Scan: Will Google have the best Text-to-Video AI 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 article_search gdelt_news wikipedia

Sub-questions (Fermi decomposition)

Which model currently ranks #1 (non-AutoEval) on the arena.ai Text-to-Video leaderboard, and what is the score gap between Google's top model and the #2 model?
How often has Google held the #1 spot on the Text-to-Video arena over the past 6-12 months, and how long have competitors (Kling, Sora, Seedance, Hailuo, Wan) displaced it when they did?
What major text-to-video model releases (Veo 4, Sora 3, Kling 3, Seedance 2, Grok Imagine, Wan) are announced or rumored for release before the end of September 2026?
What are the current Polymarket prices for each company in this market group (Google, OpenAI, Kuaishou, ByteDance, xAI, MiniMax, Alibaba, Runway, Other), and do they sum to roughly 1?
Has arena.ai changed its leaderboard methodology (e.g., AutoEval labeling, style control) in ways that could affect which model ranks #1 at check time?

Planner reasoning

This resolves on the arena.ai Text-to-Video leaderboard rank #1 on Sept 30, 2026; Google's Veo models have historically held the top slot but face competition from Kling (Kuaishou), OpenAI Sora, xAI, ByteDance Seedance, MiniMax Hailuo, Runway, and Alibaba Wan. Key uncertainties are the current standing, the margin over #2, and whether a competitor releases a new model before September. Polymarket price is the primary anchor; sibling markets in the same group give implied probabilities for each company.

2. Research Parallel tool calls

tool	status	time	items	summary
polymarket_direct	OK	0.9s	1	## This Market's Polymarket Data Will Google have the best Text-to-Video AI at the end of September 2026? - Current price (probability): 88.50% - 7-day price change: +0.50% - 30-day price change: +7.50% - Total volume: $15,592 (USD notional) - Price range: 39.00% - 92.50% - Data points: 34 days
polymarket_related	OK	5.4s	0	Scanned 100 active Polymarket markets, kept 0 matches. keyword 'best text-to-video': 0 markets \| keyword 'text-to-video AI': 0 markets \| keyword 'Veo': 0 markets \| keyword 'Sora': 0 markets \| keyword 'Kling': 0 markets
kalshi_related	OK	5.3s	2	2 related markets / summaries. keyword 'text-to-video': no matches \| keyword 'video AI': ok \| keyword 'best AI video model': ok
claude_news	OK	34.6s	12	Here are key findings on the Text-to-Video AI landscape as of September 2026: Arena.ai (LMArena) Leaderboard Status - The official arena.ai Text-to-Video leaderboard, updated Sep 4, 2026, shows 668,045 votes across 48 models, with scores ranging from a min of 1006 to a max of 1515 . (Specific
article_search	OK	0.7s	35	Fetched 35 unique articles across 4 queries (lookback=60d). 'arena.ai text-to-video leaderboard top model': 10 articles \| 'Google Veo video model release 2026': 10 articles \| 'Kling Seedance Sora video model leaderboard ranking': 10 articles \| 'new text-to-video AI model launch': 10 articles
gdelt_news	OK	196.5s	30	GDELT: 30 articles across 5 queries (lookback=45d). 'Veo text-to-video leaderboard': error GDELT rate-limited after retries (429) \| 'Seedance video model': 10 hits \| 'Kling AI video model': 10 hits \| 'Sora video model update': error GDELT rate-limited after retries (429) \| 'Grok Imagine video': 10 h
wikipedia	OK	0.2s	4	Fetched 4 Wikipedia entries (0 missing pages).

