Scan: Will OpenAI have the best AI model on LiveBench (Coding) at the end of September 2026?

1. Planner Fable 5.1 · Fermi decomposition + tool plan

Tools chosen: polymarket_direct polymarket_related kalshi_related claude_news claude_news gdelt_news wikipedia code_execution

Sub-questions (Fermi decomposition)

Which company currently holds the top Coding score on the latest LiveBench release, and what is the margin between the top model and the best OpenAI model?
How often has the #1 Coding spot on LiveBench changed hands over the past 12 months, and what fraction of that time did OpenAI hold it?
What frontier model releases (GPT-5.x/o-series, Claude, Gemini, Grok) are announced or rumored for August-September 2026, and how quickly does LiveBench typically add new models after release?
What is the current Polymarket price for OpenAI in this market group, and do the sibling markets (Google, Anthropic, xAI, Other) sum to roughly 100% after de-vigging?
Does the LiveBench Coding category structure (e.g., agentic coding tasks, cost-per-task display) currently favor a particular lab's model design?
Are there any known issues with LiveBench availability or methodology changes planned before the September 30 check date?

Planner reasoning

This is a Polymarket question about which lab tops the LiveBench Coding leaderboard on Sept 30, 2026. The key drivers are the current leaderboard standings (who leads and by how much), the expected release cadence of frontier models from OpenAI vs Anthropic/Google/xAI before end of September, and historical volatility of the top spot on this specific benchmark. The Polymarket price is the primary anchor, with sibling markets in the same event group (Google, Anthropic, xAI) useful for de-vigging.

2. Research Parallel tool calls

tool	status	time	items	summary
polymarket_direct	OK	0.4s	1	## This Market's Polymarket Data Will OpenAI have the best AI model on LiveBench (Coding) at the end of September 2026? - Current price (probability): 44.00% - 7-day price change: +8.50% - 30-day price change: +12.50% - Total volume: $16,023 (USD notional) - Price range: 16.00% - 50.00% - Data
polymarket_related	OK	2.9s	2	Scanned 100 active Polymarket markets, kept 2 matches. keyword 'LiveBench': 0 markets \| keyword 'best AI model': 2 markets \| keyword 'coding': 0 markets
kalshi_related	OK	2.7s	1	1 related markets / summaries. keyword 'LiveBench': no matches \| keyword 'best AI model': no matches \| keyword 'OpenAI': ok
claude_news	OK	30.2s	7	Based on available search results, here are key findings (note: some sources conflate the distinct LiveBench benchmark with the unrelated LiveCodeBench benchmark — I've flagged this where relevant): - LiveBench overall (not coding-specific) leaderboard, June 2026 snapshot: LiveBench re
claude_news	OK	22.3s	8	Based on my research, here are the key findings: - Anthropic currently leads coding benchmarks (not OpenAI): Coding results are the strongest part of the release: #1 on LiveCodeBench (90.52%), #2 on Terminal-Bench 2.1 (85.02%, up from 80.52% for Fable 5 and behind only GPT-5.6 Sol at 85.77%)
gdelt_news	OK	164.6s	10	GDELT: 10 articles across 3 queries (lookback=45d). 'LiveBench coding leaderboard': 10 hits \| 'OpenAI new model coding benchmark': error GDELT rate-limited after retries (429) \| 'Claude Gemini Grok coding benchmark top': error GDELT rate-limited after retries (429)
wikipedia	FAILED	0.2s	0	http error: Server error '503 Service Unavailable' for url 'https://en.wikipedia.org/w/api.php?action=query&format=json&prop=extracts&exintro=true&explaintext=true&redirects=1&titles=LiveBench%7COpenA
code_execution	OK	36.6s	0	Note: No actual sibling-market price data or leaderboard history was provided in this conversation, so the figures below use representative/illustrative inputs to demonstrate the methodology. Replace the raw prices and historical leader-months with real research data to get a live answer. - **D

