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

date	predicted	market price	confidence
2026-09-04	80%	73%	49%

1. Planner Fable 5.1 · Fermi decomposition + tool plan

Tools chosen: polymarket_direct polymarket_related kalshi_related claude_news claude_news article_search gdelt_news wikipedia code_execution

Sub-questions (Fermi decomposition)

Which company's model currently holds the #1 Overall score on LiveBench (latest release, category 'All'), and what is the point gap between #1 and the best Anthropic model?
What major frontier model releases (Anthropic Claude, OpenAI GPT, Google Gemini, xAI Grok, others) are expected or rumored between now and September 30, 2026, and what has each lab's recent LiveBench performance been?
Historically, how often has Anthropic held #1 on LiveBench Overall, and how long do leadership periods last before another lab overtakes?
How frequently does LiveBench publish new releases/question refreshes, and do those refreshes tend to reshuffle the top of the leaderboard (e.g., by changing category weights or adding new models)?
What do the sibling Polymarket markets (OpenAI, Google, xAI, Other) imply after normalization, and do they sum consistently with the Anthropic price?
Are there any recent reports of LiveBench evaluating Anthropic's latest models late or excluding certain models (e.g., due to cost or API access), which could affect what appears on the leaderboard at check time?

Planner reasoning

This is a leaderboard-snapshot question about LiveBench Overall on Sept 30, 2026, where the key drivers are (a) who currently tops LiveBench and by what margin, (b) expected model releases from Anthropic, OpenAI, Google, xAI, and others before the check date, and (c) how LiveBench's release cadence/methodology changes could reshuffle rankings. The Polymarket price is the primary anchor, with sibling markets (OpenAI, Google, xAI, Other) giving a normalized cross-check.

2. Research Parallel tool calls

tool	status	time	items	summary
polymarket_direct	OK	0.5s	1	## This Market's Polymarket Data Will Anthropic have the best AI model on LiveBench (Overall) at the end of September 2026? - Current price (probability): 88.00% - 7-day price change: -1.00% - 30-day price change: +35.50% - Total volume: $18,228 (USD notional) - Price range: 27.00% - 94.00% - D
polymarket_related	OK	3.3s	0	Scanned 100 active Polymarket markets, kept 0 matches. keyword 'LiveBench': 0 markets \| keyword 'best AI model': 0 markets \| keyword 'Anthropic': 0 markets \| keyword 'Claude': 0 markets
kalshi_related	OK	3.1s	1	1 related markets / summaries. keyword 'LiveBench': no matches \| keyword 'best AI model': no matches \| keyword 'Anthropic': ok
claude_news	OK	36.1s	14	Note: Search results included several low-quality/SEO-aggregator sites that appear to contain fabricated or unreliable model names (e.g., "Claude Fable," "GPT-6 Astra," "Muse Spark") — treat specific score figures with caution. Findings below prioritize the LiveBench.ai primary source and cross-chec
claude_news	OK	24.4s	10	Here are key findings on the state of frontier models and LiveBench standings relevant to this question: LiveBench standings (as of Sept 2026): - Claude Fable 5.1 currently leads the published LiveBench snapshot with 83.4% livebench overall score as of a June 25, 2026 snapshot (source: benchl
article_search	OK	0.6s	28	Fetched 28 unique articles across 3 queries (lookback=45d). 'LiveBench leaderboard top model': 10 articles \| 'Anthropic Claude new model release benchmark': 10 articles \| 'GPT Gemini Grok new model release LiveBench score': 10 articles
gdelt_news	OK	156.4s	10	GDELT: 10 articles across 3 queries (lookback=30d). 'LiveBench Anthropic Claude': error GDELT rate-limited after retries (429) \| 'Claude new model release 2026': error GDELT rate-limited after retries (429) \| 'Gemini GPT Grok frontier model release': 10 hits
wikipedia	OK	0.2s	3	Fetched 3 Wikipedia entries (0 missing pages).
code_execution	OK	21.6s	0	Note: Actual live Polymarket quotes weren't included in the prompt, so the figures below use a representative placeholder snapshot (raw YES prices roughly consistent with recent market chatter) to demonstrate the de-vig methodology. Swap in real quoted prices to get the final number. - **Raw si

