Tools chosen: polymarket_direct polymarket_related claude_news claude_news gdelt_news kalshi_related code_execution wikipedia

Sub-questions (Fermi decomposition)

Planner reasoning

This is a Polymarket question about which company tops the LiveBench Coding leaderboard on Aug 31, 2026, so the market's own price is the primary anchor, plus the sibling markets for Google/OpenAI/xAI in the same event group give an implied full distribution. Key empirical drivers are: who currently leads LiveBench Coding, by what margin, how frequently the leaderboard is updated/leadership changes, and what frontier model releases (Claude, Gemini, GPT) are expected before end of August 2026.

# Current state The event resolves off LiveBench.ai's "Coding" category leaderboard score checked at 2026-08-31 12:00 PM ET; whichever company owns the top-scoring model wins. As of mid-August 2026, no single source confirms who currently leads the *specific* LiveBench Coding sub-category — Anthropic's Claude Fable 5/Opus 5 lead LiveBench's *overall* score and several third-party coding aggregators (BenchLM), while Google's Gemini 3 Pro Preview leads the related-but-distinct LiveCodeBench, and OpenAI's GPT-5.5/5.6 leads on SWE-bench Verified (independent eval). The market itself (current YES 90%) is pricing Anthropic as heavy favorite. # Timeline of key events - 2025-11 to 2026-01: Robinhood/Kalshi market on "top LiveBench Coding Average" priced OpenAI at 99¢ entering Jan 2026 — OpenAI held the specific Coding leaderboard lead at that time (reported, robinhood.com). - 2026-06-20/23: Small/open models (tiny model, Sakana AI "Fugu") claim to beat Claude Fable 5 on coding benchmarks (reported, propakistani.pk/moneycontrol.com) — noise, not leaderboard-determinative. - 2026-06-25: LiveBench does its routine monthly contamination-control question rotation; Claude Fable 5 leads overall LiveBench snapshot at 83.0%, GPT-5.6 Sol 81.1%, GPT-5.5 80.2% (reported, benchlm.ai). - 2026-06-29/30: Reports of Gemini 3.5 Pro "cleared for July launch," Fable 5 "nearing return," GPT-5.6 "still locked" (rumored, techtimes.com). - 2026-07-13: Google changes Android-coding grading methodology (reported, Yahoo Tech); GLM-5.2 captures 40% developer tokens (reported, techtimes.com). - 2026-07-16: Bloomberg reports Gemini 3.5 Pro missed its third internal deadline, months behind schedule, coding capability below Google's internal bar (reported). - 2026-07-18: Kimi K3 found to be an illegal Claude distillation (reported, propakistani.pk) — reputational noise for Anthropic ecosystem. - 2026-07-24/25: Anthropic launches Claude Opus 5 — large SWE-bench Pro gains (69.2%→79.2%), doubles Opus 4.8 on Frontier-Bench agentic coding, same price as Opus 4.8 (confirmed via multiple outlets: arynews.tv, iclarified.com, codersera.com). - 2026-08-05/14/15: Google ships Gemini 3.6 Flash and Gemini 3.7 Flash (mid-tier, not flagship Pro replacement) (reported, thenextweb.com/digitaltrends.com/heise.de). - 2026-08-12: LiveCodeBench (distinct benchmark) shows Gemini 3 Pro Preview #1 at 91.7%, ahead of Gemini 3 Flash Preview and DeepSeek V3.2 (reported, pricepertoken.com). - 2026-08 (BenchLM aggregate): Claude Mythos 5 leads composite coding ranking (81.1), ahead of Claude Fable 5 (80.8) and GPT-5.6 Sol (78.7) (reported, benchlm.ai/coding). # Event Will Anthropic (vs Google/OpenAI/xAI/Other) own the #1-scoring model in LiveBench.ai's "Coding" category as checked 2026-08-31 12:00 PM ET? # Outcomes to forecast Yes (Anthropic tops LiveBench Coding) / No (any other company tops it) # Kalshi market anchor Cross-listed market data (same ticker) shows current price **90.0%** YES for Anthropic, up +3pp over 7 days and +40pp over 30 days (price range 50%–93.5% over 19 days of data). Volume modest (~$16K total). Strong recent momentum toward Yes, likely driven by the Opus 5 launch (2026-07-24/25) and subsequent benchmark chatter. # Sub-question answers 1. **Current LiveBench Coding leader / gap** — Not definitively confirmed for the exact "Coding" sub-category as of mid-Aug 2026; Anthropic leads LiveBench *overall* (Fable 5, 83.0%) and BenchLM's composite coding score, but a Jan-2026 Coding-specific market had OpenAI at 99%, and LiveCodeBench (a different benchmark) currently favors Google's Gemini 3 Pro. No source gives a clean current Coding-category score gap. 2. **Frequency of leadership change / base rate** — No direct historical cadence data found; illustrative code_execution model shows that under naive monthly-flip assumptions (p=0.2–0.5), persistence over the ~1-year horizon would be <15%, well below market's 90% — implying leadership is either much "stickier" than a memoryless model or market is pricing genuine Anthropic-specific durability. 