Kalshi AI — Mispricing Audit (v2, experimental)
Report date: 2026-06-09 · Horizon: 45 days · Suggested capital: $1,000 · Author: automated deep-research audit
1. How this was researched
Mode: theme-match. Kalshi has no AI category (categories present: Sports, Crypto, Elections, Economics, Politics, Science and Technology, Companies, …), so markets were selected by keyword across all categories. SQL searched trading_events.title and trading_markets.title for: openai, anthropic, claude, " ai ", artificial intelligence, chatgpt, gpt-, gemini, grok, xai, nvidia, deepseek, llama, superintelligen-, agi, llm, sam altman, sora, waymo — 215 raw rows, hand-curated down to ~20 genuinely AI-linked event families after dropping sports false-positives and "what will X say" mention markets.
The dominant fact of the day: Anthropic released Claude Fable 5 (its first public "Mythos-class" model, with a restricted sibling Mythos 5 via Project Glasswing) on the morning of June 9, 2026 — hours before this audit ran. Every AI-leaderboard market on Kalshi repriced violently intraday: KXTOPMODEL-CLAUT crashed 96→61, KXAISPIKE-1525 spiked 15→31, KXCODEAI flipped from ChatGPT-favorite to Claude-favorite. This audit is largely an exercise in separating launch-day hype from leaderboard mechanics, using the actual resolution sources (arena.ai, deepswe.datacurve.ai) as ground truth.
Pipeline: read-only DB mirror for the universe + 14-day intraday history (market_snapshots, 5-min bars) → live Kalshi orderbooks for entry context → primary-source web diligence (live leaderboards, vendor announcements, Polymarket cross-check). Stage-1 filters: volume24h ≥ 1,500, spread ≤ 5¢, not fully priced (5¢ < price < 95¢), no mention markets, no already-published picks (none of the 10 existing picks are AI markets; none re-appear below).
2. Markets reviewed
| Market | Question | YES bid/ask | Vol 24h | Outcome of review |
|---|---|---|---|---|
| KXTOPMODEL-26JUN13-CLAUT | claude-opus-4-6-thinking top Arena model Jun 13 | 60 / 61 | 16,054 | PICK 2 — BUY YES |
| KXCODEAI-26JUN30-CLAU | Claude top of Datacurve DeepSWE Jun 30 | 57 / 61 | 6,660 | PICK 1 — BUY NO |
| KXAISPIKE-26B-1550 | Any model ≥1550 on LMSYS before Jul 1 | 11 / 12 | 8,890 | PICK 3 — BUY NO |
| KXLLM1-26JUN30-GOOG | Google top-ranked LLM Jun 30 | 9 / 10 | 2,334 | PICK 4 — BUY NO |
| KXCODEAI-26JUN30-CHAT | ChatGPT top of DeepSWE Jun 30 | 41 / 46 | 1,142 | REJECT — thin book, duplicate thesis |
| KXAISPIKE-26B-1525 | Any model ≥1525 on LMSYS before Jul 1 | 28 / 33 | 10,503 | REJECT — efficiently repriced |
| KXLLM1-26JUN30-A | Anthropic top-ranked LLM Jun 30 | 90 / 92 | 10,318 | REJECT — ≤2¢ edge vs Polymarket |
| KXCLAUDE-MYTH-26JUL01 | Anthropic releases "Mythos" before Jul 1 | 2 / 3 | 181,587 | REJECT — definitional trap |
| KXFRONTIER-FRON-26JUL01 | AI solves a Frontier Math open problem before Jul 1 | 22 / 29 | 44 | REJECT — unfillable |
| KXANTHROPICRESCIND-26JUN | DoD/WH rescinds Anthropic supply-chain designation | 18 / 25 | 0 | REJECT — dead book |
| KXCLAUDE-CLAU-26JUL01 | Anthropic releases Claude 5 before Jul 1 | 97 / 100 | 14,351 | Filtered — fully priced (Fable 5 counted) |
| KXLLM1-26JUN13-ANTH | Anthropic best AI this week | 97 / 99 | 5,797 | Filtered — fully priced |
| KXGPT-OPEN-26JUL01 | GPT-6 before Jul 1 | 4 / 6 | 21 | Filtered — fully priced / illiquid |
| KXGROK-GROK5 / KXIPO* / KXAGICO / KXWAYMO | Grok 5, OpenAI/Anthropic/Waymo IPO, AGI markets | ≤2 | — | Filtered — fully priced |
| KX{A100,H100,H200,B200,RTX5090}{W,Q} | NVIDIA compute-price ladders (10 events) | — | <250 | Filtered — spreads up to 90¢, near-zero volume |
| KXFEDMENTION / KXMELANIAMENTION / KXHEARINGMENTION | "Will X say AI…" mention markets | — | — | Filtered — announcer-noise by rule |
3. Picks — detail and thesis
Pick 1 — KXCODEAI-26JUN30-CLAU · BUY NO @ 0.45 limit · HIGH CONVICTION
"Top Coding AI this month?" — resolves YES for Claude if Claude is the highest-ranked DeepSWE-score model on Datacurve DeepSWE on Jun 30, 2026, 10:00 AM ET.
