# Kalshi AI — Mispricing Audit — 2026-07-07

Date: 2026-07-07
Source: https://kalshi-audits.pages.dev/2026-07-07-kalshi-ai-mispricing-audit-r2

---
# Kalshi AI — Mispricing Audit

Report date 2026-07-07 · Horizon 45 days · Capital $1,000 · Methodology **v3.1** (conviction-weighted, EV-agnostic sizing; resolver-priced probabilities) · Picks in this report count toward the public track record and are cleared for auto-copy.

## Headline: zero picks — 100% cash CASH IS THE POSITION

**Fifteen AI-theme markets passed the liquidity screens; ten survived to full diligence; none cleared the v3 conviction and entry-band gates.** The 60–90¢ favorite band — the only band where this program has historically made money — contained three candidates, and each failed on the merits: one had already run 50¢ on a single-source launch leak, one carries an unpriceable resolver risk (will Arena re-add claude-fable-5?), and one crashed 25¢ in 72 hours on resolver information that is not publicly verifiable. Every remaining "edge" my models found lives in sub-35¢ tails — exactly the band the resolved v1/v2 record shows going ~0-for. Under v3 rules the correct book this week is no book. All ten diligenced candidates are logged below as shadow-tracked rejects.

## 1. How this was researched

**Mode: theme-match.** Kalshi has no `AI` category (categories present: Sports, Crypto, Elections, Climate and Weather, Entertainment, Financials, Economics, Mentions, Politics, Commodities, Science and Technology, Companies, World, Social, Health, Transportation), so markets were selected by keyword across all categories. Search terms over event and market titles: _openai, ai, artificial intelligence, anthropic, claude, gpt, chatgpt, gemini, grok, xai, nvidia, deepseek, mistral, agi, llm, superintelligence, sam altman, fable, ai model_, plus a second sweep for _superintel, altman, data center, chip, semiconductor, robot, compute, frontier, model, mythos, agent_. (The term "llama" was dropped after it matched ATP tennis player Pablo Llamas Ruiz.) This returned ~400 active markets closing within 45 days, dominated by NVIDIA GPU compute-price ladders (~300 legs) and AI release/leaderboard markets.

Stage-1 screens (volume24h ≥ 1,500, spread ≤ 5¢, not priced ≥95¢/≤5¢, no Mentions markets, no already-published feed picks) cut this to 7 strict passes plus 3 near-misses carried into diligence. For each survivor I pulled full rules text (mirror DB + Kalshi public API `rules_primary/secondary`), 14-day price history from 5-minute snapshots, the live orderbook, and primary-source news. Four tickers from the existing feed (KXGPT-OPEN-26JUL09, KXMATHAI-26JUL31-GEMI, KXTOPMODEL-26JUL31-CLAUF, KXFABLERESTORE-27-26JUL17) were excluded from candidacy up front.

**The week's context, in one paragraph (it drives every verdict below):** Claude Fable 5 launched Jun 9 and topped every Arena leaderboard (~1525), was suspended worldwide Jun 12 under a US export-control directive, and was removed from Arena's leaderboards. The directive was lifted Jun 30 and Anthropic restored Fable 5 globally on Jul 1 — but Arena has _not_ re-added it (daily changelog silent through Jul 7). Meanwhile OpenAI previewed GPT-5.6 (Sol/Terra/Luna) on Jun 26 to ~20 trusted partners behind a government safety review, with a leaked internal GA target of Jul 7–9 — and Kalshi's GPT ladder puts ~75% on GA landing exactly Jul 9. Google's Gemini 3.5 Pro is priced ~80% to ship before Jul 31. Nearly every liquid AI market this week is a derivative of these three unresolved events.