3. Evidence Brief Sonnet 5 · 8079 chars

# Current state LiveBench's latest published release (2026-06-25) reportedly shows an Anthropic model ("Claude Fable 5.1") in the #1 Overall slot, per secondary aggregator mirrors (benchlm.ai) — LiveBench.ai's own site was not directly quoted with a numeric table. Since late June 2026, competitive pressure has intensified: OpenAI's GPT-6 Astra (launched Sept 3, 2026) has reportedly taken the LiveBench Reasoning sub-category, while Anthropic's Claude Fable 5.1 still reportedly leads Mathematics — meaning Overall leadership is close and could tip either way before the Sept 30, 2026, 12:00 PM ET check. LiveBench's stated ~6-month refresh cadence suggests the next official refresh lands near Dec 2026, so the June 25 snapshot (or an unannounced interim update) will likely be what's checked at resolution. # Timeline of key events - 2024 (mid): Claude 3.5 Sonnet debuts as #1 on original LiveBench Overall, ~6% ahead of the field (confirmed, LiveBench paper). - 2025-11-24: Anthropic releases Claude Opus 4.5, first to break 80% SWE-bench Verified (reported). - 2026 (through mid-year): Anthropic iterates Opus 4.6 → 4.8 (reported, no specific dates given). - 2026-05: Anthropic valued at $965B in Series H round (confirmed, Wikipedia). - 2026-06-25: Latest confirmed LiveBench "release" date; Anthropic model reported #1 Overall (~83.4%) per benchlm.ai mirror (reported, not independently verified on livebench.ai itself). - 2026-07-24: Claude Opus 5 launches, reportedly topping boards on deep reasoning/agentic work (reported). - 2026-08-12: Grok 4.6 launches — positioned as value/mid-tier, not a leaderboard topper (reported). - 2026-08-2x: Google Gemini 3 Pro/Flash Preview lead LiveCodeBench (coding-specific, not Overall) (reported). - 2026-09-01: Anthropic launches "Claude Fable/Mythos 5.1" (reported). - 2026-09-03: OpenAI launches GPT-6 Astra; reportedly takes LiveBench Reasoning (92.65 vs Claude's 91.65) and ARC-AGI-3 (reported, single aggregator source, name unverified against official OpenAI naming). - 2026-09-05: Artificial Analysis publishes Intelligence Index v4.2, showing Claude Fable 5.1 #1 (confirmed by GDELT headline; details reported). # Event Will Anthropic's Claude hold the #1 spot on LiveBench's Overall leaderboard when checked on Sept 30, 2026 at 12:00 PM ET? # Outcomes to forecast Yes / No # Kalshi market anchor No direct Kalshi order-book price was returned by kalshi_direct in this research pass. The only quantitative anchor available is the **Polymarket price for this identical ticker: 88% YES**, up sharply from a 30-day range low of 27% (30d change +35.5%), essentially flat over 7 days (-1pp), on thin volume (~$18.2K total, 50 data points). This is a large, recent repricing toward "Yes," likely driven by Anthropic's Opus 5 / Fable 5.1 releases (Jul–Sep 2026). # Sub-question answers 1. **Current #1 holder / gap** — Reported (benchlm.ai mirror of LiveBench's June 25, 2026 release): Anthropic's Claude Fable 5.1 leads Overall at 83.4%, Claude Fable 5 second (83.0%), Muse Spark 1.3 third (81.6%) — gap to #2 is Anthropic-internal (~0.4pt); gap to nearest non-Anthropic competitor ~1.8pt. Not independently confirmed on livebench.ai's live table. 2. **Frontier releases/performance** — Anthropic: Opus 4.5→4.8→Opus 5 (Jul 24)→Fable 5.1 (Sep 1), reported strong across LiveBench Overall/Math and Artificial Analysis indices. OpenAI: GPT-6 Astra (Sep 3) reportedly leads LiveBench Reasoning and ARC-AGI-2/3. Google: Gemini 3 Pro/Flash strong on LiveCodeBench (coding), not confirmed Overall leader. xAI: Grok 4.6, described as value/mid-tier, not a leaderboard topper. 3. **Historical #1 frequency** — Claude 3.5 Sonnet held #1 at LiveBench's 2024 launch; qualitative reporting suggests Anthropic has "strong track record" of Overall leadership, but no rigorous month-by-month count was found in primary sources (the "4 of 12 months" figure from code_execution is explicitly a placeholder/illustrative example, not real data — do not treat as fact). 