Benchmark Almanac

MultiMedia-TerminalBench

Terminus-MM + Gemini 3.1 Pro 37.1% tops its leaderboard GPT-5.2 16.2% leads as a model on its own Gemini 3.1 Pro 12.4% leads in Terminus 2
Not in model cards

Agents inspect and edit media in persistent terminal workspaces. The 105 tasks require verifiable output files across five areas of professional multimedia work.

In May 2026, the best Gemini 3.1 Pro setup reached 37.1% binary success and 46.9% partial credit. GPT-5.2 with Codex CLI solved 16.2%.

Made by
Chiyeong Heo and colleagues
Released
May 2026
Kind
Video understanding
Level
Expert
Scored
Checked automatically
Answers
Agent task
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Images, Video, Audio → Actions, Images, Video, Audio
Interaction
Agentic
Environment
Terminal
Refresh
One fixed set
Run by
Anyone
Occupations
Arts, design, entertainment, sports and media
Strongest model tested
Gemini 3.1 Pro #40 on the capability index
Last model added
February 2026
Scores refreshed
November 20th, 2026
21
Entries in its paper, 5 models we track

Terminus-MM + Gemini 3.1 Pro tops the leaderboard with 37.1% 2 models share the top score of 16.2% among models released by February 2026 Gemini 3.1 Pro leads with 12.4% in Terminus 2

Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran on its own with several settings, we keep its best. Every run is in the next chart. Agents other companies built around a model get a pale shade of that model's lab color. Gray marks blends and entries not in our model list, such as fine-tunes and older models. None of these count toward leaders. The best model on its own is GPT-5.2, at 16.2%. Agents built around a model, such as Terminus-MM + Gemini 3.1 Pro, are their own entries. We don't count them as the model's score. This view shows only models we track, each at its best. This view keeps only runs in Terminus 2, a harness no model’s maker built. A maker’s own agent, such as Claude Code with a Claude model, is left out.

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

Source: The MultiMedia-TerminalBench paper

#EntryScoreSettingDateSource
1 Terminus-MM + Gemini 3.1 Pro 37.1% T+I+A+V Feb 2026 Official leaderboard
2 Terminus-IA + Gemini 3.1 Pro 33.3% T+I+A Feb 2026 Official leaderboard
3 Terminus-IV + Gemini 3.1 Pro 33.3% T+I+V Feb 2026 Official leaderboard
4 Terminus-V + Gemini 3.1 Pro 28.6% T+V Feb 2026 Official leaderboard
5 Terminus-AV + Gemini 3.1 Pro 27.6% T+A+V Feb 2026 Official leaderboard
6 Terminus-A + Gemini 3.1 Pro 25.7% T+A Feb 2026 Official leaderboard
7 Terminus-V + Gemini 2.5 Flash Sep 2025 22.9% T+V Sep 2025 Official leaderboard
8 Terminus-MM + Gemini 2.5 Flash Sep 2025 22.9% T+I+A+V Sep 2025 Official leaderboard
9 Terminus-AV + Gemini 2.5 Flash Sep 2025 21.9% T+A+V Sep 2025 Official leaderboard
10 Terminus-IV + Gemini 2.5 Flash Sep 2025 16.2% T+I+V Sep 2025 Official leaderboard
11 GPT-5.2 16.2% T+I, Codex CLI Dec 2025 Official leaderboard
12 Claude Sonnet 4.6 16.2% T+I, Claude Code Feb 2026 Official leaderboard
13 Terminus-IA + Gemini 2.5 Flash Sep 2025 13.3% T+I+A Sep 2025 Official leaderboard
14 Gemini 3.1 Pro 12.4% T, Terminus-2 Feb 2026 Official leaderboard
15 Terminus-KIRA + GPT-5.2 11.4% T+I Dec 2025 Official leaderboard
16 Terminus-A + Gemini 2.5 Flash Sep 2025 10.5% T+A Sep 2025 Official leaderboard
17 Terminus-KIRA + Gemini 3.1 Pro 10.5% T+I Feb 2026 Official leaderboard
18 Qwen3.5-122B-A10B 10.5% T, Terminus-2 Feb 2026 Official leaderboard
19 Terminus-KIRA + Qwen3.5-122B-A10B 9.50% T+I Feb 2026 Official leaderboard
20 Gemini 2.5 Flash Sep 2025 6.70% T, Terminus-2 Sep 2025 Official leaderboard
21 Terminus-KIRA + Gemini 2.5 Flash Sep 2025 6.70% T+I Sep 2025 Official leaderboard
#ModelScoreSettingDateSource
1 GPT-5.2 16.2% T+I, Codex CLI Dec 2025 Official leaderboard
2 Claude Sonnet 4.6 16.2% T+I, Claude Code Feb 2026 Official leaderboard
3 Gemini 3.1 Pro 12.4% T, Terminus-2 Feb 2026 Official leaderboard
4 Qwen3.5-122B-A10B 10.5% T, Terminus-2 Feb 2026 Official leaderboard
5 Gemini 2.5 Flash Sep 2025 6.70% T, Terminus-2 Sep 2025 Official leaderboard
#ModelScoreSettingDateSource
1 Gemini 3.1 Pro 12.4% T, Terminus-2 Feb 2026 Official leaderboard
2 GPT-5.2 10.5% T, Terminus-2 Dec 2025 Official leaderboard
3 Qwen3.5-122B-A10B 10.5% T, Terminus-2 Feb 2026 Official leaderboard
4 Gemini 2.5 Flash Sep 2025 6.70% T, Terminus-2 Sep 2025 Official leaderboard

Rankings by part

We read 17 parts from the benchmark’s paper in 2 groups: capability and meta-category. Each ranks models on one part of the test. GPT-5.2 leads 8 of them, and 3 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.