3. Evidence Brief Sonnet 5 · 8099 chars

# Current state As of Sep 2026, the arena.ai Text-to-Video leaderboard's #1 spot is genuinely contested and conflicting across sources — some report a Google model (Veo 3.1/Gemini Omni Flash "preliminary") at #1, others report Kuaishou's Kling v3 (score 1934) leading, with Alibaba's "Happy Horse 1.0" (1816) and ByteDance's Seedance close behind. No single authoritative snapshot of the actual arena.ai page (non-AutoEval filtered) was captured in research. Kalshi-direct price for this ticker was not retrieved; the same-ticker Polymarket price is 88.5% YES (Google), reflecting bettor confidence despite the ambiguous underlying data. # Timeline of key events - 2024-05: Google announces Veo (Veo 1) [confirmed, Wikipedia]. - 2025-05: Veo 3 released with audio generation [confirmed, Wikipedia]. - 2025-09 (end): OpenAI releases Sora 2 to select US/Canada users [confirmed, Wikipedia]. - ~2025-Q4/2026-Q1: LMArena Video Arena launches; early rankings show Veo3 (audio on) #1, Veo3/Veo3-fast #3, Hailuo 02 & Seedance 1.0 pro #5, Kling 2.1 Master #6 [reported, claude_news/X]. - 2026-04-26: OpenAI shuts down the Sora consumer app [confirmed, Wikipedia]. - 2026-05-19/20: Google I/O 2026 — unveils "Gemini Omni" unified multimodal platform; no formal "Veo 4" branding announced [confirmed/reported, claude_news]. - 2026-06-30: Two new Google models (per agentpedia, likely Gemini Omni Flash variants) launch, initially on preview/pre-release access [reported, claude_news]. - 2026-08 (Kuaishou Q2 earnings): Kling AI revenue tops RMB850M, up 200%+ YoY, signaling strong commercial momentum for Kling [confirmed, TechNode/gdelt_news]. - 2026-08-12: Seedance 2.5 cited in deepfake-blocking context (Deadline); TikTok/Dreamina roll out Seedance 2.5 [confirmed, gdelt_news]. - 2026-09-04: arena.ai leaderboard shows 668,045 votes, 48 models, score range 1006–1515 (exact #1 name not captured) [reported, claude_news]. - 2026-09-24 (upcoming): OpenAI's Sora API sunsets, removing OpenAI entirely from contention [reported, claude_news/Wikipedia]. - 2026-09 (ongoing): Multiple conflicting rankings published — llm-stats.com claims Kling v3 #1 (1934) over Happy Horse (1816) and Seedance 2.0 Fast (1747); codesota.com claims Veo 3.1 (1381) #1 over Veo 3.1 Fast (1378) and Sora 2 Pro (1367) [both rumored/unverified, conflicting]. # Event Will Google's model hold the #1 rank on arena.ai's Text-to-Video leaderboard (non-AutoEval, Models filter) at check time Sep 30, 2026, 12:00 PM ET? # Outcomes to forecast Yes (Google), No (any other company) # Kalshi market anchor No kalshi_direct data was returned in research (tool not queried/no result). Cross-listed Polymarket price for this identical ticker: **88.5% YES**, up +0.5% (7d) and +7.5% (30d), trading range 39%–92.5% over 34 days, modest volume (~$15.6K). This is a thinly traded market with a wide historical range, suggesting past volatility/uncertainty that has since resolved toward Google in bettor sentiment — but this conflicts with several factual news sources. # Sub-question answers 1. **Current #1 model / score gap** — Unclear/contested. One source (llm-stats.com) says Kling v3 (Kuaishou) leads at 1934 vs. Google not in top 3. Another (codesota.com) says Veo 3.1 leads at 1381 vs. Veo 3.1 Fast (1378) and Sora 2 Pro (1367). A third (agentpedia) implies a Google model (Gemini Omni Flash) is #1 in Text-to-Video but tagged "Preliminary." No single verified live snapshot of arena.ai itself was captured. 2. **Historical Google #1 tenure** — Early LMArena Video Arena rankings (2025) had Veo3 at #1. Since then, Chinese models (Kling, Seedance, Alibaba's Happy Horse) have reportedly displaced Google at various points; exact duration/frequency of swaps is not documented in research. 3. **Major releases before Sep 2026** — No "Veo 4" launched; Google rebranded to "Gemini Omni" (announced I/O 2026, two new models launched 2026-06-30, still "preliminary" on Arena). Sora effectively discontinued (app shut 2026-04-26, API sunsets 2026-09-24). Seedance 2.5 released and adopted by TikTok/Dreamina (Aug 2026). Kling v3 apparently released/leading per one source. No Grok Imagine video-specific update found (only Grok Imagine Image 2.0, Aug 2026). 