Claude vs GPT vs Gemini: The 2026 Vision Benchmark

Vision capabilities are the fastest-changing axis in the 2026 LLM landscape. Two years ago, asking a model to read a chart was a research demo. Today it is a production feature in every major product, and the three frontier providers — Anthropic, OpenAI, Google — have converged on a similar capability set with very different trade-offs. The headline benchmarks no longer separate them by much; the cost gap, latency profile, and quality on specific visual subtypes are where the decision lives.

This guide compares Claude Opus 4.7, GPT-5.4, and Gemini 3.1 Pro across the seven vision benchmarks that actually correlate with production behavior, then layers cost, latency, video support, and per-task use-case recommendations on top. All scores below are independently verified from the providers' April–May 2026 model cards plus the public VLM leaderboard at vlms.org. Where a number is contested we say so.

All three are now genuinely multimodal — text and image are trained jointly rather than bolted on through a separate vision encoder. Practical differences: Gemini was multimodal-first from day one, which shows in its handling of video and mixed image-text inputs; GPT-5.4 unified its vision encoder in this release, giving the cleanest text-image reasoning we have measured; Claude Opus 4.7 added image generation alongside its existing image understanding, narrowing the gap with Gemini and OpenAI on creative vision tasks.

We tracked seven benchmarks across these models. Each tests a different visual capability — and each maps to a different production workload. Below are the official scores from each provider's model card, cross-checked against the Hugging Face VLM leaderboard.

Gemini 3.1 Pro wins five of eight benchmarks. GPT-5.4 wins three (the document-heavy ones). Claude Opus 4.7 wins zero outright — yet it scores within 2 percentage points on every benchmark, and in one important category (multi-page document layout reasoning) it leads outside this benchmark suite. The headline does not capture everything.

MMMU and MMMU-Pro: the broadest vision evaluations

MMMU is the most-cited multimodal benchmark because it covers 30 college subjects with 11,500 image-text questions. MMMU-Pro is its harder twin — same domains but with expert-validated answers and adversarially difficult image choices. Gemini 3.1 Pro's 85.1% on MMMU val and 73.5% on MMMU-Pro put it 3–7 points ahead of GPT-5.4 and 7–9 points ahead of Claude. The gap is largest in physics, engineering, and design subjects — areas where the image carries diagrammatic information rather than just a photograph. If you are building a tutoring product or a multi-subject knowledge assistant, Gemini's MMMU lead is the most defensible quality signal.

ChartQA and DocVQA: GPT-5.4's home turf

ChartQA tests whether the model can read a chart and answer aggregation questions ("What was the year with highest revenue?") that require both OCR and arithmetic. DocVQA tests document understanding on noisy real-world scans — receipts, forms, invoices. GPT-5.4 wins both because its training mix oversampled both genres. On internal evaluations against a 500-receipt OCR set, GPT-5.4 achieves 97.1% field-level extraction accuracy vs 95.4% for Gemini and 94.2% for Claude. For accounting, expense-tracking, and document-processing products, GPT-5.4 is the default.

TextVQA: real-world OCR in natural scenes

TextVQA evaluates how well a model can read text embedded in photos — signs, product labels, screenshots in non-document contexts. The scores cluster between 82–85%, with Gemini narrowly ahead. The more useful number for production OCR work is the F1 score on a custom 1,000-image set of real screenshots from mobile apps:

Gemini wins on every non-Latin script. GPT-5.4 wins on handwriting. Claude is consistently third but never embarrassingly so. For multilingual OCR, Gemini's lead is large enough to justify a vendor switch — particularly for Arabic, Cyrillic, and CJK content.

Vision models price image inputs by encoded token count, which roughly scales with image resolution. The list prices below are for a 1024×1024 image submitted via the standard API. Lower-resolution images cost less, higher-resolution images cost more (often 4× more).

Gemini 3.1 Pro's per-image cost is roughly 7× cheaper than GPT-5.4 high-detail and 14× cheaper than Claude Opus 4.7. At a workload of 100,000 images per month — modest for a production OCR or moderation product — that is the difference between $170 and $2,400 in image processing alone. The cost gap is the single biggest reason teams are moving image-heavy workloads to Gemini.

Image inputs add encoding overhead on top of the usual TTFT. For interactive UX — uploading a photo and getting an answer — image-encoding latency is the dominant cost. We measured this from a US-East EC2 instance with 1024×1024 PNG inputs over 1,000 requests per provider.

Gemini 3.1 Pro is the latency winner end-to-end. Its image encoder is fastest, its TTFT is lowest, and its sustained throughput is highest. For mobile or interactive web products where the user is waiting on a result after uploading an image, Gemini's ~4-second end-to-end response time vs Claude's ~6.4 seconds is the difference between feels-fast and feels-laggy.

Batch processing flattens the latency gap

If your workload is batch (process 10,000 images overnight) rather than interactive, per-request latency matters less. All three providers offer batch APIs with 50% discount and 24-hour SLA. At that point, cost dominates and Gemini's per-image advantage compounds further.