Audio-Visual Alignment

5 models we track have a score here.

  1. 1 Claude Sonnet 4.6 ↑1 12.7%
  2. 2 GPT-5.2 ↓1 10.9%
  3. 3 Gemini 2.5 Flash Sep 2025 ↑2 9.10%
  4. 4 Gemini 3.1 Pro ↓1 9.10%
  5. 5 Qwen3.5-122B-A10B ↓1 3.60%

The best score, 37.1%, costs $1.23 a task

Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

Agents other companies built, in a pale shade of the model’s lab Cost frontier One model at each effort level
0% 10% 20% 30% 40% $0.01 $0.10 $1.00 $10 Cost per task, log scale Terminus-MM + Gemini 3.1 Pro, T+I+A+V Terminus-V + Gemini 3.1 Pro, T+V Terminus-MM + Gemini 2.5 Flash, T+I+A+V 0% 10% 20% 30% 40% $0.01 $0.10 $1.00 $10 Cost per task, log scale Terminus-MM + Gemini 3.1 Pro, T+I+A+V Terminus-V + Gemini 3.1 Pro, T+V

Source: The MultiMedia-TerminalBench paper

ModelEffortHarnessScoreCost per taskOutput tokens
Terminus-MM + Gemini 3.1 Pro T+I+A+V Terminus-MM 37.1% $1.23
Terminus-IA + Gemini 3.1 Pro T+I+A Terminus-IA 33.3% $1.74
Terminus-IV + Gemini 3.1 Pro T+I+V Terminus-IV 33.3% $1.28
Terminus-MM + Gemini 3.1 Pro Without modality masking Terminus-MM 32.4%
Terminus-V + Gemini 3.1 Pro T+V Terminus-V 28.6% $1.12
Terminus-AV + Gemini 3.1 Pro T+A+V Terminus-AV 27.6% $1.03
Terminus-A + Gemini 3.1 Pro T+A Terminus-A 25.7% $1.99
Terminus-V + Gemini 2.5 Flash T+V Terminus-V 22.9% $0.14
Terminus-MM + Gemini 2.5 Flash T+I+A+V Terminus-MM 22.9% $0.10
Terminus-AV + Gemini 2.5 Flash T+A+V Terminus-AV 21.9% $0.14
Terminus-MM + Gemini 2.5 Flash Without modality masking Terminus-MM 17.1%
GPT-5.2 T+I Codex CLI 16.2% $7.12
Claude Sonnet 4.6 T+I Claude Code 16.2% $1.74
Terminus-IV + Gemini 2.5 Flash T+I+V Terminus-IV 16.2% $0.18
Terminus-IA + Gemini 2.5 Flash T+I+A Terminus-IA 13.3% $0.23
Gemini 3.1 Pro T Terminus-2 12.4% $0.77
Terminus-KIRA + GPT-5.2 T+I Terminus-KIRA 11.4% $1.67
GPT-5.2 T Terminus-2 10.5% $0.82
Qwen3.5-122B-A10B T Terminus-2 10.5% $0.10
Terminus-A + Gemini 2.5 Flash T+A Terminus-A 10.5% $0.13
Terminus-KIRA + Gemini 3.1 Pro T+I Terminus-KIRA 10.5% $2.06
Terminus-KIRA + Qwen3.5-122B-A10B T+I Terminus-KIRA 9.50% $0.23
Gemini 2.5 Flash Sep 2025 T Terminus-2 6.70% $0.12
Terminus-KIRA + Gemini 2.5 Flash T+I Terminus-KIRA 6.70% $0.23

Number one over time

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

OpenAI
Benchmark paper Not tested since May 2026
GPT-5.2 16.2%

Source: The MultiMedia-TerminalBench paper

Scores over time

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

Plot with other benchmarks
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) GPT-5.2 Gemini 2.5 Flash 0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) GPT-5.2 Gemini 2.5 Flash

Source: Official MultiMedia-TerminalBench leaderboard

In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Gemini 3.1 Pro GPT-5.2 Gemini 2.5 Flash 0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Gemini 3.1 Pro

Source: Official MultiMedia-TerminalBench leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Audio and video

All 695 in Audio and video
  • Video-MMMU In use 300 expert-level lecture videos with questions testing perception, comprehension and adaptation of knowledge. Accuracy. Qwen 3.8 Max 88.7%
  • LVBench In use Hours-long video understanding with multiple-choice questions on reasoning, grounding and summarization. Gemini 4 Argon 91.7%
  • MMVU In use 3,000 expert-annotated questions on videos across 27 disciplines. Multiple-choice and open-ended accuracy. Qwen 3.8 Max 82.4%
  • MLVU In use Multi-task long video understanding (minutes to hours) with multiple-choice and generation tasks. Accuracy. Qwen 3.8 Max 90.8%
  • FLEURS In use Models transcribe read speech in 102 languages built from FLoRes-101 sentences. Model cards usually report word error rate, so lower scores are better. Phi-4-multimodal 4.00%
  • VoiceBench In use Spoken instructions with speaker, environment and content variation. Voice assistant responses scored by GPT judge or rules. Qwen3.5-Omni-Plus 93.1%

Data audit

9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 9 months ago, in February 2026, for Qwen3.5-122B-A10B. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
  • Pass Independent runs Scores from the benchmark’s paper (21 entries).
  • Pass Models scored 5 models we track, plus 16 other entries on its leaderboard.
  • Warning Newest score The newest came 9 months ago, in February 2026, for Qwen3.5-122B-A10B.
  • Note Frontier models None of the top 10 has a score. The strongest tested is Gemini 3.1 Pro (#40).
  • Pass Headroom The best is 16.2%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 24 rows for models we track at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check