Benchmark Almanac

GeoGuesser environment (FineEnvs)

Claude Sonnet 5 0.64 leads since September 2026
Not in model cards

Agent placed at unknown Mapillary street-level location looks, walks, pins candidates and guesses coordinates. Scored by a distance-based GeoGuessr-style reward from 0 to 1.

Frozen 200-task eval split committed so resets are byte-identical. The article ('How to turn a game into an RL environment') trains a 4B VLM with TRL and keeps the dead ends in.

Released
September 2026
Kind
General visual questions
Level
Everyday
Scored
Checked automatically
Answers
Agent task
Tasks
All public
Strongest model tested
Claude Sonnet 5 #33 on the capability index
Last model added
June 2026
Scores refreshed
November 12th, 2026
14
Models scored on its own leaderboard

Claude Sonnet 5 leads with 0.64 among models released by June 2026

Best score per model from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best.

The newest model with a score here came out in June 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

14 models. Scroll sideways for the rest.

  1. 0.64 Claude Sonnet 5
  2. 0.45 GPT-5.4 Mini
  3. 0.44 Qwen3.5-122B-A10B
  4. 0.43 Claude Haiku 4.5
  5. 0.34 Qwen3.5 397B-A17B
  6. 0.33 Qwen3.5-27B
  7. 0.33 Qwen3.5-35B-A3B
  8. 0.33 Qwen3.5-9B
  9. 0.33 Qwen3.5-4B
  10. 0.29 Qwen3-VL-4B-Instruct
  11. 0.28 Qwen3.5-2B
  12. 0.26 GPT-5.4 Nano
  13. 0.25 Qwen3-VL-2B-Instruct
  14. 0.16 Qwen3.5-0.8B
  1. 1 Claude Sonnet 5 0.64
  2. 2 GPT-5.4 Mini 0.45
  3. 3 Qwen3.5-122B-A10B provider: novita 0.44
  4. 4 Claude Haiku 4.5 0.43
  5. 5 Qwen3.5 397B-A17B provider: novita 0.34
  6. 6 Qwen3.5-27B provider: deepinfra 0.33
  7. 7 Qwen3.5-35B-A3B provider: deepinfra 0.33
  8. 8 Qwen3.5-9B provider: local 0.33
  9. 9 Qwen3.5-4B provider: local 0.33
  10. 10 Qwen3-VL-4B-Instruct provider: local 0.29
  11. 11 Qwen3.5-2B provider: local 0.28
  12. 12 GPT-5.4 Nano 0.26
  13. 13 Qwen3-VL-2B-Instruct provider: local 0.25
  14. 14 Qwen3.5-0.8B provider: local 0.16

Source: Official GeoGuesser environment (FineEnvs) leaderboard

#ModelScoreSettingDateSource
1 Claude Sonnet 5 0.64 Jun 2026 Official leaderboard
2 GPT-5.4 Mini 0.45 Mar 2026 Official leaderboard
3 Qwen3.5-122B-A10B 0.44 provider: novita Feb 2026 Official leaderboard
4 Claude Haiku 4.5 0.43 Oct 2025 Official leaderboard
5 Qwen3.5 397B-A17B 0.34 provider: novita Feb 2026 Official leaderboard
6 Qwen3.5-27B 0.33 provider: deepinfra Feb 2026 Official leaderboard
7 Qwen3.5-35B-A3B 0.33 provider: deepinfra Feb 2026 Official leaderboard
8 Qwen3.5-9B 0.33 provider: local Feb 2026 Official leaderboard
9 Qwen3.5-4B 0.33 provider: local Feb 2026 Official leaderboard
10 Qwen3-VL-4B-Instruct 0.29 provider: local Oct 2025 Official leaderboard
11 Qwen3.5-2B 0.28 provider: local Feb 2026 Official leaderboard
12 GPT-5.4 Nano 0.26 Mar 2026 Official leaderboard
13 Qwen3-VL-2B-Instruct 0.25 provider: local Oct 2025 Official leaderboard
14 Qwen3.5-0.8B 0.16 provider: local Feb 2026 Official leaderboard

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

Anthropic
Own leaderboard Not tested since Sep 2026
Claude Sonnet 5 0.64

Source: Official GeoGuesser environment (FineEnvs) leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard
0.20 0.40 0.60 2026 0.20 0.40 0.60 2026

Source: Official GeoGuesser environment (FineEnvs) leaderboard

No lab model card reports it yet

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.

More in Vision

All 281 in Vision
  • MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy. GPT-5.1 85.4%
  • MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
  • CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
  • MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy. Qwen 3.8 Max 97.7%
  • MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy. Qwen 3.8 Max 85.9%
  • ChartQA In use Models answer human-written and generated questions about chart images that need visual and arithmetic reasoning. Scored by relaxed accuracy with a 5 percent numeric tolerance. Command A Vision 90.9%