●ChaosWatch
New here? Start here →
Model Arena · depth shallow

Sub-Saharan Africa · 2026-07-04

Season 1 · 2026-06-25 → 2026-07-12 · 90 region-days (18 dates × 5 regions) · scored 2026-08-05 → 2026-08-06 · closed pilot, not a live feed

One day, one region, several language models — each handed exactly the same headlines and readings, then asked how tense things would be a week later. Because the input is identical, any difference between their answers is a difference between the models.

This page shows what they said. Whether any of it was right is a separate question, answered on the Leaderboard — and the short version there is that none of them detectably beat guessing. Scores run from 0 (calm) to 100 (catastrophe); the Leaderboard puts those scores and what actually happened on one standard scale before comparing them, which is why its error numbers sit around 1 rather than on the 0–100 scale.

How 'identical evidence' is guaranteed

Every model in a run receives the same rendered prompt over the same evidence payload, and the payload's hash is stored with the run and displayed below. If two models had been shown different material the hashes would differ, so the fairness of the comparison is checkable rather than asserted.

Models that crashed, timed out, or returned unparseable output are persisted as rows and ranked here alongside the successes.

Asia-PacificEuropeMiddle East & North AfricaSub-Saharan AfricaNorth America
2026-07-122026-07-112026-07-102026-07-092026-07-082026-07-072026-07-062026-07-052026-07-042026-07-032026-07-022026-07-012026-06-302026-06-292026-06-282026-06-272026-06-262026-06-25
Consensus
20.0
the middle of 5 answers, with the highest and lowest set aside
Completion
100.0%
5 scored / 5 attempted — failures stay in the denominator
Evidence
4eae5e84…9718
10 headline(s), shared by every model

Ranking

prompt version p1depth shallow

Models that failed are listed here too, marked and unscored, rather than quietly dropped. A model that only answers when the question is easy would otherwise look like a model that is good at the question — every leaderboard that hides its failures is really reporting "accuracy among the runs that happened to work".

1
claude-opus-5okStable
anthropic
latency 7,996 mstokens in 1,360tokens out 276
22.0
2
deepseek-reasonerokStable
deepseek
latency 3,380 mstokens in 845tokens out 174
20.0
3
gemini-3.1-pro-previewokStable
google
latency 11,866 mstokens in 568tokens out 68
20.0
4
llama-3.3-70b-versatileokStable
groq
latency 957 mstokens in 794tokens out 83
20.0
5
claude-haiku-4-5okStable
anthropic
latency 2,966 mstokens in 935tokens out 115
18.0

Every model in this run returned a verdict, so no failure row is shown.