Interfaze: New Model Architecture Beats Gemini-3-Flash and GPT-5.4-Mini on Deterministic Tasks

Interfaze is a new model architecture from Interfaze that merges task-specific DNN/CNN models with omni-transformers, targeting high-accuracy deterministic tasks at scale. It offers a 1M token context window, 32k max output tokens, and supports text, images, audio, and file inputs with optional reasoning.

Benchmark Results

According to their benchmarks, Interfaze leads against similar pricing-tier models (Flash/mini models like Gemini-3-Flash, GPT-5.4-Mini, Claude Sonnet 4.6, and Grok-4.3) across 9 head-to-head tests:

• OCRBench V2: Interfaze 70.7% vs Gemini-3-Flash 55.8%, Claude-Sonnet-4.6 54.7%, GPT-5.4-Mini 52.7%, Grok-4.3 54.7%
• olmOCR: Interfaze 85.7% vs Gemini-3-Flash 75.3%, Claude-Sonnet-4.6 73.9%, GPT-5.4-Mini 80.1%, Grok-4.3 81.9%
• RefCOCO: Interfaze 82.1% vs Gemini-3-Flash 75.2%, Claude-Sonnet-4.6 75.5%, GPT-5.4-Mini 67.0%, Grok-4.3 25.0%
• VoxPopuli (WER, lower is better): Interfaze 2.4% vs Gemini-3-Flash 4.0%
• Spider 2.0-Lite: Interfaze 52.9% vs Gemini-3-Flash 45.2%, Claude-Sonnet-4.6 49.6%, GPT-5.4-Mini 26.7%, Grok-4.3 45.9%
• GPQA Diamond: Interfaze 89.9% vs Gemini-3-Flash 88.5%, Claude-Sonnet-4.6 89.9%, GPT-5.4-Mini 82.8%, Grok-4.3 73.6%
• MMMLU: Interfaze 90.9% vs Gemini-3-Flash 88.7%, Claude-Sonnet-4.6 84.9%, GPT-5.4-Mini 75.3%, Grok-4.3 89.7%
• MMMU-Pro: Interfaze 71.1% vs Gemini-3-Flash 67.6%, Claude-Sonnet-4.6 46.3%, GPT-5.4-Mini 40.4%, Grok-4.3 68.7%
• SOB Value Acc: Interfaze 79.5% vs Gemini-3-Flash 77.3%, Claude-Sonnet-4.6 77.9%, GPT-5.4-Mini 75.1%, Grok-4.3 78.4%

Interfaze also outperforms specialized OCR providers like Chandra OCR and Reducto, according to the source.

Pricing

Interfaze is priced at $1.50 per million input tokens and $3.50 per million output tokens — in line with Gemini-3-Flash.

Who It's For

Developers building high-volume OCR, document extraction, web search, audio transcription/speaker diarization, translation, or object/GUI detection pipelines who need deterministic accuracy without the cost of full LLMs.