Prolepsia forecastYes / no

Will Google have the best AI model at the end of December 2026?

As of 15 November 2026, Prolepsia puts the chance of yes at 73%.

73%chance of yes

Yes 73%No 27%
Prolepsia forecast made

How Prolepsia reasons

The report Prolepsia wrote with this forecast.

How forecasts are made · Track record

TL;DR

The forecast is 72.8% Yes and 27.2% No. Google is favored to finish first on the specified style-control-off Text Arena leaderboard, supported by a substantial lead and a second Google model ahead of every published rival.

Key evidence

  • Google has two layers of protection. The latest accessible qualifying edition, dated November 11, shows Gemini 4 Carbon High at 1538 ±10 from 4,273 cumulative votes, Gemini 4 Argon High at 1529 ±5 from 20,779, and Anthropic’s Opus 5.5 High at 1514 ±6 from 15,793. Carbon is preliminary, but a Carbon-specific setback would not automatically remove Google from first place (raw Text leaderboard).
  • The resolution setting matters. The market uses score-ordered Rank with style control off—not the default adjusted standings or uncertainty-based Rank Spread. Overlapping confidence intervals do not themselves create a resolving tie (ranking methodology).
  • Carbon’s availability risk has eased. Google announced broad access on November 10. That reduces concern about restricted availability affecting its continued listing; it does not represent an additional capability gain (Google announcement).
  • The strongest challenge is already real. Anthropic released Fable 5.2 on November 9. It leads the retrieved WebDev and Agent evaluations, but those measure different things from overall Text preference (Anthropic announcement, WebDev, Agent).

Uncertainties

Fable 5.2’s first mature, style-control-off Text score is the most valuable missing observation. Its absence from the November 11 edition is not evidence that it performs poorly or cannot qualify before year-end.

The forecast date is November 15, so the observed standings are not a same-day measurement. Fresh votes, score revisions, listing changes, and new releases remain live risks. Meta’s Wonder Full and OpenAI’s internal progress add competitive uncertainty, while Google also has time to improve or regain a lost lead. The market concerns ownership at the final check, not uninterrupted leadership between now and then.

Read the full report

TL;DR

The forecast is 72.8% Yes and 27.2% No for Google to own the first-place model at the December 31, 2026, noon Eastern check. Google holds the top two places in the latest accessible qualifying Text Arena edition, with both models ahead of the strongest published rival (raw Text standings). The decisive unresolved threat is Anthropic’s released Fable 5.2, whose strong results elsewhere do not establish its rank on the resolving board (Anthropic announcement).

Context

“Best AI model” has a narrow meaning in this market. Resolution uses the model owner at the top of the specified Text leaderboard, with style control off, at the contractual check. Published Rank comes first, followed by the underlying Arena score and then the company-name tiebreaker. This is not a contest over general intelligence, coding benchmarks, agent performance, or which company hosts the model. Arena distinguishes score-based Rank from uncertainty-based Rank Spread (ranking methodology).

The forecast is dated November 15, 2026. The latest accessible qualifying edition is November 11, leaving a gap between the observation and the forecast date. Google leads that edition through Carbon and Argon, while Anthropic’s newly released Fable 5.2 has results in other categories but no entry in the checked Text data (raw Text standings, official Text archive).

Evidence

The historical backbone favors persistence, but it is thin. Arena’s official archive covers 257 publication dates from May 8, 2023 through November 11, 2026. It contains 12 editions in 2023, 70 in 2024, 85 in 2025, and 90 in 2026. Across the 2025–2026 snapshots, company leadership changed six times: four in 2025 and two in 2026. Companies have therefore held first place much longer than individual model versions have (historical archive).

One approximately 46-day comparison found Google still leading at 28 of 35 eligible endpoints, a historical retention rate of 80%. That is a descriptive reference point, not the probability for this market. The observations overlap heavily and are concentrated in very few leadership spells. Different endpoint-matching rules also change the sample. The history supports incumbent strength without establishing a precise, stable retention rate (historical archive).