Mispricing thesis. The market flipped this event from ChatGPT-favorite (79¢ on Jun 2) to Claude-favorite (57–61¢) on Fable 5 launch-day hype — without engaging with what the resolution source actually measures. Datacurve's DeepSWE benchmark was explicitly built to puncture Claude's coding-leaderboard advantage, and its current board isn't close.
Evidence.
- Live deepswe.datacurve.ai board (updated Jun 7, 2026): gpt-5.5 [xhigh] 70%±3, claude-opus-4.8 [max] 58%±2, gpt-5.4 56%±2, claude-opus-4.7 54%±5. GPT-5.5 leads the best Claude by 12 points; Fable 5 is not yet listed.
- DeepSWE's launch coverage (VentureBeat, May 26; The Neuron explainer) documents that SWE-Bench Pro containers ship the repo's full
.githistory — the gold-patch commit is in the container — and "most models ignore it. Claude does not." DeepSWE's verifiers cut false-accepts to 0.3% (vs 8.5%). Claude's headline SWE-Bench Pro scores are therefore structurally inflated on exactly the benchmark Anthropic quotes for Fable 5 (80.3%, per Inc.). - For Claude to win this market, Datacurve must add Fable 5 (likely — they evaluate actively, "16 of 19 possible models" run, all on mini-swe-agent) and Fable must leapfrog ~12 points on the loophole-resistant harness. Naively mapping Anthropic's claimed +11 SWE-Bench-Pro gain (Opus 4.8: 69.2% → Fable 5: 80.3%) onto DeepSWE lands Claude at ≈69–70 — a coin-flip at the boundary at best, before any haircut for the de-fanged
.gitassist. - OpenAI is not standing still either: a GPT-6-before-Jul-1 market trades at 4–6¢, and GPT-5.5's [xhigh] config already holds the top slot.
Tail risks (cleanest loss path). Fable 5's coding gain is real and transfers: it posts ≥71% on DeepSWE in the next refresh and holds it through Jun 30. Secondary: Datacurve doesn't re-run GPT-5.5's best config and an error-bar shuffle drops it. Also note "Claude" here means any Claude model — there is no partial credit for GPT staying close.
Numbers. 550 contracts × ~44¢ = $242 cost; max payout $550; EV ≈ +$88.
Liquidity / entry context (orderbook 19:05 UTC). YES 57 bid / 61 ask (NO buyable at 43); fillable NO depth: ~182 @43, ~217 @≤44, ~560 @≤45. Vol24h 6,660; OI 13,147. Work the limit at 43–45; do not chase above 45.
Price history (14d). CLAU leg: 9¢ (Jun 2) → 17 (Jun 4) → 23 (Jun 7) → 31 (Jun 8) → 46–68 intraday today on 6.7k volume — the entire move is anticipatory; the resolution source itself has not changed.
Cluster: fable-deepswe-coding
Pick 2 — KXTOPMODEL-26JUN13-CLAUT · BUY YES @ 0.62 limit · MEDIUM CONVICTION
"Top AI model this week?" — resolves YES if claude-opus-4-6-thinking is the top-ranked AI model on Jun 13, 2026, 10:00 AM ET (LMArena text leaderboard; legs are exact arena model strings).
Mispricing thesis. This market sat at 93–99 through Jun 8, then crashed to 50–62 today on 16k volume when Fable 5 dropped. The crash prices ~35–40% odds that the incumbent is dethroned within 4 days. But dethroning requires Fable 5 to be Arena-listed, vote-qualified, and ranked above 1504 — all before Jun 13, 10:00 AM ET. Listing mechanics make that slower and less certain than the panic implies.
Evidence.