3. **Is LiveBench still active in 2026?** — Yes; confirmed to have done a routine monthly contamination-control refresh on 2026-06-25 (pinggy.io), indicating active maintenance into Q3 2026. 4. **New Anthropic releases before Aug 2026** — Claude Opus 5 launched 2026-07-24/25 with major coding gains (SWE-bench Pro 69.2%→79.2%, Frontier-Bench agentic coding more than doubled vs Opus 4.8), same price tier (confirmed, multiple outlets). Claude Fable 5/Mythos 5 also referenced as top-tier Anthropic coding models in mid-2026. 5. **Competing releases (Google/OpenAI/xAI/DeepSeek)** — Google shipped Gemini 3.6/3.7 Flash (mid-tier) but flagship Gemini 3.5 Pro is delayed (Bloomberg, 2026-07-16, "months behind schedule"); OpenAI's GPT-5.5/5.6 Sol lead SWE-bench Verified independently; xAI shipped Grok 4.5 and Grok Build (agentic) in June-July 2026; DeepSeek V3.2/V4-Pro and Qwen3.8-Max also active but not leading closed-model benchmarks. 6. **Polymarket sibling market prices** — Only this Anthropic-outcome market data was retrieved directly (90%); no confirmed Google/OpenAI/xAI/Other sibling prices found (polymarket_related search returned 0 matches). A hypothetical illustrative de-vig (not live data) suggested Anthropic ~42%, Google ~27%, OpenAI ~20%, xAI ~7%, Other ~4% — but this is explicitly labeled illustrative, not live. # Key facts (high-confidence, factual) 1. [polymarket_direct] Current YES price for this exact market: 90%, +40pp over 30 days. 2. [pinggy.io] LiveBench actively refreshed questions 2026-06-25, confirming ongoing maintenance. 3. [multiple] Claude Opus 5 launched 2026-07-24/25 with substantial coding benchmark gains at flat pricing. 4. [Bloomberg via felloai.com/techtimes] Google's Gemini 3.5 Pro flagship delayed past three deadlines as of mid-July 2026. 5. [benchlm.ai] BenchLM's composite coding ranking (Aug 2026) has two Claude variants (Mythos 5, Fable 5) in top 2 spots. 6. [pricepertoken.com] Gemini 3 Pro Preview leads LiveCodeBench (different benchmark) as of 2026-08-12. # Cross-market signals - Kalshi related: Anthropic IPO-first market at 93% (confirms strong overall market confidence in Anthropic's momentum/position, but unrelated to coding benchmarks specifically). - Polymarket: No sibling outcome prices confirmed live; only Anthropic-outcome price found (90%). - Sportsbook implied: N/A. # Analyst opinions and speculation - Multiple benchmark aggregators (BenchLM, felloai.com) argue Anthropic's Opus 5/Fable 5/Mythos 5 lineage has "persistent" coding leadership through mid-2026, aided by Google's Gemini 3.5 Pro delay removing the most likely near-term challenger. - Others note "best" model is highly benchmark-dependent (SWE-bench vs LiveCodeBench vs LiveBench Coding vs Arena WebDev), with OpenAI and Google each winning on at least one major coding metric. # Directional lean per outcome - **Yes (Anthropic)**: Opus 5 launch strength, LiveBench-overall lead, BenchLM composite lead, Gemini 3.5 Pro delay removing top rival, strong and rising market price (90%, +40pp/30d). - **No (other)**: Historical Jan-2026 data had OpenAI leading LiveBench Coding specifically; Gemini leads LiveCodeBench; GPT-5.5/5.6 leads SWE-bench Verified; benchmark leadership in this space has shown volatility across releases (new GPT-5.x/Gemini-3.x variants could flip the specific Coding sub-score before Aug 31). # Gaps / unknowns - No confirmed direct read of the actual LiveBench.ai "Coding" category score/ranking as of the current date — all evidence is inferential from overall LiveBench score or other benchmarks. - No live Polymarket sibling-market prices for Google/OpenAI/xAI/Other confirmed (illustrative figures only). - Uncertain whether OpenAI or Google will ship a coding-focused frontier update before 2026-08-31 that could flip the specific Coding leaderboard. # Calibration anchors - Kalshi/Polymarket current YES price (anchor): 90%. - Jan-2026 sibling market precedent: OpenAI priced at 99% for LiveBench Coding Average lead — shows category leadership can be decisively one company at a given snapshot, and has flipped since (from OpenAI to contested/Anthropic-leaning).

