e·EvalMapAI CAPABILITY ATLAS
As of 14 November 2026

Past & future outlook.

A quantitative look back at 176 model releases across 85 days (2026-08-21 to 2026-11-13), 15 closed forecast weeks, 337 scored probability predictions and multi-month benchmark trajectories — paired with Q4 2026 to Q1 2027 pipeline odds, lab cadence windows and statistical trend extrapolations.

Historical numbers are measured directly from EvalMap’s sealed sources. Future values combine live open forecasts with transparent statistical extrapolations of past cadence and Epoch ECI slopes — never fabricated benchmark scores.

Release cadenceForecast calibrationCapability velocityNear-term pipelineQ1 2027 outlook

Release tempo176 releases2.07/day · 58 open weights (33.0%)
Odds Brier score0.153vs markets (164 pairs): 0.164 vs 0.202
Frontier velocity (13 mo)+1.45 ECI/mo153.0 → 173.4 (R²=0.95; open +1.43/mo)
Q1 2027 trend (Mar)~177.3 ECI95% band 174.8–179.9 · open ~167.6

The past: release cadence & surprise rate

Across 85 tracked days, the frontier ecosystem shipped more than two releases per day, with open weights accounting for one-third of all launches and roughly four in five releases landing without a public week-ahead rumor.

Monthly volume & modality mix

Language leads (113, 64.2%), followed by speech (33), image (16) and video (14). Open weights: 58/176 (33.0%), including 46/113 language models.

MonthDaysReleasesRate
2026-08 (partial)11d121.09/d
2026-0930d662.20/d
2026-1031d672.16/d
2026-11 (partial)13d312.38/d

Weekly forecast ledger & likelihood accuracy

Over 15 closed weeks, 21 of 64 weekly-named cards (32.8%) shipped inside their named 7-day window, and 19 shipped later. Meanwhile 164 of 204 total releases (80.4%) arrived with no prior weekly card.

ConfidenceTargeted week hitsAll weekly appearances
high4/4 (100.0%)22/151 (14.6%)
medium9/25 (36.0%)10/26 (38.5%)
low8/35 (22.9%)8/46 (17.4%)

Implication: targeted high calls hit 100% on time, yet ~4 in 5 total launches ship without a week-ahead rumor.

Observed lab cadence (makers with ≥3 releases)

MakerTotalOpenMedian gapMean gapLatestNext projected window
Alibaba Qwen1024.5d9.6d2026-11-112026-11-21
inclusionAI (Ant Group)966.0d10.9d2026-11-052026-11-16
Google DeepMind806.0d11.1d2026-11-122026-11-23
OpenAI808.0d8.9d2026-11-042026-11-22
Tencent Hunyuan859.0d11.0d2026-11-132026-11-24
Meta714.0d13.8d2026-11-092026-11-23
Anthropic6015.0d17.2d2026-11-092026-11-26
NVIDIA666.0d8.8d2026-10-282026-11-15
Microsoft AI5026.0d18.0d2026-10-272026-11-14
Baidu413.5d3.5d2026-11-122026-11-16
DeepSeek4310.0d19.0d2026-10-172026-11-24
Z.ai4221.0d15.7d2026-10-122026-11-28

Forecast calibration & market edge

Across 337 scored daily predictions, EvalMap averaged Brier 0.153 (and beat markets 0.164 vs 0.202 on the 164 overlapping pairs). The edge comes almost entirely from avoiding hype spikes on thin prediction markets, while reliability bins show a systematic conservative bias in the middle probability range.

Reliability curve (predicted vs actual)

0%0%25%25%50%50%75%75%100%100%0–20%: pred 6.2%, actual 10.7% (n=75)20–40%: pred 29.1%, actual 50.7% (n=67)40–60%: pred 46.6%, actual 70.0% (n=50)60–80%: pred 67.9%, actual 89.7% (n=78)80–100%: pred 87.9%, actual 100.0% (n=67)
BinNPredActualBiasBrier
0–20%756.2%10.7%+4.5%0.084
20–40%6729.1%50.7%+21.7%0.303
40–60%5046.6%70.0%+23.4%0.246
60–80%7867.9%89.7%+21.9%0.149
80–100%6787.9%100.0%+12.1%0.017

Head-to-head by market source

On liquid Polymarket contracts EvalMap essentially ties the market, whereas on Manifold contracts EvalMap gains −0.190 Brier by anchoring to maker docs rather than unverified Reddit/X rumors (such as Qwen4 in Oct).

SourceNEvalMapMarketEdge
Manifold320.0880.278-0.190
Polymarket1320.1820.183-0.001

LOO calibration & log-odds recalibration: in the 20–80% middle band (n=195), actual ship rates averaged 71.3% vs 49.1% predicted (+22.2 pp). Regularized log-odds shift (a=+1.02) lowers leave-one-question-out out-of-sample Brier from 0.153 to 0.132 (-0.021; 2D Platt 0.140).