3. Evidence Brief Sonnet 5 · 7775 chars

# Current state LiveBench.ai's Coding category leaderboard was not directly retrievable in this research pass; the best available proxies (LiveCodeBench, Terminal-Bench 2.1, BenchLM composite) show Anthropic's Claude Fable 5/5.1 leading most coding-adjacent benchmarks as of late Aug/early Sept 2026, with OpenAI's GPT-5.6 Sol close behind (and ahead specifically on Terminal-Bench 2.1 and agentic tasks). No source confirmed which company currently sits #1 specifically on LiveBench's own Coding tab. # Timeline of key events - 2026-06-25 (confirmed): LiveBench overall leaderboard refresh shows Claude Fable 5 (83.0) #1, GPT-5.6 Sol (81.0) and GPT-5.5 (80.2) trailing. [failingfast.io, benchlm.ai mirror] - 2026-07-22–24 (reported): Anthropic launches Claude Opus 5, quickly rated top on multiple Arena boards incl. both coding boards. [felloai.com] - Late Jul/Aug 2026 (confirmed): OpenAI releases GPT-5.6 (Sol), touted for agentic/browsing gains (BrowseComp SOTA, ExploitBench improvements), but trailing Anthropic on head-to-head coding (Terminal-Bench 2.1: GPT-5.6 Sol 85.77 vs Fable 5.1 85.02, however GPT-5.6 Sol actually leads this metric). [openai.com] - 2026-08-04 (reported): "Fable 5 Laps Field on MirrorCode" — benchmark-design article notes GPT-5.5 score collapse tied to design changes, illustrating benchmark volatility. [techtimes.com] - 2026-08-25 (reported): LiveCodeBench (Artificial Analysis) shows Google's Gemini 3 Pro Preview #1 (91.7%), ahead of Gemini 3 Flash and DeepSeek. [artificialanalysis.ai] - Early Sept 2026 (reported): Vals.ai LiveCodeBench tracker shows Claude Fable 5 leading (89.78%), Claude Opus 5 second, Gemini variants third/fourth, GPT-5.2 Codex fifth (87.99%) — note this uses an older GPT model, possibly stale. [vals.ai] - 2026-09-02 (reported): Alibaba (Qwen3.8-Max) and Tencent (Hy4) release updates, adding competitive pressure from Chinese open models, though not shown as coding leaders. [technode.com, scmp.com] # Event Will OpenAI own the #1-ranked model on LiveBench's Coding-category leaderboard when checked Sept 30, 2026, 12pm ET? # Outcomes to forecast Yes / No # Kalshi market anchor No direct kalshi_direct pull was returned; the only matching-market data available is from polymarket_direct for this same question: **current price 44% YES**, up sharply from a 16% low (30-day range 16–50%), +8.5% over 7 days and +12.5% over 30 days, on modest volume ($16k total, 38 data points). Trend is clearly upward in OpenAI's favor recently. Treat 44% as the consensus anchor to beat. # Sub-question answers 1. **Current #1 and margin** — Direct LiveBench Coding data unavailable; proxies (LiveCodeBench, Terminal-Bench, BenchLM) mostly show Anthropic's Claude Fable 5/5.1 slightly ahead of OpenAI's GPT-5.6 Sol (margins ~1–3 points), except Terminal-Bench 2.1 where GPT-5.6 Sol edges ahead (85.77 vs 85.02). Mixed signal, leaning Anthropic overall. [claude_news via vals.ai, benchlm.ai, openai.com] 2. **12-month turnover of #1 spot** — No real historical data found; only illustrative/synthetic figures were generated by code_execution (not actual research), suggesting ~3 leadership changes/12mo hypothetically. This is NOT a factual answer — genuine turnover history is a gap. 3. **Aug–Sept 2026 releases** — OpenAI: GPT-5.6/Sol (late Jul/Aug). Anthropic: Claude Opus 5 (Jul 24), Fable 5/5.1. Google: Gemini 3 Pro/Flash Preview. Also Qwen3.8-Max (Alibaba, Sep 2) and GLM 5.2 (Zhipu) as open-weight contenders. LiveBench's own update cadence (how fast it ingests new models) was not documented in sources; its last confirmed refresh was 2026-06-25. 4. **Polymarket sibling markets** — No sibling breakdown exists for the Coding-specific market group. A different, broader "best AI model" Polymarket pair shows OpenAI 24.4% / Anthropic 74.5% (sums to ~99.9%, near-binary framing, no Google/xAI/Other shown) — not directly transferable to the Coding-specific question. Coding-market siblings unknown (gap). 