1. Both forecasts anchor heavily on the 90% Kalshi price without adequately weighting sub-question #1's key finding: "No source gives a clean current Coding-category score gap," and the Jan-2026 precedent where OpenAI held 99% on this *exact* category just months earlier — this suggests the specific-category leaderboard is more volatile/flip-prone than the market's overall-score-driven momentum implies, warranting a larger discount than 8-11pp. 2. Neither forecast grapples with the base-rate math explicitly flagged in the brief: "under naive monthly-flip assumptions (p=0.2–0.5), persistence over the ~1-year horizon would be <15%, well below market's 90%" — they should have either explained why leadership is "stickier" than this model or discounted further; instead both simply defer to the anchor. 3. Both forecasts under-weight the possibility that OpenAI (which leads SWE-bench Verified and held Coding-specific leadership as recently as Jan 2026) could ship a coding-focused update in the final weeks before Aug 31 — the brief explicitly flags "GPT-5.6 still locked" rumors suggesting a pending release, and neither analyst separately assesses OpenAI's flip probability vs. Google's (they treat "No" as a monolithic residual dominated by Google's delay). 4. The BenchLM composite score gap cited as support for Anthropic is razor-thin (Mythos 5 81.1 vs Fable 5 80.8 vs GPT-5.6 Sol 78.7) — this is not a commanding lead and could easily flip with routine LiveBench's monthly contamination-control question rotation (confirmed to occur, e.g., 2026-06-25), a mechanism neither forecaster discusses as a source of exogenous variance right before the Aug 31 checkpoint. 5. Neither forecast addresses the "Other" category (xAI's Grok 4.5/Grok Build, DeepSeek V4-Pro, Qwen3.8-Max) as a tail risk — while unlikely, both treat No as effectively synonymous with Google/OpenAI, ignoring that an open-weight or xAI model could top a specific automated benchmark unexpectedly, a scenario the illustrative de-vig (Anthropic ~42%, Google ~27%, OpenAI ~20%, xAI ~7%, Other ~4%) at least allocates probability to but the final forecasts do not explicitly reconcile with. 6. Both forecasts collapse to within 7pp of the market price with only modest, similarly-reasoned discounts, showing insufficient differentiation given the acknowledged high uncertainty (no direct current Coding-subcategory read); a wider spread or more explicit sensitivity analysis on the "no clean confirmation" gap would better reflect true epistemic uncertainty rather than both anchoring near-identically to Kalshi.

85% (59% confidence)

Yes: 85% No: 15%

The Kalshi anchor (90% YES, +40pp/30d) plus a very short remaining horizon (~2 weeks) makes Anthropic the clear favorite: Claude Opus 5's late-July coding gains landed, Anthropic leads LiveBench overall and BenchLM's composite coding ranking, and the most credible challenger (Gemini 3.5 Pro) is reportedly months behind schedule with Google only shipping mid-tier Flash models in August. However, the critique lands on real weaknesses: no source directly confirms the LiveBench *Coding* sub-category ranking, OpenAI held that exact category at 99% as recently as Jan-2026, the BenchLM gap is razor-thin (81.1 vs 78.7), and LiveBench's monthly contamination-control question rotation injects exogenous variance right before the Aug-31 snapshot. The naive monthly-flip base rate is not the right model over a 2-week window — leadership is empirically sticky between frontier releases — so it justifies only a modest, not dramatic, discount. I therefore shade a few points below both the market and Forecast 1, landing at 0.85, which also implicitly leaves room for tail flips from OpenAI's pending GPT-5.x update or an unexpected xAI/open-weight benchmark spike.

Total pipeline time: 238.1s