	55	71	32	70	45	64%	+30
Supreme Court	19¢	60	32	28	44	38	62	65	47%	+28
Epstein	4¢	30	85	30	16	12	35	2	30%	+26
Barack Hussein Obama	53¢	55	40	6	26	22	45	12	29%	−24
Weave / Weaved / Weaving	13¢	55	20	20	24	35	32	57	35%	+22
Ballroom	75¢	93	40	35	61	15	76	68	55%	−20
Newsom / Newscum	30¢	57	85	8	44	68	62	25	50%	+20
Radical Left	55¢	91	85	10	72	82	85	85	73%	+18
TDS / Trump Derangement Syndrome	32¢	55	55	22	66	38	62	45	49%	+17
Crypto / Bitcoin	6¢	55	12	10	17	22	32	12	23%	+17
Venezuela	48¢	78	10	35	24	30	70	5	36%	−12
Cheat / Cheater / Cheating	50¢	87	72	8	68	62	82	45	61%	+11
Communist / Communism	62¢	95	78	45	72	62	72	72	71%	+9
Transgender	27¢	78	15	5	31	28	60	35	36%	+9
Iran (3+ times)	35¢	45	55	13	29	28	30	8	30%	−5
Fake News	87¢	94	90	88	78	78	70	85	83%	−4
Stock Market	65¢	92	85	72	59	45	80	48	69%	+4
Sleepy Joe	53¢	78	70	62	48	35	63	45	57%	+4
Oil	46¢	82	15	10	30	62	80	25	43%	−3
Autopen	31¢	75	15	3	29	38	60	15	34%	+3
250	59¢	94	60	55	68	35	70	45	61%	+2
First Term	60¢	80	60	35	57	38	60	95	61%	+1
Movie Star	18¢	22	15	15	24	9	28	3	17%	−1

Model estimates generated July 22–23, 2026, price-blind, from a seven-model panel. These are model estimates, not predictions of fact and not financial or trading advice. Models are frequently wrong; the market price reflects real traders' money.

Why the Models Disagree: Two Different Trumps

Read down the Fable and Sonnet columns and you'll notice they're not really scoring the same speech. Fable consistently prices rally Trump: the unscripted riffer who works his entire greatest-hits catalog into any long block of remarks, which is why it has "Radical Left" at 91 and "Ballroom" at 93. Sonnet consistently prices formal-dinner Trump: a shorter, more scripted address in a room where he was nearly killed three months ago, which is why it has those same rally staples at 10 and 35. The four new panel models mostly land between the two Claudes, case by case.

That split is the honest state of the question. Nobody knows which Trump shows up Friday: the one who described his own style as "the weave", or the one delivering prepared remarks under the heaviest security in the event's history. The market price on most contracts sits right between the two personas. Which means the interesting contracts are not the ones where the models disagree with each other. The interesting ones are the handful where all seven agree and still disagree with the market.

First, a Receipt: The Market Moved Toward the Models Overnight

Before the edges, the scoreboard update. When these verdicts were generated on July 22–23, the panel's most aggressive NO call was Iran (3+ times): blend 30% against a market then at 56¢. Overnight, as reporting firmed up a shorter, tightly programmed evening, the market marked Iran-3x down to 35¢, a 21-cent move onto the models' number. The same repricing wave hit the duration ladder in our companion article, where the market walked rung-by-rung onto the panel's curve. Blind model estimates converging with next-day price action is the whole thesis of this series; it won't always work this cleanly, but it did this week.

Where the Consensus Edge Is

The overnight repricing dragged most YES prices down, but the models' strongest calls were already well above the old prices, so several consensus edges got wider, not narrower:

MAGA / Make America Great Again at 31¢. AI blend: 81%, all seven models between 72% and 92%. The largest consensus gap on the board got eight cents larger overnight. Fable: "Trump says 'MAGA/Make America Great Again' in the overwhelming majority of his prepared speeches, typically as a closing line, and he is confirmed to speak." DeepSeek adds that his first correspondents' dinner after an assassination attempt is "a dramatic moment" tailor-made for the slogan. The market's implicit bear case (a short, scripted set that never reaches for the tagline he has closed with for a decade) is exactly the scenario the skeptical-persona models already price, and even they sit at 72–75. A shorter speech still ends with a closing line.

China at 57¢. Blend: 91%, a spread of just 12 points. The closest thing to a lock the panel found, six cents cheaper than yesterday. Kimi: "Trump says 'China' in nearly every extended speech, and his July 16 address alleging Chinese election interference is the week's dominant story."

World Cup at 43¢. Blend: 82%. Fifteen cents cheaper than yesterday, with every model at 62 or above. GPT: "Trump's trophy presentation five days earlier makes the World Cup a highly salient, flattering accomplishment likely to appear in his prepared remarks." A prepared-remarks brag that survives any speech length, priced under a coin flip.