4. **Polymarket prices across the group** — Only this Google-outcome market's price was retrieved (88.5%); no data on OpenAI/Kuaishou/ByteDance/xAI/MiniMax/Alibaba/Runway/Other splits, so cannot confirm sum ≈1. polymarket_related search found 0 matching markets for other companies. 5. **Methodology changes (AutoEval, style control)** — No explicit mention of methodology changes found; however, multiple Google models are tagged "Preliminary" on Arena (per agentpedia), which is relevant to the "AutoEval" exclusion rule but distinct from it — unclear if these preliminary tags equal AutoEval exclusion. # Key facts (high-confidence, factual) 1. [Wikipedia] Sora app shut down 2026-04-26; API sunsets 2026-09-24 — OpenAI effectively exits. 2. [Wikipedia] Veo 3 (with audio) released May 2025; remains Google's officially branded video product as of research date. 3. [claude_news] Google announced "Gemini Omni" at I/O 2026 (no Veo 4); two new models launched 2026-06-30, still "Preliminary" on Arena. 4. [TechNode] Kuaishou's Kling AI revenue grew 200%+ YoY in Q2 2026, indicating strong Kling investment/momentum. 5. [claude_news] Arena.ai leaderboard updated 2026-09-04 with 668,045 votes across 48 models (exact #1 not confirmed). # Cross-market signals - Kalshi related: No direct match found for this market; unrelated SCOTUS market returned as noise. - Polymarket (same ticker): 88.5% YES, uptrend over 30d (+7.5pp), but historically dipped to 39% — high volatility/uncertainty priced in at times. - Sportsbook implied: N/A (not applicable to this event type). # Analyst opinions and speculation - claude_news synthesis: "genuinely contested/volatile ranking with reputable-seeming sources on both sides" — no consensus on current #1. - Industry roundup (per claude_news): "Veo 3.1 is no longer the outright #1... stays high for teams that want a familiar Google tool" — implies Google has slipped from top position by some measures. - llm-stats.com/codesota.com present starkly opposite #1 claims, suggesting either differing methodologies, different leaderboard snapshots in time, or unreliable secondary aggregation (neither may reflect true arena.ai filtered data). # Directional lean per outcome - **Yes (Google)**: Polymarket cross-market price strongly favors Google (88.5%, rising). Google has historically held #1 (Veo3 launch era) and has fresh "Gemini Omni" models in preliminary testing that could earn #1 once fully ranked. - **No (other)**: Multiple factual signals point to Chinese labs (Kling v3, Seedance, Alibaba) gaining significant ground — Kling revenue up 200%+, Seedance widely adopted (TikTok/Dreamina), one source explicitly puts Kling v3 at #1 with large score gap (1934 vs sub-1400 Google scores per another source). Google's newest models remain "Preliminary" (uncertain Arena eligibility/exclusion risk under AutoEval rule). # Gaps / unknowns - No verified live pull of arena.ai's actual "Models"-filtered, non-AutoEval leaderboard — the single most decisive data point is missing. - Unclear whether Google's "Preliminary"-tagged Gemini Omni models will carry the "AutoEval" label that disqualifies them per market rules — critical structural uncertainty. - No confirmation of Kalshi's own current YES price (only Polymarket data available) — brief anchor is a proxy, not true Kalshi consensus. - No data on other outcome-group prices (OpenAI, Kuaishou, ByteDance, etc.) to cross-check probability coherence. # Calibration anchors - Cross-market (Polymarket, same ticker) YES price: 88.5% (rising trend, but historically ranged 39–92.5%). - Precedent: In fast-moving AI leaderboards, top rank has flipped between Google/OpenAI/Chinese labs multiple times within months (e.g., Veo3 launch #1 → later displaced per multiple 2026 sources) — high base-rate of leadership churn argues for caution against over-anchoring on any current snapshot.

