Multimodal AI Benchmarks 2026: Vision, Audio, Code

The version of this page published on April 24, 2026 named a leader for every modality and gave each one a precise score. Those scores could not be traced to a lab, a model report or a benchmark’s own leaderboard, so this rebuild replaces them with figures that can. Where a primary source does not exist, the page now says so instead of filling the gap.

What remains is smaller and more useful. A handful of labs publish their own multimodal results in launch posts and technical reports. Each result carries a benchmark, a model variant and a setting, and those details decide whether two numbers can sit in the same comparison. This page keeps them together.

It covers four areas: image, document and screen understanding (MMMU-Pro, ChartQA, OSWorld-Verified), video (Video-MMMU, LVBench, Video-MME), audio and speech (word error rate on public test sets), and coding (SWE-Bench Pro, Terminal-Bench 2.0, LiveCodeBench Pro). For a wider view of how the leading models compare on price and agentic work, see our GPT-5.5 and Claude Opus 4.7 comparison.

Almost every number in this field is reported by the lab that built the model. That is normal, and it is also why the label beside each figure matters. A lab chooses the benchmark, the effort setting and the competitor versions it lists, and it often runs rival models itself or copies their reported results.

The MMMU-Pro rows show the problem clearly. Google’s evaluation notes say its MMMU-Pro scores average the Standard (10 options) and Vision settings. OpenAI’s table reports MMMU-Pro with and without tools. The two labs publish numbers under one benchmark name, but they are not the same measurement.

MMMU-Pro is the common reference point for image reasoning. The sourced rows put GPT-5.5, Gemini 3 Pro and Gemini 3.1 Pro between 80.5% and 81.2%, with Claude Sonnet 4.6, Claude Opus 4.6 and GPT-5.2 in the 73.9% to 79.5% range in Google’s February 2026 table. One detail stands out: in that table Gemini 3 Pro scores slightly higher (81.0%) than its successor Gemini 3.1 Pro (80.5%), so a newer model is not automatically ahead on every benchmark.

Claude Opus 4.7 does not have an MMMU-Pro figure in the sources read for this page; OpenAI’s table leaves its cell blank, and Anthropic’s post presents its vision evaluations as charts. What Anthropic does state in text is a platform change: Opus 4.7 accepts images up to 2,576 pixels on the long edge, about 3.75 megapixels, which it describes as more than three times as many as prior Claude models. For a deeper look at the model, see our Claude Opus 4.7 guide.

Chart and diagram reading has less current coverage. The only published ChartQA and AI2D scores in the sources are from the Qwen3-Omni report, where Qwen3-Omni-Instruct scores 86.8 on ChartQA against 86.7 for GPT-4o and 89.5 for Qwen2.5-VL-72B. Those rivals are 2024 and 2025 models, so the table says what the report compared, not where the 2026 frontier sits. Computer-use screens are covered by OSWorld-Verified, where OpenAI lists 78.7% for GPT-5.5 and 78.0% for Claude Opus 4.7, a gap of 0.7 points that comes from one vendor’s table.

Three video benchmarks appear in the primary sources, and they test different things. Video-MMMU asks college-level questions about video content. LVBench targets long-video understanding. Video-MME covers short, medium and long clips and reports results with and without subtitles. A high score on one says little about the others.

Google reports 87.6% on Video-MMMU for Gemini 3 Pro, run with high media resolution and temperature 0. For Gemini 4 Argon, Google reports 91.7% on LVBench and calls it state of the art; access to that model is limited, so most teams cannot yet reproduce the figure. The Qwen3-Omni report adds Video-MME and LVBench rows for an open-weight model against older Gemini and GPT-4o versions.

The official Video-MME leaderboard is the most useful reference for how the benchmark is scored, but when read on November 5, 2026 its newest entry is dated September 28, 2025 (video-SALMONN 2+ at 79.7% without subtitles), and the Gemini 1.5 Pro (75.0%) and GPT-4o (71.9%) entries are dated June 2024. It does not carry 2026 frontier results, so any page that quotes a 2026 Video-MME leader needs a lab report behind it.

Audio comparisons rest on word error rate (WER), where lower is better, and on audio-reasoning suites such as VoiceBench and MMAU. The one report with a full audio table among the primary sources is the Qwen3-Omni technical report. It says the model reaches open-source state of the art on 32 of 36 audio and audiovisual benchmarks and overall state of the art on 22.

The speech rows are consistent with that claim. Qwen3-Omni-Instruct posts 1.22% WER on Librispeech test-clean and 2.48% on test-other, ahead of Gemini-2.5-Pro and GPT-4o-Transcribe in the same table. On the 19-language Fleurs average, GPT-4o-Transcribe scores better (4.48% against 5.33%). On VoiceBench overall, Gemini-2.5-Pro (89.6) and GPT-4o-Audio (86.8) lead Qwen3-Omni-Instruct (85.5). These are the Qwen team’s own runs against models that predate the 2026 releases.

The latency claim needs care. The report gives a theoretical end-to-end first-packet latency of 234 ms for audio and 547 ms for audiovisual input under its lowest-concurrency test condition, and larger values at higher concurrency. That is an estimate for one configuration, not a measured time-to-first-token in production, and it is not an accuracy figure. Teams building voice agents should measure latency on their own stack.

No lab publishes a code-with-vision score under the names the earlier version of this page used. What exists are agentic coding benchmarks, which test code and tool use rather than screenshots. They still matter to a multimodal buyer because coding agents read screenshots, logs and terminals. The sourced figures come from two vendor tables, and each stays inside its own harness.

In OpenAI’s table GPT-5.5 leads Terminal-Bench 2.0 by more than 13 points over Claude Opus 4.7, but Opus 4.7 leads SWE-Bench Pro, 64.3% against 58.6%. Google’s February table adds a third view: its Gemini 3.1 Pro entry is 54.2% on SWE-Bench Pro (single attempt) and 68.5% on Terminal-Bench 2.0 (Terminus-2 harness). For a head-to-head on agentic coding, see our Gemini 3.5 Flash, GPT-5.5 and Opus 4.7 comparison.

The primary sources do not support naming a single winner per modality. They do support a method: choose the benchmark closest to your workload, read the setting, and rerun a sample of your own data on two or three candidates. Our AI transformation service builds those evaluation sets with engineering teams.

The April version of this page carried figures that circulate in other summaries. None could be matched to a primary source. The table below sets each against what the nearest primary document actually reports.

Model developers: launch posts and evaluation notes

• OpenAI, Introducing GPT-5.5, April 23, 2026 (API update April 24).
• Google, A new era of intelligence with Gemini 3, November 18, 2025.
• Google DeepMind, Gemini 3 Pro: Model Evaluation, Approach, Methodology and Results, November 18, 2025.
• Google DeepMind, Gemini 3.1 Pro benchmark table, read November 5, 2026, with the evaluation methodology, February 19, 2026.
• Google, Gemini 4 Argon announcement, September 30, 2026.
• Anthropic, Introducing Claude Opus 4.7, April 16, 2026, and Introducing Claude Opus 5, July 24, 2026.
• OpenAI, Introducing GPT-6.1 Sol, read November 5, 2026 (cited only for its existence).

Academic and benchmark owners

• Qwen team (Alibaba), Qwen3-Omni Technical Report, arXiv 2509.17765, September 22, 2025.
• Video-MME authors, Video-MME leaderboard, read November 5, 2026.