Kalshi AI — Mispricing Audit
2026-07-24 · Theme: AI · Horizon: 45 days · Capital: $2,500 · Methodology v3.3 (conviction-weighted, EV-agnostic sizing; Stage 2.5 resolver pricing; no tails, no low-conviction deploys). Picks count toward the public track record and are cleared for auto-copy.
TL;DR. The AI category is efficiently priced this week. Of 78 AI-themed events and 24 Stage-2 candidates, exactly one contract cleared every screen: KXTECHRANKLISTAICODE-26JUL27-KIMI (Kimi #1 on LM Code Arena at the Jul 27 snapshot) — a deep favorite backed by a verified 44-point Elo lead with 3 days to resolution. Everything else — including the single largest modeled edge found (+27¢ on Challenger AI job cuts) — fails a v3 screen and is machine-logged as a shadow-tracked reject. Deployment: $279 (11.2%); cash: $2,221 (88.8%). Cash is a position.
1 · How this was researched
- Mode: theme-match. No
AI category exists in trading_events.category (checked; categories are Sports/Crypto/Politics/…). I selected markets by keyword across all categories: openai, anthropic, claude, gpt, gemini, grok, "ai", artificial intelligence, nvidia, deepseek, llama, chatgpt, superintelligen*, xai, llm, agi against event and market titles (a first pass matching %agi% pulled in esports team names — refiltered excluding Sports/Crypto). Result: 78 AI-themed events closing within 45 days.
- Data. Read-only DB mirror for market/event/rules metadata (note: this mirror has no
market_snapshots table — the schema endpoint is authoritative — so 14-day intraday history was not available; live prices, 24h volume, open interest, and full orderbooks were taken from the Kalshi public API at audit time, ~2026-07-24 UTC).
- Resolver-first diligence (Stage 2.5). Every candidate's rules text was pulled and the named resolution source was fetched directly where reachable: LM Arena text/code/math leaderboards (Jul 21 data date), the Arena leaderboard changelog, Datacurve DeepSWE standings via llm-stats, Challenger Gray & Christmas June report, Epoch AI FrontierMath pages. The Ornn GPU compute index (resolver for all NVIDIA compute markets) is JS-rendered and returned no data to fetches — that inaccessibility alone rejected the entire GPU family.
- Calibration cross-checks. Where a weekly version of a market had already settled, I pulled its finalized result from the API and treated it as ground truth about the resolver's actual behavior (this materially changed two conclusions — see the fable-5 anomaly below).
- Exclusions. 38 already-published feed picks were hard-excluded up front; the AI-relevant ones are KXTOPMODEL-26JUL31 (CLAUF, CLAUT), KXMATHAI-26JUL31-GEMI, KXGEMINI-GEMI35P-26JUL31, KXGPT-OPEN-26SEP01, KXRTX5090WS-26AUG07-0.750. Mention/announcer markets (KXFEDMENTION, KXHEARINGMENTION, KXTRUMPSAYCOMPANY) were cut by rule.
The fable-5 resolver anomaly (why this report is conservative on leaderboard markets). The live Arena text leaderboard (Jul 21 data) shows claude-fable-5 at #1 (1507±6) over claude-opus-4-6-thinking (1505±4) — yet the market prices fable-5-top-on-Jul-27 at 1–2¢ and opus-4-6-thinking at 91–92¢, and the settled record agrees with the market: the Jul 6, Jul 13 and Jul 20 weeklies all finalized YES for opus-4-6-thinking and NO for fable-5, including Jul 20 — one day before the leaderboard data date on which I can see fable-5 ahead. I could not fully reconcile the resolver's view (Rank-UB mechanics, style-control toggle state, or a snapshot-timing artifact) with the public page. Mid-July the changelog also shows Arena recalculated fable-5's scores to count only post-Jul-1 votes. This is precisely the market family that produced this program's costliest historical loss, so every leaderboard candidate here carries explicit resolver deductions, and one (Chinese AI monthly) was rejected outright on unpriceable resolver mechanics.