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

claude-fable-5-1

40% conf

Yes	82%
No	18%

The strongest signal is the same-ticker Polymarket price of 88.5% YES (rising +7.5pp over 30d), and bettors there can observe the live arena.ai leaderboard, which the research brief could not capture directly. The conflicting secondary sources are partly reconcilable: llm-stats' Kling v3 at 1934 is outside arena.ai's reported 1006–1515 score range, suggesting it reflects a different leaderboard, while agentpedia's indication of a Google (Gemini Omni Flash, preliminary) model at #1 aligns with the market price. Risks include high churn in video leaderboards, strong momentum from Kling v3/Seedance 2.5, thin market volume, and uncertainty about whether 'Preliminary' tagging interacts with the non-AutoEval filter, so I shade modestly below the market anchor.

gpt-5.6-sol

43% conf

Yes	76%
No	24%

The 88.5% YES Polymarket price is the best available consensus anchor, and its recent uptrend suggests bettors may have fresher ranking information than the secondary sources. Current-rank evidence remains contradictory, however, with one source placing Kling v3 first while others place Veo 3.1 or Gemini Omni Flash first. Google’s historical leadership and June model releases support Yes, while strong Kling and Seedance momentum plus uncertainty over whether Google’s preliminary models qualify for the specified filter support No. I therefore discount the thin-market anchor substantially because no authoritative filtered arena.ai snapshot or Kalshi-direct price was retrieved.

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

Both forecasters anchor heavily on the Polymarket 88.5% price, but the brief flags this as "thinly traded" (~$15.6K volume) with a 39%–92.5% range over 34 days — high volatility that should warrant much more discounting than the ~6-9pp shave applied; a thin, volatile proxy market observing the same ambiguous public data as the researchers doesn't actually resolve the informational gap, since there's no evidence bettors have privileged access to the true filtered leaderboard.

Neither forecast adequately weights the structural/methodological risk flagged in the brief: Google's newest models are tagged "Preliminary," and it's explicitly unclear whether this triggers the "AutoEval" exclusion rule that would disqualify them from the market's specific "non-AutoEval, Models filter" criterion — this is called "critical structural uncertainty" in the brief but treated as a minor caveat rather than a potentially binary disqualifier.

Both forecasts converge suspiciously close together (0.82 vs 0.76) despite claiming to weigh the evidence differently, and neither seriously engages with the "Industry roundup" quote stating "Veo 3.1 is no longer the outright #1" — a fairly direct piece of contrary evidence that gets mentioned but not incorporated into probability movement.

The base-rate argument for churn ("top rank has flipped between Google/OpenAI/Chinese labs multiple times within months") is acknowledged by both but not translated into a wider probability spread — given documented leadership flips within months, a snapshot 3+ weeks out (forecast made well before Sep 30 close) should carry more uncertainty than ~76-82% implies.

Both forecasts ignore the possibility that Kalshi's actual price (never retrieved) could differ meaningfully from Polymarket's, and treat the proxy as if it were a validated anchor rather than flagging this substitution as a first-order risk to their own confidence calibration.

Neither forecast addresses the quantitative scoring evidence directly: llm-stats' Kling v3 score of 1934 with Alibaba's Happy Horse at 1816 implies a large gap over any Google model, and Forecast 1's dismissal of this ("outside arena.ai's reported 1006-1515 score range") is speculative reconciliation, not confirmed fact — this key uncertainty should lower confidence further, not be used to explain away the strongest anti-Google data point.

6. Aggregator Fable 5.1 · final calibrated forecast

79% (42% confidence)

Yes: 79% No: 21%

The same-ticker Polymarket price (88.5%, stable over 7d and rising over 30d) is the best available signal because the arena.ai leaderboard is public and bettors can observe the live filtered ranking that the research brief failed to capture; a 39% low earlier in the window likely reflects a genuine temporary displacement that has since reversed. The llm-stats Kling v3 score of 1934 is inconsistent with arena.ai's documented 1006–1515 range, so it most plausibly reflects a different leaderboard, weakening the strongest anti-Google data point. However, the critique is right that the market is thin, Kalshi-direct was not retrieved, video leaderboards churn frequently with Kling/Seedance momentum strong, and the 'Preliminary' tag on Gemini Omni models introduces filter-eligibility risk over the remaining weeks. I therefore land near the midpoint of the two forecasts, discounting the proxy anchor by roughly 10pp.

Pipeline Timing

Total pipeline time: 301.1s

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