## 2. Markets reviewed

| Ticker | Question | YES bid/ask | Vol 24h | OI | Verdict |
|---|---|---|---|---|---|
| KXGPT-OPENB-26JUL10 | GPT-5.6 released before Jul 10 | 82 / 83 | 17,332 | 35,652 | REJECT — already moved, leak-grade |
| KXLLM1-26JUL31-A | Anthropic top-ranked LLM Jul 31 | 81 / 82 | 10,445 | 34,650 | REJECT — negative edge at ask |
| KXLLM1-26JUL31-OAI | OpenAI top-ranked LLM Jul 31 | 6 / 7 | 8,309 | 31,175 | REJECT — no edge either side |
| KXTOPMODEL-26JUL13-CLAUT | opus-4-6-thinking top model Jul 13 | 92 / 93 | 3,823 | 5,448 | REJECT — resolver risk > cap |
| KXTOPMODEL-26JUL13-CLAUF | claude-fable-5 top model Jul 13 | 6 / 7 | 3,598 | 3,988 | REJECT — tail, lottery withheld |
| KXMATHAI-26JUL31-CLAU | Claude #1 Arena Math Jul 31 | 60 / 61 | 3,131 | 7,232 | REJECT — resolver unverifiable |
| KXCODEAI-26JUL31-CLAU | Claude top on Datacurve DeepSWE Jul 31 | 49 / 51 | 2,776 | 15,276 | REJECT — coin-flip band |
| KXCODEAI-26JUL31-CHAT | ChatGPT top on Datacurve DeepSWE Jul 31 | 50 / 52 | 487 | 5,435 | REJECT — coin-flip band, illiquid |
| KXGPT-OPENB-26JUL15 | GPT-5.6 released before Jul 15 | 88 / 89 | 320 | 14,489 | REJECT — zero edge, illiquid |
| KXLLM1-26JUL31-GOOG | Google top-ranked LLM Jul 31 | 13 / 14 | 1,254 | 19,158 | REJECT — two-leg parlay tail |
| KXGPT-OPENB-26JUL31 | GPT-5.6 released before Jul 31 | 96 / 97 | 1,425 | 23,308 | Stage-1 skip (≥95¢, fully priced) |
| KXCLAUDE-NXTMYTH-26AUG01 | Next Mythos-class model before Aug 1 | 4 / 5 | 2,036 | 1,771 | Stage-1 skip (≤5¢, fully priced) |
| KXCHAICUTS-26AUG06-T1 | AI is #1 Challenger job-cut reason | 43 / 48 | 42 | 708 | Stage-1 skip (volume, spread) |
| KXANTHROPICRESCIND-26JUN-26AUG01 | DoD/WH rescind announcement re Anthropic | 33 / 34 | 16 | 1,497 | Stage-1 skip (volume) |
| KX*WS / KX*MS / KX*MON ladders | NVIDIA GPU compute-price ladders (~300 legs) | — | — | — | Stage-1 skip: every liquid leg ≥95¢ or ≤5¢; mid-curve legs <1,500 vol or >5¢ spread |

## 3. Top picks

**None.** This is a deliberate output, not an omission. The v3 methodology sizes only what survives (a) a conviction tier that is defensible from source quality and rules clarity, (b) the entry-band screen that restricts 35–60¢ coin-flips and bans sub-35¢ tails outside a single 2% lottery slot, and (c) resolver-risk deductions capped at ~15 points. Ten candidates went through full diligence; the section below shows each one's resolver profile and exactly which gate it failed. The strongest temptations — and why they were declined:

- **Best favorite-band candidate:** KXGPT-OPENB-26JUL10 YES @ 83¢. Declined: the price ran 33¢→83¢ in ten days on a leaked "internal target window," the government safety-review dependency is live, and the rules require _public_ release ("outside of a closed beta") — a partner/enterprise expansion on Jul 9 would strand YES holders. My estimate: 75%, below the ask.
- **Best tail:** KXTOPMODEL-26JUL13-CLAUF YES @ 7¢ (~+6¢ modeled edge if Arena re-adds Fable 5 by Jul 13). Declined the lottery slot: after 7 consecutive days of Arena's changelog updating daily without a re-add, the thesis is a pure guess about a third party's operational timing — the exact epistemic posture that produced the historical tail losses.