2 · Markets reviewed
| Ticker | Question (short) | Yes bid/ask ¢ | Vol 24h | Verdict |
|---|
| KXTECHRANKLISTAICODE-26JUL27-KIMI | Kimi #1 on LM Code Arena, Jul 27 | 91 / 92 | 1,186 | PICK — BUY YES |
| KXCHAICUTS-26AUG06-T1 | AI = #1 job-cut reason, July Challenger report | 45 / 47 | 0 | REJECT — band screen (best modeled edge found) |
| KXCHINAAI-26JUL31-ALIB | Alibaba top Chinese AI on Arena text, Jul 31 | 74 / 75 | 1,926 | REJECT — resolver risk |
| KXCODEAI-26JUL31-CHAT | ChatGPT #1 on Datacurve DeepSWE, Jul 31 | 86 / 90 | 4,886 | REJECT — no edge at fillable ask |
| KXMATHAI-26JUL31-CLAU | Claude #1 on Arena math board, Jul 31 | 92 / 96 | 161 | REJECT — no edge at ask |
| KXTOPMODEL-26JUL27-CLAUT | claude-opus-4-6-thinking top model, Jul 27 | 91 / 92 | 2,147 | REJECT — unreconciled resolver view |
| KXFRONTIER-FRON-26AUG01 | AI solves a FrontierMath open problem by Aug 1 | 7 / 10 | 308 | REJECT (NO side) — slippage |
| KXGEMINI-GEMI35P-26AUG16 | Gemini 3.5 Pro released before Aug 16 | 64 / 66 | 208 | REJECT (NO side) — coin-flip band |
| KXB200MS-26JUL-6.500 | B200 July avg > $6.50/hr (Ornn) | 40 / 41 | 1,106 | REJECT — resolver data unverifiable |
| KXH200MS-26JUL-4.500 | H200 July avg > $4.50/hr (Ornn) | 40 / 41 | 1,428 | REJECT — resolver data unverifiable |
| KXRTX5090MS-26JUL-0.500 | RTX 5090 July avg > $0.50/hr (Ornn) | 49 / 50 | 2,148 | REJECT — resolver data unverifiable |
| KXFABLEDISABLED-27-26AUG31 | Fable 5 US access disabled again before Aug 31 | 4 / 9 | 0 | REJECT (NO side) — no edge, illiquid |
| KXANTHROPICRESCIND-26JUN-26AUG01 | DoD/WH rescinds Anthropic designation by Aug 1 | 5 / 11 | 40 | REJECT — no edge, wide spread |
| KXCLAUDE-NXTMYTH-26SEP01 | Next Mythos-class model before Sep 1 | 30 / 35 | 62 | REJECT — no evidence, illiquid |
| KXLLM1-26JUL31-A | Anthropic best AI in July | 98 / 99 | 427 | Stage-1 cut — fully priced |
| KXIPOANTHROPIC-DATE-26SEP01 | Anthropic confirms IPO before Sep 1 | 3 / 5 | 20,270 | Stage-1 cut — fully priced |
| KXCLAUDE-MYTH-26SEP01 | Anthropic releases Mythos before Sep 1 | 3 / 4 | 2,047 | Stage-1 cut — fully priced |
| KXVERARELEASE-VERARUBIN-* | NVIDIA Vera Rubin release timing | 2 / 98 | 0 | Stage-1 cut — no market |
| KXAIREVIEW-26JUN-SEP01 | Mandatory federal pre-release AI review by Sep 1 | 9 / 15 | 102 | Stage-1 cut — volume + spread |
| KXCHINAAI-26JUL27-ALIB | Alibaba top Chinese AI (weekly), Jul 27 | 93 / 94 | 1,696 | No edge at 94 — see monthly reject |
| KXTOPMODEL-26JUL31-*, KXMATHAI-26JUL31-GEMI, KXGEMINI-GEMI35P-26JUL31, KXGPT-OPEN-26SEP01 | — | — | — | Excluded — already live feed picks |
| KXFEDMENTION / KXHEARINGMENTION / KXTRUMPSAYCOMPANY | — | — | — | Excluded — mention markets |
3 · The pick
Pick 1 — KXTECHRANKLISTAICODE-26JUL27-KIMI · BUY YES @ 0.93 limit · HIGH CONVICTION
"Will Kimi be the #1 model on the LM Code Arena Leaderboard on Jul 27, 2026 at 10:00 AM ET?" — current 91¢ bid / 92¢ ask.
