Kalshi AI — Mispricing Audit

Report date 2026-07-07 · Horizon 45 days · Capital $1,000 · Methodology v3 (conviction-weighted, entry-band screened, EV-agnostic sizing) · Picks are live and cleared for auto-copy; rejects are machine-logged and shadow-tracked to settlement.

1. How this was researched

Mode: theme-match. Kalshi has no “AI” category (its categories are Sports, Crypto, Elections, Politics, Economics, Financials, Science and Technology, etc.), so the theme was resolved by keyword sweep over trading_events.title and trading_markets.title across all categories. Terms searched: AI, artificial intelligence, OpenAI, GPT, Anthropic, Claude, Fable, Mythos, Gemini, Grok, xAI, Nvidia, DeepMind, DeepSeek, Llama, Mistral, LLM, chatbot, superintelligence, Sam Altman. Sports-name false positives (tennis players named “Llamas”, “Hanatani”) were curated out by hand. The sweep surfaced ~63 events / 445 markets: five families of NVIDIA GPU rental-price ladders (A100/B200/H100/H200/RTX 5090 — ~330 low-volume strikes), model-release date ladders (GPT-5.6, GPT-6, Gemini 3.5 Pro, Claude Mythos), leaderboard races (LMArena overall/math, Datacurve DeepSWE coding), AI-policy markets (Anthropic’s DoD supply-chain designation, Fable 5 US access), IPO date ladders, and an AI-jobs market.

Stage-1 mechanical screens (24h volume ≥ 1,500; spread ≤ 5¢; not fully priced; not already a published pick; closes within 45 days) cut this to 8 strict passers plus 7 near-misses taken through full diligence. For each survivor: resolution rules pulled from the mirror DB, 14-day intraday history from market_snapshots, live top-of-book from the Kalshi public API, and primary-source news diligence. Two resolution precedents were established from already-settled rungs (queried via the public API): (a) every GPT-5.6 date rung through Jul 7 resolved NO despite the Jun 26 limited-partner release — Kalshi requires true general availability, not a gated preview; (b) KXFABLERESTORE rungs for Jul 7/Jul 10 resolved YES — Fable 5’s US access was restored ~Jul 1, which re-arms the July leaderboard races.

2. Markets reviewed

TickerQuestion (short)YES bid/askVol 24hDisposition
KXGPT-OPEN-26JUL09GPT-5.6 GA before Jul 916 / 1711,259PICK 1 — BUY NO
KXMATHAI-26JUL31-GEMIGemini #1 Arena Math, Jul 3128 / 312,120PICK 2 — BUY NO
KXTOPMODEL-26JUL31-CLAUFclaude-fable-5 top model, Jul 318 / 91,638PICK 3 — BUY YES (lottery)
KXCHAICUTS-26AUG06-T1AI #1 Challenger job-cut reason (July)43 / 48131Reject — band + volume floor (top watchlist)
KXCODEAI-26JUL31-CLAUClaude top on DeepSWE, Jul 3148 / 494,605Reject — coin-flip band
KXCODEAI-26JUL31-CHATChatGPT top on DeepSWE, Jul 3151 / 531,823Reject — coin-flip band
KXMATHAI-26JUL31-CLAUClaude #1 Arena Math, Jul 3156 / 585,416Reject — band screen (edge < 15¢)
KXMATHAI-26JUL31-CHATChatGPT #1 Arena Math, Jul 3113 / 141,680Reject — fairly priced
KXTOPMODEL-26JUL31-CLAUTopus-4-6-thinking top model, Jul 3168 / 72387Reject — unverifiable resolution view
KXGPT-OPEN-26JUL08GPT-5.6 GA before Jul 86 / 85,168Reject — same cluster, thinner edge
KXGPT-OPENB-26JUL10GPT-5.6 GA before Jul 1075 / 7714,092Excluded — already an active pick
KXGPT-OPENB-26JUL15GPT-5.6 GA before Jul 1584 / 85210Reject — no edge, volume died
KXGEMINI-GEMI35P-26JUL17Gemini 3.5 Pro before Jul 1719 / 211,233Reject — volume floor
KXGEMINI-GEMI35P-26JUL24Gemini 3.5 Pro before Jul 2467 / 74397Reject — 7¢ spread
KXANTHROPICRESCIND-26JUN-26AUG01DoD/WH rescinds Anthropic designation before Aug 133 / 3416Reject — illiquid
KXFRONTIER-FRON-26AUG01AI solves a FrontierMath Open Problem before Aug 115 / 17421Reject — volume floor
KXFABLEDISABLED-27-26JUL31Fable 5 US access disabled again before Jul 314 / 694Reject — tail, lottery slot used

Also screened out mechanically: ~330 GPU rental-price ladder strikes (nearly all under the volume floor), Claude Mythos date rungs (1–2¢, fully priced), “Best AI July” brand markets (ANTH at 96–97, fully priced), IPO date ladders (outside horizon or illiquid), and KXTRUMPSAYCOMPANY mention markets (announcer-noise category, skipped by policy).

