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.
📖 Read the full source: HN AI Agents
👀 See Also
Graphthulhu MCP Server Gives AI Agents Knowledge Graph Memory for Logseq/Obsidian
Graphthulhu is an open-source MCP server that provides AI agents with read-write access to Logseq or Obsidian vaults, storing memory as structured pages with properties and links instead of vector embeddings. After one month, the system generated 404 pages with 1,451 cross-references.
OpenClaw Integration for Indian Stock Markets: Multi-Agent Analysis and Trading Terminal
An open source trading terminal for Indian markets has been wired up as an OpenClaw skill server, allowing any OpenClaw agent to pull Indian stock market data and run full analysis over HTTP without local installation. The system uses seven specialist agents working in parallel to generate structured analysis with trade plans.
Claude Session Tracker: Auto-Save Claude Code Sessions to GitHub Issues
A new tool called claude-session-tracker automatically saves Claude Code sessions to GitHub Issues, logging every prompt and response as comments with timestamps. It creates one GitHub Issue per session linked to a Projects board and works through Claude Code's native hook system without consuming context tokens.
Mymir: An Open-Source Project Graph for Coding Agents via MCP
Mymir gives agents a graph-based project map with dependencies, decisions, acceptance criteria, and previous execution notes, delivered via MCP to avoid re-explaining state across sessions.