My probability: 95% vs market implied: ~92% (ask) / 93¢ at fillable size
Conviction tier / entry band: HIGH · deep favorite (>90¢) — thin edge by construction, sized by conviction
Cluster: kimi-k3-code-arena
Resolver: LM Code Arena leaderboard, brand-level ("Kimi" = any Kimi model), snapshot Jul 27, 2026 10:00 AM ET; ties resolve by the publisher's methodology, else shared rank paying $1/n. Priced risks: −3 pts (a brand-new frontier entrant added to Code Arena and out-scoring 1678 with enough votes inside 3 days, or a methodology surprise), −2 pts (an Arena score recalculation hitting kimi-k3 — Arena demonstrably recalculated claude-fable-5's scores mid-July), −0.5 pts (ordinary vote drift). Evidence-based ~99.5 → model_prob 95.
Mispricing thesis. This is a pay-the-cents lock-harvest, not a narrative bet: kimi-k3 leads the named resolver by 44 Elo points (1678 vs claude-fable-5's 1634, gpt-5.6-sol-xhigh 1630 — Jul 21 data), which is multiple confidence intervals; ordinary vote flow cannot close it in 3 days. The identical market settled YES last week (Jul 20), confirming the resolver already sees Kimi #1.
Evidence.
- Arena Code leaderboard (Jul 21 data): kimi-k3 #1 at 1678 (1,828 votes); #2 claude-fable-5 1634; #3 gpt-5.6-sol-xhigh 1630. A 44-point gap ≈ untouchable over a 3-day window by votes alone (arena.ai/leaderboard/code).
- Settled ground truth: KXTECHRANKLISTAICODE-26JUL20-KIMI finalized YES (and the Jul 13 weekly, before kimi-k3's Jul 16 arena listing, finalized NO) — the resolver's behavior is empirically consistent with the public page for this board, unlike the text board (see anomaly note above).
- Threat scan: no plausible >1678 entrant inside the window — Gemini 3.5 Pro is unreleased (delayed twice, rebuilt per Bloomberg via TechTimes Jul 13; Kalshi prices release-by-Jul-31 at 3–4¢); OpenAI's best code entry sits 48 points back; new arena entrants need days of votes to place (changelog).
- Brand-level wording ("Kimi", not "kimi-k3") adds a safety margin: any Kimi variant at #1 resolves YES.
Tail risks (cleanest way to lose).
- #1: Arena recalculates or re-buckets kimi-k3's score before Monday 10:00 ET (the fable-5 precedent shows recalcs happen with no notice). Priced −2.
- A surprise frontier code-model launch that both lists and out-ranks 1678 within 72 hours. Priced −3.
- Losing 93¢ to win 7¢ — one loss wipes ~13 wins; that asymmetry is why size stays at 11% despite the HIGH tier allowing 15%.
Numbers: 300 contracts @ 93¢ limit = $279.00 cost · max payout $300.00 · edge +2¢ · EV ≈ +2.2% (~$6; recorded for calibration only — did NOT set size) · ~0.5¢/contract fees at this price point
Liquidity / entry context: yes_bid 91 (103 shown), yes_ask 92 (≈17 fillable), ≈400 more at 93 → ≈417 contracts fillable at ≤93¢; volume 24h 1,186; total volume 5,810; open interest 3,154; spread 1¢.
Price history: intraday history unavailable (no snapshots table in this mirror); API last=92, previous close 91 — no adverse 48h move; market opened Jul 20 post-settlement of last week's YES.
Disclosed screen deviation: 24h volume is 1,186 vs the Stage-1 ≥1,500 line. I kept the pick because the screen's purpose (fillability) is directly verified by the book — 417 contracts rest at ≤93¢ against a 300-contract order — and total volume/OI are healthy. Flagged for the record; if you copy-trade mechanically, treat this as the one judgment call in this report.
4 · Recommended $2,500 portfolio
| # | Ticker | Action | Limit | Contracts | Cost | % cap | Max payout | Edge ¢ | EV % | Tier / band |
|---|
| 1 | KXTECHRANKLISTAICODE-26JUL27-KIMI | BUY YES | 93¢ | 300 | $279.00 | 11.2% | $300.00 | +2 | +2.2% | HIGH · >90¢ |
| Cash reserve | | $2,221.00 | 88.8% | Held — no other candidate cleared the screens |
| Total | 300 | $2,500.00 | 100% | $300.00 | +2 | +2.2% | — |
EV cents / EV % are recorded for calibration only; under v3 they do not drive contract counts (the resolved record showed stated EV was anti-predictive). Size was set by conviction tier (HIGH cap $375) and visible book depth, then trimmed to 300 to stay inside resting liquidity.