3. Picks — detail and thesis

Pick 1 — KXGPT-OPEN-26JUL09 · BUY NO @ 0.85 · HIGH CONVICTION

“Will OpenAI release a model called GPT-5.6 or greater before Jul 9, 2026?” — closes 2026-07-09 03:59 UTC (end of Jul 8 ET).

Current: NO ≈ 83–84¢ (YES 16/17)
Limit: 85¢
My prob (NO): 92% vs market ~84%
Entry band: favorite (60–90¢)
Cluster: gpt56-ga-timing

Thesis. For YES to win, OpenAI must take GPT-5.6 to general availability within ~41 hours of this report. Nothing is announced. As of today OpenAI’s own materials describe GPT-5.6 (Sol/Terra/Luna) as a limited preview for ~20 government-vetted partner organizations — “not available in ChatGPT during the preview” — after the U.S. government asked for a staggered rollout on Jun 26. Critically, the resolution precedent is already established: every date rung from Jun 26 through Jul 7 settled NO even though the model was technically shipped to partners, so only a true GA event counts. The residual 16–17¢ on this rung is spillover from the market’s (unconfirmed, self-referential) “GA on Thursday Jul 9” consensus — but a Jul 9 GA resolves this rung NO anyway, since it needs GA on Jul 7–8.

Evidence.

Tail risk (cleanest way to lose): OpenAI surprise-ships GA tomorrow (Jul 8) with no pre-announcement — it has a history of sudden drops, and competitive pressure is real (Fable 5 restored Jul 1, Gemini 3.5 Pro rumored Jul 17). I hold 8% for this.

Numbers. 176 contracts × 85¢ = $149.60 at risk (15.0% cap for HIGH tier — sized by tier/band, not EV). Max payout $176.00. Edge +7¢/contract, EV ≈ +8.2% (recorded, not used for sizing).

Liquidity (entry context). YES 16/17; buying NO fills 23 @ 84¢ + 166 @ 85¢ ≈ 189 contracts ≤ limit; vol24h 11,259; OI 13,968. Slippage ≤ 1¢ at this size.

Pick 2 — KXMATHAI-26JUL31-GEMI · BUY NO @ 0.73 · MEDIUM CONVICTION

“Will Gemini be the #1 model on the Arena AI Text Leaderboard (Math) on Jul 31, 2026 at 10:00 AM ET?”

Current: NO ≈ 72¢ (YES 28/31)
Limit: 73¢
My prob (NO): 79% vs market ~72%
Entry band: favorite (60–90¢)
Cluster: lmarena-math-jul31

Thesis. On the Jul 1 Arena Math snapshot, Claude holds both top slots — claude-opus-4-6-thinking (1518±12) and claude-fable-5 (1517±37) — while the best Gemini entry is the mid-tier gemini-3.5-flash at 1504±21, fourteen points back. For a YES, Gemini must overtake two entrenched Claude models within 24 days. The realistic path runs through Gemini 3.5 Pro, which (a) is not released, (b) has only an unconfirmed ~Jul 17 launch date after slipping from June, and (c) would then need to accumulate enough Math-category votes for a #1 ranking in under two weeks — Math-slice votes accrue slowly (fable-5, released Jun 9, still has only 243 math votes). The market’s 28¢ on Gemini overweights launch hype relative to that mechanical funnel.

Evidence.

Tail risk (cleanest way to lose): Gemini 3.5 Pro launches ~Jul 17 with its rumored “Deep Think” reasoning layer, debuts as a math monster, and Arena rushes it onto the Math board with a high debut score before month-end. New frontier models have debuted #1 before; that is why this is MEDIUM, not HIGH.

Numbers. 95 contracts × ~73¢ = $69.35 at risk (under the 7% MEDIUM cap). Max payout $95.00. Edge +6¢, EV ≈ +8.2% (recorded, not used for sizing).