Today’s direct evidence is stronger than that broad reference class. The November 11 raw Text edition places Google first and second:

Rank Model Displayed rating Cumulative votes
1 Gemini 4 Carbon High 1538 ±10 4,273
2 Gemini 4 Argon High 1529 ±5 20,779
3 Claude Opus 5.5 High 1514 ±6 15,793
4 Claude Fable 5.1 Max 1512 ±5 20,511
5 Claude Opus 5 Max 1507 ±5 35,670
6 Muse Wonder Max 1503 ±6 12,696

These are Arena rating points, not percentages. The vote counts are cumulative comparisons, not votes collected on the edition’s publication day (raw Text leaderboard).

The underlying scores put Carbon 24.720121 points ahead of the strongest published non-Google model and Argon 15.801528 points ahead. Google’s advantage is not riding on Carbon alone. A correction or withdrawal affecting only Carbon would leave Argon ahead of the currently scored opposition. Carbon’s preliminary status still matters, and neither model’s rating is a floor (official Text data).

The scoring setting also changes the picture. The default adjusted table gives different scores and a smaller Google advantage. Those standings do not govern this market. The explicit raw-scoring view matches the non-style-controlled Text data, and score uncertainty does not replace the published ordering used for resolution (raw standings, default standings, ranking rules).

Google’s November 10 broad-access announcement reduces another risk: Carbon remaining a restricted release whose availability complicates continued listing. This is evidence about access and listing durability, not a fresh improvement over the Carbon model already on the board (Google announcement).

The main counterweight is Anthropic. Fable 5.2 launched on November 9. Its announcement reports clearer communication and fewer safeguard blocks in early application testing. Those claims suggest plausible routes to better human preference, but they are vendor-reported results, not Text Arena measurements. Anthropic also continues to recommend Opus 5.5 for most everyday coding and knowledge work, which argues against assuming that Fable automatically dominates ordinary conversation (Anthropic announcement).

Independent category results make the threat concrete. In the retrieved November 11 WebDev standings, Fable 5.2 scores 1838 ±21 against Carbon’s 1768 ±15. In Agent Arena, Fable records 30.48% net improvement against Carbon’s 16.90%. The Agent measure is not a Text win rate, and WebDev points cannot be converted into Text points. The inference is narrower: Fable is a serious competitor, and its missing qualifying score is consequential (WebDev standings, Agent standings).

Other competitive paths remain open. Meta’s released Wonder Full appears in other Arena categories but is absent from the checked Text data. OpenAI’s November 3 engineering report describes a newer internal model outperforming Astra on BenchCAD, but gives neither a general release commitment nor evidence of winning conversational preference. These are reasons to retain challenger risk, not evidence that either company has already displaced Google (Arena changelog, OpenAI engineering report).

Google also remains an active competitor. Its Argon announcement on September 30 and Carbon announcement on October 28 demonstrate a recent ability to replace its own leader. That supports allowing for defense or recovery before settlement. It does not establish a scheduled successor, and no verified post-Carbon flagship commitment underpins the forecast (Argon announcement, Carbon announcement).

What's non-obvious

The key distinction is between beating Carbon and beating Google at the final check. A rival can overtake today’s leader temporarily without winning the market. Google can improve, regain first place, or retain it through Argon after a Carbon-specific setback. Conversely, Google does not need another release to win: the existing qualifying field already puts it ahead. This makes the endpoint ownership question different from a prediction about the next impressive launch.

The other distinction is between missing evidence and negative evidence. Fable’s absence from the checked Text edition leaves its relevant performance unknown. It does not erase its demonstrated strength elsewhere, and it does not prove that a Text result is imminent. Category-specific publication happens on different schedules. The right reading is a strong observed Google position paired with a live, unresolved challenger—not a settled race and not an already-observed Anthropic takeover (Arena changelog).