Cluster exposure
| Cluster | Cost | % of capital | Cap (15%) |
|---|
| kimi-k3-code-arena | $279.00 | 11.2% | $375 ✓ |
Conviction exposure
| Tier | Cost deployed | % of capital | Per-pick cap |
|---|
| HIGH | $279.00 | 11.2% | $375 (15%) |
| MEDIUM | $0 | 0% | $175 (7%) |
| LOW | — never deploys (v3.3) | — | — |
Risk profile
- Worst case: −$279.00 (−11.2% of capital) if any non-Kimi model tops LM Code Arena at Mon Jul 27 10:00 ET. Single-cluster, single-event exposure; a loss cannot flip more than 11.2% of the book.
- Best case / most likely (~95%): +$21.00 gross (~+$19.50 net of ~0.5¢/contract fees), settling within 3 days.
- Concentration: one thesis, deliberately — the alternative candidates all failed screens that exist because their archetypes lost money in the resolved record. 88.8% cash is the position.
Execution notes
- Rest a limit at 92 first for the ~17 shown; step to 93 for the remainder. Do not pay 94+ — at 94 the modeled edge (95) is inside fee noise. If the book thins before filling, take the partial and stop.
- Invalidators — cancel/exit if: (a) any Arena changelog entry touching kimi-k3 or Code Arena methodology before Jul 27; (b) a frontier code-model launch announcement (OpenAI/Anthropic/Google) before the weekend; (c) kimi-k3's displayed score drops below ~1650 or claude-fable-5 rises above ~1655 (gap < ~15 pts) on the live board.
- Watchlist triggers for held cash: Challenger July report (Aug 6, 07:30 ET) — if a next-month KXCHAICUTS market lists near 50¢ with real book depth, re-underwrite (this month's +27¢-modeled-edge version was band-rejected, and the reject is shadow-tracked); Ornn dashboard data becoming programmatically readable would unlock the entire GPU monthly-average family, where 40–50¢ strikes with 7/31 days unfixed are near-mechanical if the month-to-date series is known; a Gemini 3.5 Pro official date announcement would make the KXGEMINI ladder tradable in both directions.
5 · What I rejected and why
Every Stage-2 survivor below is machine-logged in picks.json and shadow-tracked to settlement — if the rejects outperform the pick, the screen adds nothing, and that finding feeds the next methodology revision. Rejects cut specifically by the Stage-3.5 entry-band screen are marked [BAND].
The one that hurt to cut
- KXCHAICUTS-26AUG06-T1 YES @ 47¢ — model_prob 74 — [BAND] coin-flip band without HIGH conviction. AI has been the #1 reason in Challenger's job-cut table for four consecutive months (June: 14,029 of 45,849 cuts, 31%), July's announcement flow is light (~13.5k tracked; largest single event Microsoft's AI-cited 4,664) with no visible competing mega-driver, and "AI-washing" biases attribution toward YES. Modeled edge +27¢ — the largest found in this audit. It is still a reject three ways: conviction is MEDIUM, not HIGH (the report is unpublished and Challenger's classification is unobservable until Aug 6 — a live way to be wrong); 35–60¢ band requires HIGH; and the book is a shell (7 contracts at the 47¢ ask, then 52¢; 24h volume 0). This is exactly the medium-conviction coin-flip v3 exists to refuse. If the shadow-track shows these winning, the band rule — not this call — is what should change.
Resolver-risk rejects (Stage 2.5)
- KXCHINAAI-26JUL31-ALIB YES @ 75¢ — model_prob 77 — resolver mechanics unpriceable (deductions > 15 pts). The trade only works via a mechanism I cannot verify: the top verifiable Chinese model on Arena text by score is Moonshot's kimi-k3 (1486±10, #10) — not Alibaba — after establishing that muse-spark-1.1 is Meta's, not Alibaba's (Arena changelog attributes it to ai.meta.com; one aggregator mislabels it). Alibaba's claim to #1 Chinese rides on qwen3.7-max-preview, a low-vote preview whose wide CI flatters its Rank-UB — and that advantage mechanically decays as votes accumulate toward Jul 31. Deductions: kimi-k3 vote-flip −8, UB-decay −10, new Chinese release −5. A 93–94¢ weekly (Jul 27) says the resolver currently sees Alibaba on top, but I can't price the decay rate; per Stage 2.5 that's a reject, not a discount.