## 4. Recommended $1,000 portfolio

| Position | Conviction | Entry band | Cost | % of capital | Max payout |
|---|---|---|---|---|---|
| _(no positions)_ | — | — | $0 | 0% | $0 |
| **Cash reserve** | — | — | **$1,000** | **100%** | $1,000 |

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

| Cluster | Thesis | Cost | % of capital | Cap |
|---|---|---|---|---|
| gpt-56-release | GPT-5.6 GA timing | $0 | 0% | 15% |
| arena-fable-readd | Arena re-lists claude-fable-5 | $0 | 0% | 15% |
| arena-top-slot-jul31 | Who holds Arena #1 on Jul 31 | $0 | 0% | 15% |
| arena-math-jul31 | Arena Math category #1 | $0 | 0% | 15% |
| deepswe-top-jul31 | Datacurve DeepSWE #1 | $0 | 0% | 15% |

Note the cluster structure itself was a finding: seven of the ten diligenced candidates are derivatives of just two world-states (Arena's treatment of Fable 5, and GPT-5.6's launch). Any multi-pick book this week would have been dangerously correlated even before sizing — the same shape as the v1 blow-up.

### Conviction exposure

| Tier | Per-pick cap | Deployed |
|---|---|---|
| HIGH | $150 (15%) | $0 |
| MEDIUM | $70 (7%) | $0 |
| LOW | $30 (3%) | $0 |

### Risk profile & execution notes

- **Worst case / best case / most likely:** $1,000 / $1,000 / $1,000. No market risk is carried out of this run. Blended EV: 0%; dollar edge: $0. (EV figures on rejects below are recorded for calibration only and did not drive any sizing decision, per v3.)
- **Watchlist trigger 1 — Arena re-adds claude-fable-5** (check [the changelog](https://arena.ai/blog/leaderboard-changelog/) daily): the entire KXTOPMODEL/KXLLM1 complex reprices within hours. A re-add _announcement_ before Jul 13 makes the CLAUF legs the trade; a continued freeze into late July puts the live-feed KXTOPMODEL-26JUL31-CLAUF YES pick from the prior run at risk.
- **Watchlist trigger 2 — GPT-5.6 GA language.** If Jul 9 brings "available to Pro/Plus subscribers" the JUL10 leg resolves YES and the JUL15/JUL31 rungs snap to 99; if it brings "expanding to additional enterprise partners," the JUL10 leg is a NO and the whole ladder reprices violently. The distinction is the resolver's, not the press's — read OpenAI's own wording, not headlines.
- **Watchlist trigger 3 — Datacurve evaluates GPT-5.6.** A published DeepSWE score above/below fable-5's 70% pass@1 turns KXCODEAI-26JUL31 from a coin flip into a near-certainty; the leftover leg will briefly lag the news.
- **Watchlist trigger 4 — Gemini 3.5 Pro launch** (market-implied ~70% before Jul 24). Its first-week Arena score vs the 1494–1504 Anthropic wall resolves whether KXLLM1-26JUL31-A at 81 is cheap or rich.

## 5. What I rejected and why

All ten rejects are machine-logged in `picks.json` and shadow-tracked to settlement — if the rejects outperform the (empty) book, the screens added nothing and that finding will be reported. Resolver-risk deductions follow the v3.1 Stage-2.5 format. Six of the ten were cut wholly or partly by the Stage-3.5 entry-band screen (three tails, two coin-flips, one deep favorite).

### KXGPT-OPENB-26JUL10 — GPT-5.6 before Jul 10 · would-be side: NO @ 18¢ · my prob 25% vs market 18%

**Resolver:** OpenAI's own release action, judged against Kalshi's rule: "A model called GPT-5.6 or greater… Release must be to the public, outside of a closed beta, though limiting it to a high-cost subscription tier is acceptable."
**Priced risks (on the NO side):** −45 pts leaked internal GA window Jul 7–9 corroborated by shipping artifacts (reasoning-slider controls in Codex builds), −15 pts competitive motive to launch during Anthropic's Jul 7 Fable-5 subscription-to-credits transition. Residual NO value comes from launch-slip base rates and the live government-review dependency ("broad access plausible mid-July at the earliest" — emergent.sh).