Brier score by question

QuestionNEvalMapMarketResolution
GPT-6 Terra (OpenAI)300.4600.407OpenAI GPT-6.1 Terra (2026-10-26)
A new MiniMax model (MiniMax)90.383—MiniMax MiniMax-M3.1-Flash-Preview (2026-09-27)
Muse Spark open weights (Meta)270.304—Meta Muse Spark Open (2026-10-15)
Muse Spark 1.4 or a newer Muse Spark (Meta)160.2530.378Expired / open
Kimi K3.1 (Moonshot AI)300.148—Moonshot AI Kimi K3.1 Pro (2026-10-23)
A new OpenAI model (DevDay is 29 Sep) (OpenAI)150.143—OpenAI GPT-6.1 Sol (2026-09-29)
Claude Sonnet 5.5 (Anthropic)150.1210.058Anthropic Claude Sonnet 5.5 (2026-09-28)
The Claude model after Fable 5.1 (Anthropic)300.1130.365Anthropic Claude Fable 5.2 (2026-11-09)
Grok 4.8 (xAI)160.1050.143Expired / open
Qwen4, any of the series (Alibaba)300.0990.413Alibaba Qwen Qwen4-Max-Preview (2026-11-11)
Gemini 4, any tier (Google)240.0880.019Google DeepMind Gemini 4 Argon (2026-10-13)
Watermelon (Meta)240.0690.035Meta Muse Wonder and Muse Wonder Full (Watermelon) (2026-10-12)
Step 5 open weights (StepFun named 15 Oct) (StepFun)270.035—StepFun Step 5 Preview open weights (2026-10-15)
DeepSeek V4.1-Pro (DeepSeek)160.0340.038Expired / open
Claude Haiku 5.5 (Anthropic)180.0190.005Anthropic Claude Haiku 5.5 (2026-10-07)
Mistral Large 4 open weights (Mistral AI)30.016—Mistral AI Mistral Large 4 (2026-10-29)
FastH3-Omni open weights (FastVideo named 20 Oct) (FastVideo (Hao AI Lab))30.010—FastVideo (Hao AI Lab) FastH3-Omni open weights (2026-10-20)
MiMo-V3 (Xiaomi)10.010—Expired / open
Nemotron 3.5 Bolt E4B (NVIDIA named 23 Oct) (NVIDIA)30.002—NVIDIA Nemotron 3.5 Bolt E4B (2026-10-23)

Capability velocity: closed vs open frontier

Over 13 months of Epoch ECI (2025-11 to 2026-11), both the overall frontier (+1.45 pts/mo, R²=0.95) and the open-weight frontier (+1.43 pts/mo, R²=0.95) climbed on nearly parallel linear tracks, keeping the general capability gap near 9–12 ECI points (~6.5 months) even as open weights caught up to lead Terminal-Bench.

Epoch ECI cumulative frontier & Q1 2027 statistical extrapolation

14515516517525-1126-0226-0526-0826-1127-0127-03ObservedLinear trend2025-11 · All frontier: 153.0 ECI2025-11 · Open weights: 145.4 ECI2025-12 · All frontier: 157.3 ECI2025-12 · Open weights: 146.2 ECI2026-01 · All frontier: 157.3 ECI2026-01 · Open weights: 148.8 ECI2026-02 · All frontier: 157.3 ECI2026-02 · Open weights: 148.8 ECI2026-03 · All frontier: 160.2 ECI2026-03 · Open weights: 148.8 ECI2026-04 · All frontier: 162.7 ECI2026-04 · Open weights: 151.6 ECI2026-05 · All frontier: 162.7 ECI2026-05 · Open weights: 151.6 ECI2026-06 · All frontier: 164.1 ECI2026-06 · Open weights: 153.0 ECI2026-07 · All frontier: 164.7 ECI2026-07 · Open weights: 158.6 ECI2026-08 · All frontier: 164.7 ECI2026-08 · Open weights: 158.6 ECI2026-09 · All frontier: 169.0 ECI2026-09 · Open weights: 158.6 ECI2026-10 · All frontier: 169.8 ECI2026-10 · Open weights: 161.4 ECI2026-11 · All frontier: 173.4 ECI2026-11 · Open weights: 161.4 ECI2026-12 (trend) · All: ~173.02026-12 (trend) · Open: ~163.32027-01 (trend) · All: ~174.42027-01 (trend) · Open: ~164.82027-02 (trend) · All: ~175.92027-02 (trend) · Open: ~166.22027-03 (trend) · All: ~177.32027-03 (trend) · Open: ~167.6177.3167.6
All models (observed & OLS trend)Open weights (observed & OLS trend)±1.96×RMSE band (±2.6 pts)

Statistical linear extrapolation from historical frontier envelope; not a benchmark measurement.