- KXTOPMODEL-26JUL27-CLAUT YES @ 92¢ — model_prob 90 — unreconciled resolver view; no edge. The public text board shows fable-5 #1, yet three straight weeklies finalized for opus-4-6-thinking. When I can't reproduce the resolver's ranking from the named source, I don't own the risk at 92¢ — this is the archetype of the program's worst historical loss. (Taking YES would also bet against the feed's live NO on the Jul 31 monthly.)
No-edge rejects (market already right)
- KXCODEAI-26JUL31-CHAT YES @ 92¢ fillable — model_prob 89. OpenAI holds #1 and #2 on Datacurve DeepSWE (GPT-5.6 Sol 0.727, Terra 0.696; Kimi K3 0.675), so only a new/re-scored third-party eval flips it — but Datacurve demonstrably evaluates fast (kimi-k3 already scored) and just shipped v1.1, which re-scored Claude downward after a benchmark-loophole finding. ~10–12% add-risk over 7 days ≈ the price. Shown ask 88 is 3 contracts deep; real size costs 92–93. No edge.
- KXMATHAI-26JUL31-CLAU YES @ 96¢ — model_prob 92. Claude holds #1/#3/#5/#6/#9 on the Arena math board (fable-5 1543±21) — but the board runs on tiny vote counts (761 votes at #1; #2 gemini-3.6-flash 1519±38 on 226 votes), so a few hundred adverse votes can reorder it. 92 < 96: negative edge at the ask.
- KXFABLEDISABLED-27-26AUG31 NO @ 96¢ — model_prob 94. Re-imposition of US access restrictions within 5 weeks is unlikely — but the Jun 12 export-control order proved the hazard is real, Commerce reversed itself once already this summer, and 94 < 96. Negative edge, v24 = 0.
- KXANTHROPICRESCIND-26JUN-26AUG01 NO @ 95¢ — model_prob 93. A DoD/White-House rescission of the Anthropic supply-chain designation inside 8 days is priced at 5–11¢; my ~7% is inside that. 6¢ spread, v24 = 40. Nothing to do.
- KXCLAUDE-NXTMYTH-26SEP01 NO @ 70¢ — model_prob 70. Next Mythos-tier (Mythos/Fable-branded, ex-Fable 5) release before Sep 1: no public evidence either way; Anthropic just shipped Fable 5 + Opus 4.8 + Sonnet 5, and the Oct/Nov legs (73–90¢) imply the market expects it in Sep–Oct. My number equals the price. 5¢ spread, v24 = 62.
Liquidity / slippage rejects
- KXFRONTIER-FRON-26AUG01 NO @ 93¢ — model_prob 96 — slippage kill. One verified FrontierMath open-problem solve exists (Mar 24, GPT-5.4 Pro, Epoch-confirmed); no second solve in 4 months, and Epoch verification latency alone shields an 8-day window. The trade is right — the book isn't: 1.25 contracts at 93¢, then nothing until 97¢, where the edge is gone. Stage-3 ≥3¢-slippage rule.
Entry-band screen cuts [BAND]
- KXGEMINI-GEMI35P-26AUG16 NO @ 36¢ — model_prob 43. Gemini 3.5 Pro has slipped twice (June → Jul 17 → "August window"), with Bloomberg reporting a scrapped-and-rebuilt base model — I lean under-65% for release-before-Aug-16 vs the market's 64–66¢, but a 36¢ NO entry is the coin-flip band without HIGH conviction. Reject by rule; also v24 = 208.
- KXB200MS-26JUL-6.500 / KXH200MS-26JUL-4.500 @ 40–41¢, KXRTX5090MS-26JUL-0.500 @ 49–50¢ — no model_prob formed. July monthly averages with 24/31 days already fixed are the closest thing to free money on this board if you can read the resolver — but all resolve off the Ornn hourly index, whose dashboard is JS-rendered and returned no data to any fetch attempted (4 endpoints probed). Without the month-to-date series these are coin-flips I can't underwrite, and the band requires HIGH + ≥15¢. Watchlist: whoever gets Ornn data programmatically owns this family.
- Sub-35¢ tails (KXCODEAI-KIMI at 2–3¢, KXTOPMODEL-CLAU7 at 0–1¢, KXANTHROPICRESCIND-AUG01 YES at 5–11¢, etc.): hard reject, no exceptions, per v3.2.
6 · Sources