- Preview was Jun 26 to ~20 trusted partner orgs behind a government safety review — that is a closed beta and did not resolve the ladder (JUL07/JUL08 rungs died at 0–4¢).
- The ladder prices GA on Jul 9 _specifically_ at ~75% (JUL09 rung 7/8 vs JUL10 rung 82/83). Ten days ago the JUL10 rung traded 33¢; the move is complete. Todays's intraday range was 57–87 — this is a rumor casino now.
- **Why rejected:** the positive-EV side (NO) is a sub-35¢ tail fading a dated, partially-corroborated leak — LOW conviction by construction, and the lottery slot was not spent here. Entry context: YES 82/83, ~330 contracts fillable near my would-be NO limit, vol 17.3k/24h, OI 35.6k.

### KXTOPMODEL-26JUL13-CLAUT — opus-4-6-thinking top model Jul 13 · would-be side: YES @ 93¢ · my prob 84% vs market 93%

**Resolver:** arena.ai Text leaderboard snapshot Jul 13, 10:00 AM ET, with the "Remove Style Control" toggle checked; ranking by Rank (UB), ties broken by Arena Score, then votes, then release date (quoted from `rules_secondary`).
**Priced risks:** −12 pts Arena re-adds claude-fable-5 (frozen at 1509±9 / 4,350 votes, it retakes #1 on the score tie-break the moment it's listed), −3 pts GPT-5.6 GA Jul 9 + fast-tracked listing beats 1504 in four days, −1 pt intra-Anthropic drift (opus-4-7-thinking at 1502 is 2 pts behind, but both scores have been static). Total deductions ≈ 16 pts — **over the 15-pt Stage-2.5 cap**.

- The market prices the fable-5 re-add at 6–7¢ (the CLAUF leg) and everything else at ~0 — every other leg of this mutually-exclusive event sits at 0/1¢. The whole contract is a bet on one unannounced ops decision by Arena.
- Evidence cuts both ways: Arena's removal reason (export directive + Anthropic suspension) is fully mooted since Jun 30/Jul 1, and Arena publicly called Fable 5 "the most powerful model we've ever tested"; but the changelog has updated daily through Jul 7 (muse-image added that day) with no re-add.
- **Why rejected:** unpriceable resolver risk over the cap, and at a 93¢ ask my 84% estimate implies negative edge anyway. Entry context: 92/93, ~11,800 contracts at ask, vol 3.8k, OI 5.4k.

### KXTOPMODEL-26JUL13-CLAUF — claude-fable-5 top model Jul 13 · would-be side: YES @ 7¢ · my prob 13% vs market 7%

**Resolver:** same as above. This is the other side of the same coin: YES requires Arena to re-add fable-5 _and_ have it ranked by the Jul 13 snapshot; with its frozen 1509 score and the score tie-break, listing ≈ winning.

- Modeled +6¢ on a 7¢ contract (+86% EV) — and that is precisely why v3 exists: this is a LOW-conviction, sub-35¢ tail whose "edge" is an unfalsifiable guess about a third party's internal timing. The one permitted 2% lottery slot was withheld because no evidence distinguishes my 13% from the market's 7%.
- Also note: the live feed already carries this thesis via KXTOPMODEL-26JUL31-CLAUF YES from a prior run; adding the Jul 13 rung would stack correlated exposure on the same world-state.

### KXLLM1-26JUL31-A — Anthropic has top-ranked LLM Jul 31 · would-be side: NO @ 19¢ · my prob 29% vs market 19%

**Resolver:** Arena text leaderboard, Rank (UB) with score/votes/release-date tie-breaks, Jul 31 10:00 ET. Brand-level: _any_ Anthropic model at #1 counts — and Anthropic currently holds the top four listed slots (opus-4-6-thinking 1504, 4-7-thinking 1502, 4-6 1499, 4-7 1494; fable-5 delisted). A fable-5 re-add only strengthens YES.
**Priced risks (against YES):** −22 pts Gemini 3.5 Pro ships (its own Kalshi ladder implies ~80% before Jul 31) and clears the 1504 wall (~30% conditional — gemini-3.1-pro-preview sits 18 pts back at 1486), −8 pts GPT-5.6 tops text Arena within three weeks of GA (gpt-5.5-high is 23 pts back; 5.6 is agent/coding-positioned).