4. **Refresh cadence/reshuffling** — LiveBench states ~6-month refresh cycle (23 tasks, 7 categories); last confirmed release June 25, 2026, implying next scheduled refresh ~Dec 2026, after this market's close — meaning the Sept 30 check likely uses the June release or an unscheduled interim update, reducing (but not eliminating) reshuffle risk before resolution. 5. **Sibling Polymarket markets** — No real sibling data retrieved; polymarket_related found 0 matching markets. The code_execution figures (Anthropic 34.3% de-vigged vs. this ticker's actual 88%) are illustrative placeholders and starkly inconsistent with the real Polymarket price — treat as unreliable, not a genuine cross-check. 6. **LiveBench evaluation lag/exclusion of Anthropic models** — No evidence found of exclusion; if anything, LiveBench's comparison view reportedly features Anthropic models prominently as of the June 2026 release. # Key facts (high-confidence, factual) 1. [polymarket_direct] This exact ticker trades at 88% YES on Polymarket, up from 27% a month ago. 2. [Wikipedia] Anthropic valued at $965B (May 2026 Series H); releases Claude in Haiku/Sonnet/Opus/Fable tiers plus restricted Mythos. 3. [claude_news/benchlm.ai] Latest LiveBench release dated 2026-06-25 reportedly shows an Anthropic model #1 Overall. 4. [claude_news/felloai] GPT-6 Astra (OpenAI, Sept 3 2026) reportedly leads LiveBench Reasoning sub-category, not confirmed Overall. 5. [claude_news] LiveBench's stated refresh cadence is ~6 months, implying no new official leaderboard update before Sept 30, 2026 close. # Cross-market signals - Kalshi related: "OpenAI or Anthropic IPO first — Anthropic" trades 91% (unrelated to this market but signals general market confidence in Anthropic's trajectory). - Polymarket (same ticker): 88% YES, strong 30-day uptrend. - Sibling Polymarket markets (OpenAI/Google/xAI/Other): not found/retrieved; no valid cross-check available. # Analyst opinions and speculation - Aggregator sites (benchlm.ai, felloai.com) suggest leadership is fragmented by sub-category — Anthropic strongest on Math/Overall aggregate, OpenAI strongest on Reasoning/ARC-AGI, Google strongest on coding-specific (non-LiveBench-Overall) benchmarks. - Multiple model names ("GPT-6 Astra," "Claude Fable," "Muse Spark") are unverified against official lab announcements — flagged explicitly by research as possibly unreliable/SEO-generated naming; treat scores with caution. # Directional lean per outcome - **Yes (Anthropic)**: Supported by reported June 2026 LiveBench snapshot showing Anthropic #1, continued rapid Anthropic releases (Opus 5, Fable 5.1), Polymarket pricing at 88%, and historical benchmark strength since 2024. Opposing: OpenAI's GPT-6 Astra reportedly gaining in Reasoning just weeks before close; narrow point gaps (~1.8pt) mean a single OpenAI/Google release could flip Overall. - **No (other labs)**: Supported by GPT-6 Astra's reasoning gains and fragmented category leadership; Google's coding strength. Opposing: no confirmed reports of any non-Anthropic model actually topping Overall (only sub-category leads reported); LiveBench's slow refresh cycle reduces chance of a new ranking capturing a late OpenAI surge before Sept 30 check. # Gaps / unknowns - No direct livebench.ai table data retrieved (only third-party mirrors); scores may be stale or mislabeled. - No verified sibling Polymarket market data (OpenAI/Google/xAI/Other) to cross-check Anthropic's implied probability. - Model naming ("Claude Fable," "GPT-6 Astra") unconfirmed against official company announcements — could reflect confused/hallucinated aggregator content. - Unclear whether LiveBench will publish an interim update before Sept 30 close. # Calibration anchors - Polymarket YES price for this exact market: 88% (primary quantitative anchor; no separate Kalshi price available). - Precedent: Claude 3.5 Sonnet led LiveBench decisively at 2024 launch; Anthropic reported to still hold Overall #1 as of the most recent (June 2026) release despite OpenAI's September reasoning-category gains.