5. **Category structure bias** — Some evidence GPT-5.6 favors agentic/tool-use tasks (BrowseComp, ExploitBench) while Anthropic's models score higher on raw code-generation/SWE-bench-style tasks; LiveBench's Coding category composition (agentic vs. static) not documented in sources — gap. 6. **LiveBench availability/methodology risk** — No planned changes found; one article flagged that benchmark design changes previously caused a "GPT-5.5 score collapse" on a different benchmark (MirrorCode), showing sensitivity to methodology shifts, but nothing specific to LiveBench Coding before Sept 30. # Key facts (high-confidence, factual) 1. [polymarket_direct] This exact market currently prices OpenAI YES at 44%, up from a 16% low, with clear upward momentum over 30 days. 2. [openai.com] OpenAI's latest model is GPT-5.6 ("Sol"), released ~late July/August 2026. 3. [multiple: vals.ai, benchlm.ai, felloai.com] Anthropic's Claude Fable 5/5.1 and Claude Opus 5 lead most independent coding-adjacent benchmarks (LiveCodeBench, Terminal-Bench composite, BenchLM) as of Aug–Sept 2026. 4. [artificialanalysis.ai] Google's Gemini 3 Pro Preview leads LiveCodeBench specifically (91.7%) as of Aug 25, 2026 snapshot. 5. [failingfast.io] LiveBench's last confirmed refresh (overall, not coding-specific) was 2026-06-25, with Claude Fable 5 #1 overall. # Cross-market signals - Kalshi related: No direct LiveBench/coding markets found; only tangential OpenAI-related Kalshi markets (IPO race, US equity stake, Apple lawsuit) with no bearing on model rankings. - Polymarket (this market mirror): 44% YES, rising trend. - Polymarket (broader "best AI model," different market): OpenAI 24.4% vs Anthropic 74.5% — suggests market participants view Anthropic as favorite for general model supremacy, though this is a distinct, non-coding-specific market. - No sportsbook-style odds applicable. # Analyst opinions and speculation - Claude_news synthesis: "leaning against" OpenAion holding LiveBench Coding #1, citing Anthropic's broad benchmark dominance in coding-adjacent tests, though caveats that no source directly confirmed LiveBench Coding-category standings. - Some sources note OpenAI's relative strength in agentic/tool-use coding tasks (Terminal-Bench, ExploitBench) — if LiveBench's Coding category weights agentic tasks heavily, this could favor OpenAI. # Directional lean per outcome - **Yes (OpenAI)**: Rising Polymarket price (16%→44%) suggests market sentiment shifting toward OpenAI; GPT-5.6 Sol leads on Terminal-Bench 2.1 and agentic coding subtasks. - **No (not OpenAI)**: Majority of coding-benchmark proxies (LiveCodeBench, BenchLM, Vals.ai) show Anthropic (Fable 5/5.1, Opus 5) ahead of OpenAI; Google's Gemini 3 also competitive/leading on LiveCodeBench specifically. Direct LiveBench Coding data not confirmed for either side. # Gaps / unknowns - No direct read of current livebench.ai Coding tab was obtained (data conflated with LiveCodeBench across sources). - Real historical turnover rate of LiveBench Coding #1 spot is unknown (only illustrative/fake data was generated). - No genuine sibling-market prices for the Coding-specific Polymarket/Kalshi group (Google, Anthropic, xAI, Other). - LiveBench's exact task composition and update/ingestion cadence for new model releases is undocumented. - No confirmed info on planned LiveBench methodology changes before Sept 30, 2026 check. # Calibration anchors - Kalshi/Polymarket current YES price for this exact market: **44%** (trending up). - Related broader-market anchor: OpenAI 24.4% vs Anthropic 74.5% on "best AI model overall" (different, non-coding market) — suggests OpenAI's coding-specific odds may reasonably sit above its "best overall" odds, consistent with the 44% figure, given OpenAI's specific coding/agentic strengths.

