GPT-5.4 Image 2

OpenRouter launched an "Image benchmarks" site on August 25, 2026 that compares multiple models side-by-side using identical prompts, with results sortable by cost and generation time. The default "pour white wine to the brim" prompt tests prompt fidelity — only "GPT Image 2" and "GPT-5.4 Image 2" satisfied the instruc The benchmark also covers text rendering accuracy, image editing precision, and other challenge families such as "Closed Umbrellas" (counting and state negation). Separately, OpenAI rolled out a ChatGPT Images 2.0–powered custom sticker feature for mobile apps on the same day, supporting up to nine stickers from photos

GIGAZINE's August 25, 2026 report covers OpenRouter's "Image benchmarks" platform, which displays identical-prompt outputs from a range of image generation AIs. In the default wine-glass prompt, only "OpenAI: GPT Image 2" and "OpenAI: GPT-5.4 Image 2" filled the glass to the brim, with other models failing to match the The site lets visitors compare text rendering performance, switch prompts (e.g., "Closed Umbrellas" — 12 people with closed umbrellas in heavy rain), and review image editing results such as object removal. Results are sortable by descending cost or generation time, and clicking an output surfaces a button to obtain th

OpenOCR published a benchmark ranking page explicitly titled "OpenAI: GPT-5.4 Image 2," reporting an average score of 99.3% across 11 documents on the ocr-shared-11-normalized/v2 corpus. The evaluation was measured on 2026-08-07 and published 2026-08-22, with all 11 scores reproducible from on-page transcripts using ch Per-document results include a grocery receipt at 100%, a café receipt at 98.2%, a handwritten letter at 100%, and handwritten notes at 100%, with average latency of 2.5 seconds. The provider offering used for the measurement is explicitly identified as openrouter-openai-gpt-5.4-image-2, with corpus SHA256 b8f240c12ac5