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

claude-fable-5-1

35% conf

Yes	80%
No	20%

The primary quantitative anchor is Polymarket at 88% YES for this identical ticker, up sharply after Anthropic's reported Opus 5 / Fable 5.1 releases, and the most recent LiveBench snapshot (June 25, 2026, via mirror) reportedly shows Anthropic #1 Overall with a ~1.8pt lead over the nearest non-Anthropic model. However, OpenAI's GPT-6 Astra (Sept 3) reportedly already leads the Reasoning sub-category, the Overall gap is narrow, and LiveBench does add new models to its leaderboard between question refreshes, so a late flip is plausible. Much of the brief's model naming and scoring is unverified or possibly SEO/hallucinated content, so I shade modestly below the market rather than fully trusting the 88% on thin volume (~$18K).

gpt-5.6-sol

64% conf

Yes	86%
No	14%

The reported June 25 snapshot points toward Yes because Anthropic leads Overall by roughly 1.8 points over the nearest non-Anthropic competitor. Recent releases also modestly favor Yes: Anthropic reportedly remains strong across Overall and Mathematics, while OpenAI and Google have only demonstrated leadership in narrower categories. LiveBench's approximately six-month refresh cadence further favors the incumbent, since the next scheduled refresh is likely after resolution. I anchor near the identical Polymarket ticker's 88% YES price but reduce to 86% because the current table and model names rely heavily on unverified secondary aggregators, and an interim update could incorporate OpenAI's recent gains.

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

Both forecasters anchor heavily on Polymarket's 88% and converge to 0.80–0.86, but neither seriously stress-tests the "thin volume ~$18K, 50 data points" caveat explicitly flagged in the brief — a 30-day move from 27%→88% on such thin liquidity could reflect noise/manipulation rather than robust information, warranting more downward adjustment than either applied.

Neither forecast adequately weights sub-question 3 (historical #1 frequency), which the brief explicitly warns is unverified/placeholder data ("4 of 12 months... do not treat as fact") — yet both implicitly lean on "Anthropic has strong track record" language without flagging that this reference class is essentially unsupported.

Both treat the June 25 snapshot as the likely resolution basis due to the ~6-month refresh cadence, but neither explores the scenario where LiveBench publishes an unscheduled interim update incorporating GPT-6 Astra's Reasoning gains before Sept 30 — the brief notes this explicitly as a live possibility ("reducing but not eliminating reshuffle risk"), and it's the single biggest tail risk for a flip.

Neither forecaster discusses the internal Anthropic model-naming inconsistency (Claude Fable 5.1 vs Fable 5 vs Opus 5 vs "Mythos") or what happens if LiveBench's table treats these as separate/different entries — if aggregation across Anthropic's own multiple entries is what produces the "#1 and #2" internal gap, a single consolidated non-Anthropic model could leapfrog more easily than the point-gap suggests.

Both forecasts settle in a narrow 0.80–0.86 band with little daylight between them despite F1 stating "shade modestly below the market" while landing at 0.80 (8pp below) and F2 landing at 0.86 (2pp below) — this is a case where the two forecasters agree qualitatively but their quantitative confidence levels (0.35 vs 0.64) diverge sharply without clear justification for why F2 is nearly twice as confident given they cite the same unverified evidence.

Neither forecast explicitly addresses the possibility that LiveBench's own live table (never directly retrieved) could already differ from the benchlm.ai mirror — given the brief's repeated caveat "not independently verified on livebench.ai itself," this is a structural verification gap that should push uncertainty higher than a simple 12-14% No allocation reflects.

6. Aggregator Fable 5.1 · final calibrated forecast

83% (50% confidence)

Yes: 83% No: 17%

The best available anchor is the identical Polymarket ticker at 88% YES, corroborated directionally by the reported June 25 LiveBench snapshot showing Anthropic #1 Overall with a ~1.8pt lead over the nearest non-Anthropic model, and by the ~6-month refresh cadence that makes a pre-resolution reshuffle less likely. The critique correctly notes the thin liquidity (~$18K) behind the 27%→88% move, the reliance on unverified aggregator mirrors with suspicious model naming, and the live possibility of an interim LiveBench update incorporating GPT-6 Astra's Reasoning gains — these warrant shading below the market more than F2 did but not as far as a coin-flip. No source reports any non-Anthropic model actually topping Overall, only sub-category leads, so the incumbent advantage stands. I land at 0.83, between the two forecasts and modestly below the market to reflect structural verification gaps.

Pipeline Timing

Total pipeline time: 252.6s

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