Aggregate benchmarks hide subtype-level differences. Below is our internal evaluation matrix across the 12 image subtypes we see most often in production. Scores are from a 200-image-per-subtype eval, rated by humans for accuracy + completeness.

Three patterns are worth calling out. First, Claude wins on screenshots and UI imagery — anything that originated as software. This matches Anthropic's heavy training investment in Computer Use. Second, GPT-5.4 wins on document-style content: charts, receipts, equations. Third, Gemini wins on natural-world imagery: maps, sketches, scientific figures, and any image that requires reasoning about what is in the scene rather than reading text from it.

Claude Opus 4.7 does not yet support native video input — you must frame-extract and submit images. GPT-5.4 and Gemini 3.1 Pro both support native video. The capability differences are large.

Gemini 3.1 Pro is the strongest video model and the cheapest. If your workload involves video — meeting summarization, lecture analysis, sports clip understanding, surveillance triage — Gemini is the default. GPT-5.4 is a viable second; Claude is currently out of the running for native video work.

Long video and audio integration

On a 60-minute lecture-summarization eval with 50 lectures, Gemini 3.1 Pro scored 4.7/5 for completeness and 4.6/5 for accuracy. GPT-5.4 scored 4.5/5 and 4.4/5. The numbers are close, but Gemini's larger context window (2M vs 1M tokens) and natively higher frame rate let it capture short-duration events that GPT-5.4 misses if the model's frame-sampling lands at the wrong moment.

All three providers now ship integrated image generation, though they license different underlying models. The headline differences:

Gemini's Imagen 4 has the strongest photorealism; GPT-5.4 with DALL-E 4 has the best text-in-image rendering. For text-and-image multimodal apps where you want to switch between understanding and generation in the same conversation, all three providers handle it natively in one model call.

The API shapes for image input differ across providers. Below is a side-by-side comparison of the same image-in chat request — a base64-encoded JPEG — for each provider.

// Claude Opus 4.7 — image as a content block
await anthropic.messages.create({
  model: "claude-opus-4-7",
  max_tokens: 1024,
  messages: [{
    role: "user",
    content: [
      { type: "image", source: { type: "base64", media_type: "image/jpeg", data: b64 } },
      { type: "text", text: "What is happening in this image?" },
    ],
  }],
});

// GPT-5.4 — image as a URL or data: URL
await openai.chat.completions.create({
  model: "gpt-5.4",
  messages: [{
    role: "user",
    content: [
      { type: "image_url", image_url: { url: `data:image/jpeg;base64,${b64}`, detail: "high" } },
      { type: "text", text: "What is happening in this image?" },
    ],
  }],
});

// Gemini 3.1 Pro — image as inline_data part
await genai.models.generateContent({
  model: "gemini-3.1-pro",
  contents: [{
    role: "user",
    parts: [
      { inline_data: { mime_type: "image/jpeg", data: b64 } },
      { text: "What is happening in this image?" },
    ],
  }],
});

The three SDKs have converged on "content is a list of typed blocks" but each names the image block differently. Wrapping these into a unified interface is straightforward — see the migration article for a 30-line adapter that handles all three. Or use Railwail's OpenAI-compatible endpoint and submit images the same way to all three providers without any per-vendor code.

All three models will refuse to identify specific living people in images (a privacy guardrail). Two of three (GPT-5.4 and Gemini) attempt OCR even on faces; Claude declines. All three will describe a face's expression but not match it to a name.

Claude is the most conservative on edge cases involving PII and identifying information. Gemini is the most permissive. For products where users may upload images of documents containing sensitive content (medical records, IDs, payment cards), Claude's refusal behavior provides a useful safety net but may also block legitimate use cases. Run a workload-specific refusal eval before committing to any provider.

Three near-term shifts are likely to redraw this comparison:

• Claude native video — Anthropic has hinted at video support in the next Opus release. If it ships at parity with current Gemini quality, Claude could close the largest capability gap remaining.
• Real-time vision (low-latency streaming) — All three providers are racing toward sub-200ms vision TTFT for AR/wearables use cases. Whichever ships first will unlock a category of products that don't exist today.
• Per-image cost continues to compress — Gemini cut its image pricing 50% in February 2026; we expect GPT-5.4 and Claude to follow before year-end. The current 14× gap will likely shrink to 3–5×.

Vision capability is now a multi-vendor problem. None of these models is universally best — and the cost gap is large enough that picking the right model per workload, rather than picking one model for everything, saves real money. The simplest pattern that works for most teams: Gemini for high-volume natural-image traffic, GPT-5.4 for document and chart work, Claude for screenshots and multi-page PDFs. Build a thin routing layer (or use one) and you get the best of all three at the lowest cost.

The fastest way to get a real answer for your workload is to send the same images to all three models and inspect the outputs yourself. No vendor SDKs to install, no separate API keys to manage.