- That arithmetic makes YES worth ~69 vs an 82 ask (reject) and NO worth ~29 vs a 19 cost (+10¢). But the entire NO edge hangs on my ~30% "Gemini clears the wall" conditional, for which I have no grounding beyond version-bump base rates — if it's 15%, the market is exactly right.
- **Why rejected:** NO is a sub-35¢ tail with an undefendable core assumption; YES has negative edge. Neither side is a v3 trade. Entry: 81/82, NO fillable ~2,900 contracts ≤21¢, vol 10.4k, OI 34.7k. Contract drifted 91→81 over the week as Gemini-3.5 expectations firmed — the market is already migrating toward my number, which is further evidence the residual is noise.

### KXMATHAI-26JUL31-CLAU — Claude #1 on Arena Math Jul 31 · would-be side: YES @ 61¢ · no committed probability

**Resolver:** "Arena AI Text Arena Leaderboard (Math)" category view, Jul 31 10:00 ET, publisher tie-breaks.
**Unpriceable:** I could not verify the current Math-category #1 from any reachable source — Arena's category view is client-rendered, the daily-snapshot GitHub mirror carries no category data (and its overall file has been frozen since Jul 1, still showing delisted fable-5 at #1), and every third-party aggregator I checked was weeks stale.

- The contract spiked to ~87 on Jul 3–4 then crashed to 56–61 by Jul 7 while the GEMI leg rose to 28/30 — a 25¢+ move on information I cannot see. Both the "resolver state unverifiable" and "price already moved in 48h" screens fire.
- **Why rejected:** Stage-2.5 rule 4 — a definitional/resolver state you cannot verify is a reject regardless of apparent value. Buying the favorite side of someone else's information is how the 35–60¢ band lost money historically. Entry: 60/61, ~1,600 at ask, vol 3.1k, OI 7.2k.

### KXCODEAI-26JUL31-CLAU / -CHAT — Datacurve DeepSWE #1 on Jul 31 · would-be sides: YES @ 51¢ / YES @ 52¢ · my probs 52% / 48%

**Resolver:** deepswe.datacurve.ai leaderboard, Jul 31 10:00 ET. Current official board (v1.1, updated Jul 1): claude-fable-5 70%±4 (#1, still listed despite the June suspension; 73 of 2,260 trials incomplete), gpt-5.5 67%±6 (#2), claude-opus-4.8 59% (#3).
**Priced risks (on CLAU YES):** −40 pts Datacurve evaluates GPT-5.6 after its ~Jul 9 GA (they re-ran the board Jul 1 and evaluate frontier launches promptly) and it clears 70% — GPT-5.6 is explicitly agentic-coding-positioned (Sol Ultra: 91.9% TerminalBench 2.1) and GPT-5.5 already sits 3 pts back, −5 pts fable-5's incomplete trials complete/re-run and move its score.

- Net: ~52% for CLAU — indistinguishable from the 49/51 market. CLAU crashed 72→49 in 24h as the GPT-5.6 launch window firmed; the two legs now trade as the coin flip they are (CLAU 51 + CHAT 52 + rest ≈ 0).
- **Why rejected:** 35–60¢ coin-flip band requires HIGH conviction plus a ≥15¢ edge; this has neither, on either leg. CHAT additionally fails the volume screen (487/24h). Entry (CLAU): 49/51, ~660 at ask, vol 2.8k, OI 15.3k.

### Remaining logged rejects (summary)

- KXGPT-OPENB-26JUL15 YES @ 89 — my 88%: zero edge, and 320 contracts/24h fails the liquidity screen despite 14.5k OI. Cluster gpt-56-release.
- KXLLM1-26JUL31-GOOG YES @ 14 — my 22%: positive EV on paper but a sub-35¢ two-leg parlay (Gemini 3.5 Pro ships AND clears the Anthropic wall); also 1,254/24h volume. Cluster arena-top-slot-jul31.
- KXLLM1-26JUL31-OAI NO @ 94 effective — my 92%: negative edge in the >90¢ deep-favorite band; GPT-5.6 topping text Arena is unlikely but not 6-cents unlikely. Cluster arena-top-slot-jul31.