Comedy / Comedian at 33¢. Blend: 67%, with an honest asterisk. The models scored this when the entertainment lineup was unconfirmed, and part of their case was Trump as the evening's only headliner. This morning's reporting confirms mentalist Oz Pearlman returns; that premise is half-aged. What survives it: a mentalist is still not a comedian, the two-years-without-a-comic storyline is still sitting on a tee, and Sonnet's driver was never the lineup: "Two straight dinners without a booked comedian is a live storyline Trump... would likely needle directly." We'd still take the models' side here, with reduced conviction versus the table's raw +34.

And one clean NO-side call remaining:

Barack Hussein Obama at 53¢. Blend: 29%. The panel isn't buying the majority-priced YES. The contract needs the full three-word formulation, and as GLM notes, Trump "typically defaults to just 'Obama'." The market looks like it's pricing "attacks Obama" when the contract demands the specific rhetorical flourish. This is now the board's widest overpriced contract that isn't a model-split trap.

The Traps: Big Edges That Aren't

Transparency matters more than a clean story, so here is where the table would mislead you:

• "Event does not qualify" (5¢) is excluded from our table because three of our seven models flunked it. It's a meta-contract (YES means the dinner doesn't happen or doesn't qualify under Kalshi's terms), and GLM, Kimi, and DeepSeek each output 95-99% while their own written reasoning says the opposite: they scored it as if it were a mention contract, then inverted the sign. The models that parsed it correctly put it at 6%, right at the market. We're showing you this because a blend is only as good as the comprehension underneath it, and this contract broke three models' comprehension.
• Hottest (67¢, blend 36%) looks like a big NO edge, but the panel is split 8-to-80: Fable and Kimi know "hottest country in the world" is a signature line he has used from Davos to the State of the Union; five models treat it as obscure. The market firmed this contract up two cents overnight; it is siding with the two models that have the receipts. We'd call the market closer to right than our own blend here.
• Kamala (30¢), Hoax (16¢), Radical Left (55¢) all show fat positive edges, all are really the rally-vs-formal persona bet in disguise (Sonnet at 8, 5, and 10 against teammates in the 80s). If rally Trump shows up these hit; if formal Trump shows up they don't. That's a speech-tone bet, not a mispricing. Though at Hoax's new 16 cents, the persona bet itself has gotten cheap.
• Epstein (4¢, blend 30%) is one contrarian in disguise: Opus at 85% argues he'd address it to defuse it; five of seven models are at 35 or below on his long track record of avoidance. The 4-cent market is siding with the majority.
• Affordability (29¢, blend 62%) is a real signal but softer than it looks: the three Claudes and DeepSeek are at 75-85 on the grounds that "making America affordable again" is his stated top domestic priority and anchored his State of the Union economic case; GPT, GLM, and Kimi are at 34-42 betting the dinner speech skips the economic stump. Half persona bet, half real edge, and eight cents cheaper than yesterday.

The Best Value on the Board

If you take one thing from the models: MAGA at 31 cents is the mispricing. It is the only contract where the largest gap and the strongest consensus coincide, seven models spanning three different AI labs, including the ones that assume a subdued formal speech, all land between 72% and 92% against a 31-cent price that got cheaper overnight. Runners-up on the YES side: World Cup at 43¢ (blend 82%, every model 62+) and China at 57¢ (blend 91%, the panel's tightest agreement). On the NO side: Barack Hussein Obama at 53¢, where the market appears to be pricing the topic when the contract demands the specific three-word phrase. The panel's other big NO call. Iran (3+ times), was 56¢ when the models made it; the market took it to 35¢ before we published.

To be explicit: these are model estimates for information and entertainment, not financial advice and not a recommendation to trade any contract. The models were graded against real settlements all summer and they are wrong regularly, the Traps section above shows three of them failing a contract in this very article. Event contracts are for adults 18+ where legal.

FAQ

Is Kalshi legal in my state?

Availability varies by state and changes over time, so check your eligibility directly on the platform before doing anything. Kalshi is a federally regulated exchange (event contracts are CFTC-regulated and are not sports betting), but access rules differ by jurisdiction and you must be 18 or older. That is the state of play as of July 2026.

Are the model verdicts advice?

No. Model verdicts are statistical estimates of how a market might resolve, not financial advice, and nothing here is a recommendation to buy or sell any contract.

How exactly does a mention count?

Per Kalshi's rules, each market resolves YES if Trump says the listed word or phrase as part of the July 24, 2026 Correspondents' Dinner. Variants listed in the contract name (e.g., "Cheat / Cheater / Cheating") count for that contract, and the "Iran" contract specifically requires three or more mentions. Kalshi's settlement rules govern edge cases, read the contract before assuming a partial or paraphrased utterance counts.

Why didn't the models see the market prices?

Anchoring. Our summer-long grading shows that models shown the market price cluster tightly around it, which makes their estimate an echo rather than an opinion. Price-blind estimates are what make the edges in the table meaningful, and also what makes them riskier: sometimes the crowd knows something the models don't.

Will these prices move before Friday?

They already have, 29 of 34 contracts repriced overnight as the program details firmed up, several by 10 cents or more. The table above is the morning-of-event snapshot, July 24; expect more movement right up to the podium. For the rest of the picture, see the companion pieces on how long the speech runs and the reschedule market itself.