- Live arena.ai text leaderboard (board update Jun 5): #1 claude-opus-4-6-thinking 1504±4, #2 claude-opus-4-7-thinking 1501±5, #5 muse-spark 1489, #8 claude-opus-4-8-thinking 1482±8. 365 models, 6.7M votes.
- Listing-lag precedent: Opus 4.8 released May 28 (Anthropic, GitHub changelog) and first appeared on the Jun 5 board — ~8 days. Fable 5 released Jun 9 (TechCrunch) leaves ≤4 days to listing + vote accumulation + a board refresh.
- Debut-rank precedent: Anthropic's last two flagship releases both failed to take Arena #1 — 4-7-thinking debuted at 1501 (below 1504), 4-8-thinking at 1482. Arena preference-ELO does not reward raw capability one-for-one; Fable 5's safety fallbacks (it blocks cyber/bio queries and serves Opus 4.8's answer instead — TechCrunch) are a further vote drag, even at the reported ≤5% deferral rate.
- Even if no refresh lands before Jun 13 10:00 ET, the market resolves on the standing board — where CLAUT is #1.
Tail risks (cleanest loss path). Fable 5 has been collecting Arena votes under a pseudonym (the muse-spark entry proves codename testing is live practice) and is revealed at #1 in a refresh before Jun 13. Second path: a routine refresh flips 4-7-thinking (1501±5) above 4-6-thinking (1504±4) — a 3-point gap inside the error bars (~10% on its own).
Numbers. 340 contracts × ~61.4¢ = $209; max payout $340; EV ≈ +$22.
Liquidity / entry context. YES 60/61; fillable ~210 @61, ~341 @≤62. Vol24h 16,054; OI 13,322. 4-day hold.
Price history (14d). 90 (Jun 6) → 96 (Jun 7) → 96 (Jun 8) → 50–96 today; a ≥20¢ single-day move on event news, volume-confirmed.
Cluster: fable-arena-debut
Pick 3 — KXAISPIKE-26B-1550 · BUY NO @ 0.90 limit · MEDIUM CONVICTION
"Will any AI model have a score of at least 1550 before Jul 1, 2026 on the LMSYS leaderboard?"
Mispricing thesis. 1550 requires a +46-point jump over the all-time Arena ceiling. The ceiling has been pinned at ~1500–1504 for three-plus months through four frontier launches (opus 4.7, opus 4.8, gpt-5.5-high, muse-spark — none moved it). Today's YES spike (5→11, intraday high 24, on 8.9k volume) is Fable-launch froth applied to a strike the launch almost certainly can't reach.
Evidence.
- Arena top score Jun 5: 1504 (arena.ai). March 2026 top: ~1504 (LocalAI Master archive) — net ceiling movement over the quarter: ~0 points.
- Anthropic flagship debuts: 4-7-thinking +(−3) vs ceiling, 4-8-thinking −22. Preference-ELO compresses hard at the top; a +46 debut would be unprecedented in this data.
- The sibling 1525 strike (a +21 jump) trades 28/33 — the market itself only gives ~30% to a far easier threshold.
Tail risks (cleanest loss path). The rules count any appearance before Jul 1 — a provisional, wide-error-bar Fable 5 debut score that briefly prints ≥1550 before settling lower would resolve YES. New-model provisional scores do run hot. This is the single risk that keeps the position medium-sized.
Numbers. 155 contracts × ~90¢ = $140; max payout $155; EV ≈ +$4.7.
Liquidity / entry context. NO buyable: ~360 @89, ~2,700 @≤90 (YES bids 11¢/$39.55, 10¢/$236). Vol24h 8,890; OI 30,121.
Price history (14d). 3–8¢ range May 29–Jun 8; today 5→24→11 on 8.9k volume.
Cluster: fable-arena-debut (shares Fable-debut exposure with Pick 2 — jointly capped)
Pick 4 — KXLLM1-26JUN30-GOOG · BUY NO @ 0.92 limit · LOW CONVICTION
"Best AI in Jun 2026?" — Google leg; resolves YES if Google has the top-ranked LLM on Jun 30, 2026.
Mispricing thesis. Google's best Arena entry (gemini-3.1-pro-preview, 1488) sits 16 points behind Anthropic's #1, behind three other Anthropic models, and now must also out-leapfrog an incoming Fable 5 — within 21 days. 9–10¢ is a point or two of leftover Gemini hope; the larger cross-market book prices it at 7%.