Uncertainties

The largest gap is Fable 5.2’s first mature raw overall Text result, including its configuration, vote count, and interval. That observation would replace indirect evidence with a measurement that directly bears on settlement. Until then, the transfer from technical and agent strength to conversational preference remains uncertain.

The current standings also need refreshing. November 11 is the latest accessible edition, not a verified November 15 roster. Carbon remains preliminary, and fresh comparisons, revisions, endpoint changes, or listing decisions can alter the board. Historical archive comparisons do not eliminate those risks, especially when methodology and sampling practices change over time (raw Text leaderboard, historical archive).

Finally, release timing and leaderboard admission are separate hurdles. A strong model must become eligible, receive enough evaluation, and lead at the actual check. There is no verified public evaluation queue establishing every pending entrant. The evidence therefore supports Google as the clear favorite while leaving substantial room for a qualifying rival breakthrough before settlement.

What changed since the last forecast

This forecast is not compared to the previous run because the answer shape, model stack, or engine settings changed. Treat it as a fresh baseline.

Forecast history

How Prolepsia’s forecast has moved10 forecasts, from 57% on 6 Oct 2026 to 73% on 15 Nov 2026.6 Oct 2026: 57%9 Oct 2026: 60%11 Oct 2026: 60%16 Oct 2026: 63%21 Oct 2026: 63%26 Oct 2026: 59%31 Oct 2026: 73%5 Nov 2026: 73%10 Nov 2026: 71%15 Nov 2026: 73%
DateForecast
73% chance of yes
71% chance of yes
73% chance of yes
73% chance of yes
59% chance of yes
63% chance of yes
63% chance of yes
60% chance of yes
60% chance of yes
57% chance of yes

Sources Prolepsia read

  1. raw Text standingsarena.ai
  2. Introducing Claude Fable 5.2 and Claude Mythos 5.2 \ Anthropicwww.anthropic.com
  3. Text leaderboardlmarena.ai
  4. text/ · lmarena-ai/leaderboard-dataset at mainhuggingface.co
  5. Gemini 4 Carbon is now availableblog.google
  6. Accelerating engineering for the physical world | OpenAIopenai.com
  7. Muse Wonder: A Frontier Model For Personal & Professional AI Agentsresearch.meta.ai
  8. OpenAI shelves new AI model release over safety concerns - AOLwww.aol.com
  9. Claude Platform release notes - Claude Platform Docsplatform.claude.com
  10. openai-python/src/openai/types/shared/chat_model.py at 5b43cceea4a25bc1b9ed1b08f087408b47981083 · openai/openai-python · GitHubgithub.com

The question

Imported from Polymarket. This yes/no question comes from a single Polymarket market.

How it resolves

This market will resolve according to the company which owns the model which has the highest arena rank based off the Chatbot Arena LLM Leaderboard (https://lmarena.ai/) when the table under the "Leaderboard" tab is checked on December 31, 2026, 12:00 PM ET. Results from the "Rank" section on the Leaderboard tab of https://lmarena.ai/leaderboard/text with the style control off will be used to resolve this market. Models will be ordered primarily by their leaderboard rank at the market’s check time. If two or more models are tied on rank, they will be ordered by their Arena score, including any underlying, unrounded, granular values reflected in the data below the leaderboard. If a tie remains, alphabetical order of company names as listed in this market group will be used as a final tiebreaker (e.g., if the two models are tied by exact arena score, “Google” would be ranked ahead of “xAI”). This market will resolve based on the company that occupies first place under this ranking system. The resolution source for this market is the Chatbot Arena LLM Leaderboard found at https://lmarena.ai/. If this resolution source is unavailable at check time, this market will remain open until the leaderboard comes back online and resolve based on the first check after it becomes available. If it becomes permanently unavailable, this market will resolve based on another resolution source.