Rolling ECI backtestWindowsMAERMSE
+1 mo91.341.51
+2 mo81.551.79
+3 mo71.422.00

Artificial Analysis 48-day jump

Between 2026-09-27 and 2026-11-14 (48 days):

MetricStartLatestGainCurrent leader
AA Intelligence v557.664.6+7.0Claude Fable 5.2 (max with fallback)
Terminal-Bench Hard59.6%67.2%+7.6Kimi K3.1 Pro (open)

JEEP 3D & LMArena frontier

JEEP 3D (83 scored builds) tops out at 58/100 (Pareto 26.10 Preview), leaving 42 points of uncaptured agentic 3D coding headroom. LMArena (2026-10-13) tops at 1525 Elo (gemini-4-argon-high).

JEEP top buildsScore
Pareto 26.10 Preview58
Claude Sonnet 5.548
Mistral Large 4 Preview48
MiMo 2.6 Flash46

The future: near-term pipeline (Nov 2026 – Jan 2027)

10 active odds questions and 11 expected cards track the immediate release pipeline through 31 January 2027.

Active odds watchlist

Modelby 30 Novby 31 Decby 31 Jan
DeepSeek V4.1-Pro (DeepSeek)12%
mkt 8%		25%
mkt 18%		65%
mkt 90%
Grok 4.8 (xAI)65%
mkt 54%		88%
mkt 91%		92%
mkt 94%
Muse Spark 1.4 or a newer Muse Spark (Meta)45%
mkt 49%		72%
mkt 77%		85%
no mkt
MiMo-V3 (Xiaomi)8%
no mkt		24%
no mkt		45%
no mkt
Hy4 open weights (Tencent Hunyuan)65%
no mkt		85%
no mkt		92%
no mkt
ERNIE 5.5 Flash open weights (Baidu)82%
no mkt		94%
no mkt		97%
no mkt
Next OpenAI language model (GPT-6.2) (OpenAI)65%
mkt 72%		88%
mkt 93%		95%
no mkt
Hy4 Flash (Tencent Hunyuan)35%
no mkt		65%
no mkt		82%
no mkt
Kardashev-0.7 (Banbury Road)18%
no mkt		48%
no mkt		68%
no mkt
Fern 0.1 (Mirendil)18%
no mkt		45%
no mkt		65%
no mkt

Open expected cards

ModelLikelihoodBucketWindow / signal
Grok 4.8 (xAI)mediumfurther out‘later this week’ (Musk, 10 Nov)
Qwen-Image 3.1 (Alibaba)highfurther outlater this year
DeepSeek V4.1-Pro (DeepSeek)highfurther outday not named
A new OpenAI model (none named) (OpenAI)mediumthis week‘very soon’, none named
MiMo-V3 (Xiaomi)highfurther outannounced 23 Sep, no date
Kardashev-0.7 (Banbury Road)highfurther outannounced 5 Oct, no date
Fern 0.1 (Mirendil)highfurther outannounced 26 Oct, no date
Genie 4 (Google DeepMind)highfurther outannounced 27 Oct, no date
Hy4 open weights (Tencent Hunyuan)highfurther out‘in the coming days’
Hy4 Flash (Tencent Hunyuan)highfurther outannounced 4 Nov, no date
ERNIE 5.5 Flash open weights (Baidu)highfurther outby 25 Nov

Q1 2027 trajectory (through March 2027)

Extending observed Sep–Nov release rates and 13-month capability slopes through Q1 2027 gives baseline volume and capability bands under steady scaling.

Projected release volume (Sep–Oct pace ~2.18/d)

PeriodDaysBaselineRange (1.9–2.4/d)
2026-11 (rest, 17d)17d3732–41
2026-1231d6859–74
2027-0131d6859–74
2027-0228d6153–67
2027-0331d6859–74

Projected Epoch ECI frontier (linear trend ±95% residual band)

MonthAll trend95% bandOpen trendGap
2026-12173.0170.4–175.5163.39.6
2027-01174.4171.9–177.0164.89.7
2027-02175.9173.3–178.4166.29.7
2027-03177.3174.8–179.9167.69.7

If AA’s 48-day pace continues at 0.5×–1.0× slope, Q1 2027 leaders would sit near 74.6–84.6 AA Intelligence and 78.0–88.9% Terminal-Bench Hard.

Lab cadence survival priors (conditional release odds)

Empirical survival probabilities computed by tools/odds.py prior from observed inter-release intervals and elapsed days since each maker’s latest release.

MakerElapsedMean gapby 30 Novby 31 Decby 31 Jan
Alibaba Qwen3d9.6d78%99%99%
Tencent Hunyuan1d11.0d83%99%99%
Google DeepMind2d11.1d66%99%99%
inclusionAI (Ant Group)9d10.9d77%99%99%
OpenAI10d8.9d90%99%99%
Meta5d13.8d61%98%99%
Undisclosed (stealth)6d13.0d69%99%99%
Anthropic5d17.2d69%97%99%
NVIDIA17d8.8d84%99%99%
Microsoft AI18d18.0d75%96%99%