## 6. Sources

- Kalshi live market data & rules: [KXGPT-OPENB-26JUL10](https://api.elections.kalshi.com/trade-api/v2/markets/KXGPT-OPENB-26JUL10), [KXTOPMODEL-26JUL13-CLAUT](https://api.elections.kalshi.com/trade-api/v2/markets/KXTOPMODEL-26JUL13-CLAUT), [KXLLM1-26JUL31-A](https://api.elections.kalshi.com/trade-api/v2/markets/KXLLM1-26JUL31-A), [KXMATHAI-26JUL31-CLAU](https://api.elections.kalshi.com/trade-api/v2/markets/KXMATHAI-26JUL31-CLAU), [KXCODEAI-26JUL31-CLAU](https://api.elections.kalshi.com/trade-api/v2/markets/KXCODEAI-26JUL31-CLAU) (+ orderbooks); read-only mirror DB for snapshots/history.
- [OpenAI — Previewing GPT-5.6 Sol](https://openai.com/index/previewing-gpt-5-6-sol/) (Jun 26 preview; ~20 partner orgs; government safety review); [GPT-5.6 Preview System Card](https://deploymentsafety.openai.com/gpt-5-6-preview); [Wikipedia — GPT-5.6](https://en.wikipedia.org/wiki/GPT-5.6) (Sol/Terra/Luna, TerminalBench 2.1 scores).
- [TradingKey (Jul 2026)](https://www.tradingkey.com/analysis/stocks/us-stocks/262016012-opanai-ipo-gpt-ai-anthropic-ipo-fable5-tradingkey) — "internally locked target window between July 7 and July 9"; Fable 5 subscription→pay-as-you-go transition Jul 7. [TechTimes (Jul 6)](https://www.techtimes.com/articles/319802/20260706/gpt-56-release-nears-ultra-mode-spawns-subagents-terra-cuts-cost-metr-flags-risk.htm); [emergent.sh GA-timing analysis](https://emergent.sh/news/gpt-5-6-release-date); [explainx.ai availability analysis](https://explainx.ai/blog/when-will-gpt-5-6-sol-terra-luna-be-available-everyone-2026).
- [Anthropic — Redeploying Claude Fable 5](https://www.anthropic.com/news/redeploying-fable-5) (export controls lifted Jun 30; global restoration Jul 1; usage-credit transition Jul 7). [Fable 5 suspension timeline](https://ssntpl.com/blog-claude-fable-5-access-suspended/) (Jun 12 export-control directive). [PCWorld on subscriber reaction](https://www.pcworld.com/article/3181897/claude-subscribers-are-furious-over-fables-new-restrictions.html).
- [Arena.ai on X](https://x.com/arena/status/2065620808773611997) — Fable 5 removal announcement ("#1 across Agent, Text, and Code Arena"). [Arena leaderboard changelog](https://arena.ai/blog/leaderboard-changelog/) (no re-add through Jul 7). [Arena text leaderboard](https://arena.ai/leaderboard/text) + [oolong-tea-2026/arena-ai-leaderboards](https://github.com/oolong-tea-2026/arena-ai-leaderboards) daily JSON snapshots (top-12 scores; snapshot staleness itself documented above).
- [Datacurve DeepSWE leaderboard](https://deepswe.datacurve.ai/) (v1.1, Jul 1: fable-5 70% #1) and [DeepSWE v1.1 blog](https://deepswe.datacurve.ai/blog/deepswe-v1-1) (isolated verification; incomplete fable trials footnote); [VentureBeat on DeepSWE v1](https://venturebeat.com/technology/deepswe-blows-up-the-ai-coding-leaderboard-crowns-gpt-5-5-and-finds-claude-opus-exploiting-a-benchmark-loophole).

Data: Kalshi public trade API (live orderbooks, rules) and a read-only mirror of Kalshi market snapshots (5-minute bars); news via web search with primary sources preferred. All probabilities are subjective estimates committed before knowing outcomes; every contract discussed can resolve to zero. Rejects are shadow-tracked to settlement for selection-screen calibration. This report is research, not investment advice. Methodology v3.1 — conviction-weighted sizing, entry-band screens, 15% cluster caps, EV-agnostic sizing, resolver-priced probabilities.