# Which AI is best at predicting the future? The Sikt Bench preview

> Sikt Bench tests how well AI models forecast real events, using only what was knowable at the time. The first preview ranks six models on 232 events.

Nov 7, 2026 · AI forecasting · Sikt Intelligence · https://www.siktintelligence.com/blog/which-ai-is-best-at-predicting-the-future

**In the first Sikt Bench preview, the AI model that came closest to predicting real events was Opus 5.5, with a Sikt Score of 55.9.** The test covered six models and 232 events across geopolitics, finance, the economy, AI and elections. Every model saw only what could be known at the time, with no web access and no market prices. But no model won every category, and the preview is too small to call a final ranking. Here are the results, how the test works, and how to read them.

## The Sikt Bench leaderboard

[Sikt Bench](/bench) (also written SiktBench) is Sikt Intelligence's benchmark for AI forecasting. Preview v0.1 ranks the models by their **Sikt Score**: 100 is a perfect forecast, and 0 is no better than an even split across the possible outcomes.

| Rank | Model | Sikt Score | Chance on the outcome that really happened |
|---|---|---|---|
| 1 | Opus 5.5 | 55.9 | 71.4% |
| 2 | Haiku 5.5 | 50.2 | 69.0% |
| 3 | GPT-6 Astra | 47.9 | 66.8% |
| 4 | GPT-6 Sol | 47.0 | 67.1% |
| 5 | Qwen3.8-27B | 39.2 | 62.3% |
| 6 | GPT-6 Luna | 30.2 | 61.5% |

The score rewards two things: picking the outcome that happened, and putting the right amount of confidence on it. A model that says 90% on everything it believes is punished hard when it is wrong. A model that says 50% on everything scores zero.

The last column is easier to picture: the **chance each model gave to the outcome that really happened**, averaged over every question, whether that chance was high or low. Across the six models it ran from 61.5% (GPT-6 Luna) to 71.4% (Opus 5.5). It counts confidence too: a 90% call that comes true adds more than a 55% one, and a confident miss pulls the average down. The board is ranked by the Sikt Score, which compares each model with an even split on the same questions.

## No model wins everything

The overall order hides big differences by topic. Opus 5.5 led geopolitics & policy and finance; GPT-6 Astra led macro & central banks and elections; Opus 5.5 and GPT-6 Astra tied in AI & compute.

| Category | Opus 5.5 | Haiku 5.5 | GPT-6 Astra | GPT-6 Sol | Qwen3.8-27B | GPT-6 Luna |
|---|---|---|---|---|---|---|
| Geopolitics & policy | **67** | 63 | 38 | 46 | 38 | 1 |
| Finance | **61** | 56 | 55 | 40 | 38 | 36 |
| Macro & central banks | 44 | 38 | **50** | 45 | 37 | 39 |
| AI & compute | **63** | 55 | **63** | 59 | 47 | 52 |
| Elections | 48 | 43 | **50** | 48 | 38 | 39 |

The spread inside a single model is often larger than the gap between models. GPT-6 Luna, last overall, scored 1 on geopolitics but 52 on AI & compute. So "which AI is best at forecasting?" depends on what you are forecasting. A model that reads central bank decisions well can still misjudge a ceasefire.

## How Sikt Bench works

Most AI benchmarks ask questions whose answers are already on the internet. A forecasting test cannot do that: a model that has seen the result can simply remember it. This problem, [lookahead bias](/blog/lookahead-bias-ai-forecasting), ruins many impressive-looking backtests.

Sikt Bench closes that door in four ways:

- **Real, recent events.** Every question is a real event that settled between late September and early November 2026: wars and diplomacy, prices and earnings, economic data and rate decisions, model releases, and elections.
- **A frozen archive.** Each model gets the same archive of news and data, cut off at the date of the forecast. It has no web access, so it cannot look up what happened next.
- **No market prices.** The models never see what traders were paying. Every number on the board is the model's own judgment.
- **A memory check.** Each model is also asked the bare question with no archive. If it does about as well from memory alone, that shows up.

Each event is forecast at several dates before its deadline, from a month out down to two days before (finance from a week out), and the score averages those dates. Related questions about one event, such as the strikes on one stock or the candidates in one race, count as **one vote**, so a busy week of earnings or an election night cannot dominate the result.

Every model gets the same prompt, the same tools and the same archive. Qwen3.8-27B, an open model, runs through a standard chat API instead of a vendor agent, with its thinking capped at 16k tokens per question.

## How the Sikt Score is calculated

The Sikt Score starts from the [Brier score](/blog/brier-score-explained), the standard measure of forecast accuracy: the squared gap between the forecast and what happened, averaged over every question. Lower is better.

**Sikt Score = 100 × (1 − the model's Brier score ÷ the Brier score of an even split).**

An even split is the forecaster who gives every outcome the same chance and learns nothing. Halving its error gives a score of 50. A model that does worse than an even split, by being confidently wrong, gets a negative score.