Liquidity (entry context). YES 28/31; buying NO fills ~15 @ 72¢ + 176 @ 73¢ ≈ 191 contracts ≤ limit; vol24h 2,120; OI 5,749.

Pick 3 — KXTOPMODEL-26JUL31-CLAUF · BUY YES @ 0.09 · LOW CONVICTION ASYMMETRIC LOTTERY — 2% CAP

“Will claude-fable-5 be the top-ranked AI model on Jul 31, 2026 at 10:00 AM ET?” (Arena leaderboard; ties under Rank (UB) break to highest Arena Score; resolution uses the “Remove Style Control” view.)

Current: YES 8/9¢
Limit:
My prob (YES): 25% vs market ~9%
Entry band: <35¢ tail — explicit lottery slot (1 per run)
Cluster: lmarena-top-jul31

Thesis. The market pays 11:1 against the model that is already #1. On the Jul 1 Arena text snapshot, claude-fable-5 sits first at 1509±9 — above claude-opus-4-6-thinking (1504±4) — and the market’s stated tiebreak (highest Arena Score) favors it. Its 4,350-vote count is thin because US access was disabled for most of June; access was restored ~Jul 1 (Kalshi’s own FABLERESTORE rungs settled YES), so July brings a large US vote influx to a model with restoration-hype demand. If its score merely holds through Jul 31, this contract wins at 9¢.

Evidence.

Tail risk (cleanest way to lose) — and why this is a lottery, not a favorite: the market resolves on the style-control-removed view, which I could not retrieve; the persistent 8–9¢ price despite a visible #1 ranking implies informed traders may know fable-5 ranks below opus-4-6-thinking there. Second risk: the post-restore vote flood regresses its thin 1509 downward. Third: GPT-5.6 or Gemini 3.5 Pro debuts #1 before Jul 31, beating every Claude strike at once. Size is capped at 2% of capital accordingly.

Numbers. 222 contracts × 9¢ = $19.98 at risk (2.0% lottery cap — the only sub-35¢ position this run). Max payout $222.00. Edge +16¢, EV ≈ +178% (recorded, explicitly NOT used for sizing — v3 sizes by tier and band only).

Liquidity (entry context). YES 8/9; ~500 contracts offered at 9¢ (NO bid 91 × 500); vol24h 1,638; OI 4,493. Fill is trivial at this size.

4. Recommended $1,000 portfolio

#TickerSideLimitContractsCostConviction / band (sets size)Max payoutEdge ¢EV %
1KXGPT-OPEN-26JUL09NO85176$149.60HIGH · favorite → 15% cap$176.00+78.2
2KXMATHAI-26JUL31-GEMINO7395$69.35MEDIUM · favorite → 7% cap$95.00+68.2
3KXTOPMODEL-26JUL31-CLAUFYES9222$19.98LOW · tail lottery → 2% cap$222.00+16178
Total deployed$238.93Cash held: $761.07 (76.1%)$493.00~22 blended

Edge and EV columns are recorded for calibration only. They did not set position sizes — the resolved v1/v2 track record showed stated EV was anti-predictive, so v3 sizes strictly by conviction tier and entry band. The blended EV is dominated by the lottery leg, which is precisely why it is capped at 2%.

Cluster exposure (cap: 15% of capital per cluster)

ClusterLegsCost% of capitalCap
gpt56-ga-timingPick 1$149.6015.0%15% — at cap
lmarena-math-jul31Pick 2$69.356.9%15%
lmarena-top-jul31Pick 3$19.982.0%15%

Correlation note: Picks 2 and 3 both ride Arena state on Jul 31 but in opposite failure modes — a Gemini 3.5 Pro mega-debut is the one event that hurts both. Even treated as a single cluster they total $89.33 (8.9%), inside the 15% cap. The GPT-5.6 cluster deliberately holds only the Jul 9 rung; the Jul 8 rung and Jul 15 rung were rejected rather than stacking correlated date legs (the exact v1 failure mode).

Conviction exposure

TierCost deployed% of capitalPer-pick cap
HIGH$149.6015.0%15%
MEDIUM$69.356.9%7%
LOW / lottery$19.982.0%2% (tail rule)

Risk profile

Execution notes

5. What I rejected and why

Every Stage-2 survivor that didn’t become a pick is machine-logged in picks.json and shadow-tracked to settlement — if the rejects outperform the picks, the screen adds nothing, and we want to know that. Entry-band cuts are marked [band]: this is the v3 discipline making itself visible.

6. Sources