Evidence.
- arena.ai Jun 5: Anthropic holds ranks 1–4; Google's best is #6 (1488).
- Polymarket's equivalent market ($12.7M volume, resolves on lmarena.ai, style control off): Anthropic 91.1%, Google 7%, OpenAI 2.7%.
- This leg is also the only pick in the book that benefits from Fable 5 strength — partial directional offset to Picks 1–3.
Tail risks (cleanest loss path). Google ships full Gemini 3.1/3.2 before Jun 30 and it debuts ≥1505+ above all Anthropic entries (preview→full upgrades have historically added points). Note this same event would likely also kill Pick 2 — a real cross-cluster correlation, sized accordingly small.
Numbers. 110 contracts × ~91.5¢ = $101; max payout $110; EV ≈ +$1.7.
Liquidity / entry context. NO buyable: ~106 @91, ~23,900 @92 (YES bids 9¢/$9.57, 8¢/$1,904). Vol24h 2,334; OI 19,670.
Price history (14d). Google leg decayed 24 (May 29) → 14 (Jun 3) → 9 today — steady drift against Google, no single repricing event.
Cluster: arena-anthropic-dominance
4. Recommended $1,000 portfolio
| # | Market | Side | Limit | Contracts | Cost (≈) | Max payout | My P | EV ¢/ct | EV % | Conviction |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | KXCODEAI-26JUN30-CLAU | NO | 45 | 550 | $242 | $550 | 60% | +16 | +36% | HIGH |
| 2 | KXTOPMODEL-26JUN13-CLAUT | YES | 62 | 340 | $209 | $340 | 68% | +7 | +11% | MEDIUM |
| 3 | KXAISPIKE-26B-1550 | NO | 90 | 155 | $140 | $155 | 93% | +3 | +3.3% | MEDIUM |
| 4 | KXLLM1-26JUN30-GOOG | NO | 92 | 110 | $101 | $110 | 93% | +2 | +1.6% | LOW |
| Total deployed | $692 | $1,155 | Dollar edge ≈ +$117 | +16.9% | ||||||
| Cash reserve: $308 (30.8%) — above the ~10% floor; held back deliberately because all fillable depth at acceptable slippage is already consumed and the books are thin. | ||||||||||
Cluster exposure (cap: 35% of capital = $350 per cluster)
| Cluster | Underlying thesis | Picks | Cost | % of $1,000 | Cap check |
|---|---|---|---|---|---|
| fable-arena-debut | Fable 5 does NOT take over LMArena quickly/spectacularly | 2, 3 | $349 | 34.9% | ✔ at cap |
| fable-deepswe-coding | Fable 5 does NOT leapfrog GPT-5.5 on DeepSWE | 1 | $242 | 24.2% | ✔ |
| arena-anthropic-dominance | No non-Anthropic model takes Arena #1 by Jun 30 | 4 | $101 | 10.1% | ✔ |
Risk profile
- Honest concentration disclosure: clusters 1 and 2 are formally separate (different resolution sources, different mechanisms — Arena preference-ELO vs agentic-coding pass rate; opus-4-6-thinking is #1 on Arena while scoring 28% on DeepSWE) but share one actor: all three picks are short "Fable 5 exceptionalism." Joint worst case — Fable lists on Arena ≤Jun 13 at #1, prints ≥1550, and tops DeepSWE — loses Picks 1–3: −$591 (Pick 4 wins +$9; net ≈ −$582, −58% of capital). Estimated joint probability ≈ 8–10%.
- Worst case (everything loses incl. a Gemini shock): −$692. Best case (all win): +$463. Most likely outcome (independence approximation, P = .60/.68/.93/.93): Picks 3+4 win, Pick 1 wins, Pick 2 a coin-leaning-win → modal P&L ≈ +$250 to +$460; probability all four win ≈ 35%.
- Timing asymmetry: Pick 2 resolves Jun 13 — it is an early, cheap read on the Fable-debut thesis. If it loses because Fable lists at #1, immediately reassess Pick 3 (sell NO if Fable's provisional score is ≥1530) and Pick 1 (Fable strength is correlated).
Execution notes
- Limit orders only; all four books are thin enough that market orders pay 2–5¢. Rest at: NO 43–45 (Pick 1), YES 61–62 (Pick 2), NO 89–90 (Pick 3), NO 91–92 (Pick 4). Partial fills are acceptable and expected on Picks 1–2.