The categories are weighted, because some kinds of events matter more to the people who use forecasts, and some are much more common in the data:

| Category | Weight | What it covers |
|---|---|---|
| Geopolitics & policy | 35% | Wars, sanctions, diplomacy, US policy |
| Finance | 25% | Prices, earnings, short-dated markets |
| Macro & central banks | 20% | Economic data, rate decisions |
| AI & compute | 10% | Model releases, rankings, compute |
| Elections | 10% | US and international races |

Elections carry 10% even though the 2026 midterms produced many questions: one election night moves dozens of races at once, so they tell us less than their number suggests.

## How to read the preview

The preview is a first look, not a final ranking. Read it with four things in mind:

1. **It is small.** 232 events over about six weeks, and many of them were moved by the same news. A different six weeks could reorder models whose scores are close.
2. **Close scores are close.** The models in the middle of the board are separated by a few points, which is inside what chance alone can produce at this size. A gap like the one between first and last is not.
3. **It is a replay.** The events are real and the information is frozen at the right date, but the forecasts were made after the fact. Live forecasting, where nobody knows the answer yet, is the harder test. That is why Sikt also forecasts in public, as in [the 2026 midterms](/blog/how-sikt-forecast-the-2026-midterms).
4. **Calibration matters as much as hits.** A good forecaster's 70% calls come true about 7 times in 10. More on that in [forecast calibration](/blog/forecast-calibration).

Before a full release, each model will be run several times on the same events to measure how much its score moves by chance, and the method will be reviewed outside Sikt. Sikt Intelligence's own forecaster is not on the board.

## How Sikt Bench compares with other AI forecasting benchmarks

[ForecastBench, Prophet Arena and Metaculus FutureEval](/blog/ai-forecasting-benchmarks) ask models about events that have not happened yet and wait for the answer. That is the cleanest test, but it is slow, and each model sees a different internet.

Sikt Bench adds the other half. Because every model reads the same frozen archive, a new model can be tested as soon as it is released, on events that have already settled, with exactly the information its rivals had. Both kinds of test are needed. A benchmark on settled events answers "which model reads the evidence best?"; a live one answers "does it still work when nobody knows the answer?"

## Can AI predict the future?

Better than chance, clearly. Every model on the board beat an even split from evidence alone, and the best ones cut its error by about half. But AI forecasting is not magic: the scores show large gaps between topics, and even the top score is far from 100. For the wider question, see [can AI predict the future?](/blog/can-ai-predict-the-future) and [AI vs. superforecasters](/blog/ai-vs-superforecasters).

## Key takeaways

- In the Sikt Bench preview, the AI model that forecast real events best was Opus 5.5, with a Sikt Score of 55.9.
- No model won every category, and the gap between topics inside one model is often larger than the gap between models.
- On average, the models gave the outcome that really happened a chance from 61.5% (GPT-6 Luna) to 71.4% (Opus 5.5).
- Every model saw only a frozen archive from the date of the forecast: no web, no market prices.
- The Sikt Score runs from 0 (an even split) to 100 (perfect) and is built on the Brier score.
- It is a preview: 232 events over about six weeks, with close scores that could change order.

## FAQ

### Which AI is best at predicting the future?

In the Sikt Bench preview v0.1, the top model was Opus 5.5, with a Sikt Score of 55.9, out of six models scored on 232 real events. No model won every category, so the best AI depends on what you want to forecast.

### How often is AI right when it predicts the future?

In the Sikt Bench preview, the average chance a model gave to the outcome that really happened ran from 61.5% (GPT-6 Luna) to 71.4% (Opus 5.5), counted over every question, whether that chance was high or low. Sikt Bench ranks the models by the Sikt Score, which compares that accuracy with an even split on the same questions.

### What is Sikt Bench?

Sikt Bench (SiktBench) is Sikt Intelligence's benchmark for AI forecasting. It asks AI models to forecast real events using only a frozen archive of what was known at the time, with no web access and no market prices, and scores them against what actually happened.

### What is a Sikt Score?

A measure of forecast accuracy from 0 to 100. It compares a model's Brier score with the Brier score of an even split across outcomes: 100 is a perfect forecast, 0 is no better than an even split, and a negative score is worse than an even split.

### How does Sikt Bench stop AI models from cheating?

Every question is a recent event, each model gets an archive cut off at the date of the forecast, it has no web access and sees no market prices, and a separate memory check tests whether it already knows the answer.

### Is Sikt Intelligence's own AI on Sikt Bench?

No. The preview ranks public AI models only. Sikt Intelligence's own forecaster publishes its forecasts in public instead, next to the market, as in [the 2026 midterms](/blog/how-sikt-forecast-the-2026-midterms).