Interfaze

Interfaze is a multimodal AI model engineered for deterministic accuracy in OCR, speech-to-text, and structured data extraction. Built by JigsawStack (YC P26), it targets developers and enterprises that need verifiable, consistent outputs with confidence scores and bounding boxes. The model handles text, images, audio, files, and video via a Chat Completion API compatible with OpenAI, Vercel AI, LangChain, and n8n. Key features include state-of-the-art OCR (scoring 70.7% on OCRBench V2, top among competitors), speech-to-text with word-level accuracy, multilingual support in 100+ languages, code sandboxing, and configurable guardrails. The architecture combines specialized models and tools, routing each request to the best model for accuracy and speed. It offers a 1M token context window, 32K max output tokens, streaming, reasoning, and function calling. Infrastructure costs (sandbox, browser engine) are folded into token counts, with free caching included. Pricing is pay-as-you-go at $1.50 per million input tokens and $3.50 per million output tokens. A self-hosted option is available on request. Recent updates include OCR improvements, a new diffusion audio ASR model, Gemini 3.5 benchmarks, and Postgres LLM integration. Benchmarks show it leading on GPQA Diamond (92.4%), MMMLU (90.9%), and SOB Value Acc (80.5%) versus models like GPT-5.4-Mini and Claude-Sonnet-5. Compared to general-purpose models like GPT-4o or Claude, Interfaze prioritizes accuracy and verifiability over creative generation. It is best suited for production-grade extraction tasks where auditability matters, not for open-ended chat or creative work. Its deterministic outputs and confidence scores make it a strong choice for regulated industries and data pipelines requiring rule-based validation.