- Fees: Kalshi quadratic fees peak mid-curve — ≈1.7¢/contract at 44¢ (Pick 1), ≈1.6¢ at 61¢ (Pick 2), ≈0.6¢ at 90¢. Pick 1's edge survives comfortably; Picks 3–4 require maker fills (KXLLM1 pays maker rebates: quadratic_with_maker_fees).
- Invalidation triggers: (a) Fable 5 appears on arena.ai before Jun 12 → exit Pick 2 at market, re-evaluate 3; (b) Datacurve announces a Fable 5 DeepSWE run ≥68% → cut Pick 1 to half; (c) credible Gemini 3.2 launch date before Jun 30 → exit Pick 4.
- Watchlist (not deployed): KXFRONTIER-FRON-26JUL01 (22/29, dead book — bid 20 for a Fable-math-result lottery if liquidity appears); KXANTHROPICRESCIND-26JUN (DoD/Anthropic policy reversal, zero volume — would reprice violently on a single headline).
5. What I rejected and why
In v2 every Stage-2-surviving reject below is machine-logged in picks.json and shadow-tracked to settlement, so whether this screen adds selection skill is a testable claim, not an anecdote.
| Market | Side I'd have taken | Price | My P | Why rejected |
|---|---|---|---|---|
| KXCODEAI-26JUN30-CHAT | YES | 46 | 55% | Same thesis as Pick 1, worse execution: ~3 contracts offered at 46, real size at 49 where the edge halves. One thesis, one (better) leg. |
| KXAISPIKE-26B-1525 | NO | 72 | 70% | The crowd repriced this within hours of the launch (15→31 on 10.5k contracts) to almost exactly my number. Price already moved on the thesis — edge eaten. The clean illustration that the 1550 strike, not this one, kept the froth. |
| KXLLM1-26JUN30-A | YES | 92 | 93% | ≤2¢ edge, and Polymarket's $12.7M equivalent book sits at 91.1% — the largest AI prediction book in existence disagrees with my number by more than my edge. Respect it. |
| KXCLAUDE-MYTH-26JUL01 | YES | 3 | 8% | Definitional trap. Anthropic shipped "Claude Mythos 5" today — but only to Glasswing-vetted security orgs. Does gated deployment count as "releases a model called Mythos"? ~490k contracts across both deadline variants litigated this to ≤3¢ today. A 30:1 ticket on a resolver's reading of one verb is not an edge. |
| KXFRONTIER-FRON-26JUL01 | YES | 29 | — | 7¢ spread on 44 contracts of daily volume — unfillable without ≥3¢ slippage. No probability committed; never reached full diligence. Watchlist. |
| KXANTHROPICRESCIND-26JUN | YES | 25 | — | The most interesting story in the category (DoD has a supply-chain-risk phaseout directive against Anthropic, per the contract rules text) and a completely dead book: zero 24h volume, OI 85. Unfillable. Watchlist. |
6. Sources
- arena.ai — LLM text leaderboard (fetched Jun 9; board update Jun 5, 2026) — resolution-source data for Picks 2–4
- Datacurve DeepSWE leaderboard (fetched Jun 9; updated Jun 7, 2026) — resolution-source data for Pick 1
- Anthropic — Introducing Claude Opus 4.8 (May 28, 2026) — listing-lag baseline
- GitHub Changelog — Opus 4.8 GA (May 28) and Fable 5 GA (Jun 9)
- TechCrunch — Anthropic releases Claude Fable 5 (Jun 9, 2026; fetched) — safety fallback to Opus 4.8, ≤5% deferral, Glasswing terms
- Inc. — Fable 5 release — 80.3% SWE-Bench Pro claim
- VentureBeat — DeepSWE crowns GPT-5.5, finds Claude exploiting a benchmark loophole (May 26, 2026)
- The Neuron — DeepSWE verifier-quality explainer
- Polymarket — Best AI model end of June (fetched Jun 9) — cross-market calibration; confirms lmarena.ai resolution convention
- CNBC and 9to5Google — Fable/Mythos launch coverage (Jun 9, 2026)
- LocalAI Master — LMArena rankings archive — March 2026 ceiling reference (secondary source; used only for ceiling-stability claim)
- Kalshi read-only DB mirror (
trading_events,trading_markets,market_snapshots5-min bars) and Kalshi public API orderbooks, fetched 2026-06-09 ~19:05–19:20 UTC