4. Forecasters Fable 5.1 + GPT-5.6 Sol · in parallel

claude-fable-5-1

33% conf

Yes	41%
No	59%

The only direct market anchor is Polymarket at 44% YES with strong upward momentum (16%→44% over 30 days), which suggests traders may be reacting to GPT-5.6 Sol's competitiveness. However, most coding-adjacent proxies (Vals.ai LiveCodeBench, BenchLM, LiveBench's June overall refresh) show Anthropic's Claude Fable 5/5.1 and Opus 5 slightly ahead of OpenAI, with Google's Gemini 3 Pro also leading on at least one coding benchmark, so OpenAI faces two credible competitors for a single #1 slot. GPT-5.6 Sol's edge on Terminal-Bench 2.1 and agentic tasks provides a real path to Yes if LiveBench's Coding tab weights such tasks, and the ~4-week window leaves room for a new OpenAI release or LiveBench ingestion. Absent a direct read of the LiveBench Coding tab, I stay close to the market anchor but shade slightly toward No given the preponderance of proxies favoring Anthropic.

gpt-5.6-sol

46% conf

Yes	42%
No	58%

The 44% YES market price is the best consensus anchor, although modest volume and a recent rise from 16% indicate substantial volatility. Current-position proxies lean No because Anthropic leads most coding-adjacent benchmarks by roughly 1–3 points, while Google also leads one LiveCodeBench snapshot; OpenAI’s strongest counterevidence is its narrow Terminal-Bench 2.1 lead. Unknown leaderboard turnover, LiveBench model-ingestion timing, category composition, and methodology changes prevent a strong adjustment, while the broad set of recent competitor releases further favors No at the margin. I therefore shade the forecast slightly below the 44% anchor rather than moving aggressively without direct LiveBench Coding standings.

5. Devil's Advocate Sonnet 5 · red-team critique

Both forecasters converge tightly around 41-42% with minimal daylight, essentially just parroting the Polymarket anchor without independently weighing that the anchor itself is thin (only $16k volume, 38 data points) and highly volatile (16%→50% range in 30 days) — this warrants a wider confidence interval or more deviation from the anchor, not near-identical point estimates.

Neither forecast adequately grapples with the three-or-more-way competition: Anthropic (Fable 5/5.1, Opus 5) AND Google (Gemini 3 Pro at 91.7% on LiveCodeBench) are both plausible #1 holders, meaning "No" is actually a composite of at least two strong competitors plus Chinese open models (Qwen3.8-Max, GLM 5.2) — the true OpenAI "Yes" probability may deserve more downward pressure than a simple binary framing suggests, since Google's Gemini 3 Pro lead is arguably the most recent (Aug 25) hard number available on a LiveCodeBench-style metric.

Both forecasters underweight sub-question 5 (category structure bias) — LiveBench Coding's actual task composition (agentic vs. static code-gen) is undocumented, yet this is the single biggest lever that could flip the answer given GPT-5.6 Sol's specific strength in agentic/Terminal-Bench tasks; this genuine uncertainty should widen the distribution rather than just be noted as a caveat.

Neither forecast discusses LiveBench's ingestion lag risk sufficiently: the last confirmed refresh was 2026-06-25, over 3 months before the close date, and if LiveBench fails to ingest GPT-5.6 Sol, Claude Opus 5, or Gemini 3 Pro by Sept 30, the "current" leaderboard could reflect stale rankings (e.g., June's Fable 5 #1) — this scenario could be assessed to favor No more strongly (frozen Anthropic lead) or introduce unpredictable resolution ambiguity that neither model flags as a distinct tail risk.

The broader "best AI model overall" Polymarket sibling (OpenAI 24.4% vs Anthropic 74.5%) is inconsistent with the coding-specific 44% anchor, and neither forecaster explores why coding-specific odds for OpenAI would be nearly double its overall odds — this gap deserves explicit reconciliation (e.g., is the coding market mispriced relative to the overall market, or does OpenAI genuinely have a differentiated coding-specific edge that doesn't show up in "overall" assessments?).

Both forecasts anchor to Polymarket's 44% while treating it as roughly equivalent to a "Kalshi YES price," but the brief explicitly states no direct Kalshi pull was available — neither forecaster flags that Kalshi's own market (if it exists and prices differently) could diverge from Polymarket's thin, volatile 38-data-point market, and using Polymarket as if it were an interchangeable Kalshi-anchor is a methodological shortcut both share without justification.

6. Aggregator Fable 5.1 · final calibrated forecast

40% (40% confidence)

Yes: 40% No: 60%

Both forecasts sit just under the 44% Polymarket anchor, and the critique correctly notes that anchor is thin ($16k, 38 data points) and volatile, so it deserves less deference than a liquid market would. The substantive evidence leans No: most coding proxies (Vals.ai LiveCodeBench, BenchLM, LiveBench's June overall refresh) show Anthropic ahead, Google's Gemini 3 Pro leads at least one recent coding snapshot, and 'No' aggregates multiple credible competitors for a single #1 slot. Offsetting this, the sharp 16%→44% rise plausibly reflects traders observing something favorable to OpenAI (e.g., GPT-5.6 Sol's agentic strength, or an actual LiveBench Coding tab update not captured by the brief), and the undocumented task composition of LiveBench Coding is a genuine lever that could favor OpenAI. Netting these, I shade slightly further below the anchor than either forecaster, to 40% Yes.

Pipeline Timing

Total pipeline time: 282.6s

Per-tool research timings shown in the Research section above.