Announcements
Introducing Incredible Small 1.0 — Agentic AI Model That Can Take 1000+ Actions Simultaneously and Work With Large Datasets
Sep 10, 2025
Today, we’re introducing our first Agentic Model, designed to take actions in your apps. While others optimize for academic intelligence, we’ve optimized this model for real work. Incredible Small 1.0 is a significant leap in performance over other SOTA models in its performance to efficiently perform workflows, process and analyze data, while showing near-zero hallucination on data. Small 1.0 is available as a research preview at https://www.incredible.one, and also as an API at https://web.incredible.one/for-developers.
"Despite very limited resources, we've created models that compete with world-leading AI companies where it matters most: getting real work done," said Philip Alm, co-founder and CEO of Incredible. "We spent three years solving a different problem than everyone else. While others optimize for general intelligence with billions in funding, we optimize for how much real work models can actually complete.”
VIDEO: Incredible Small 1.0 Taking a Meeting Transcript in Google Docs and creating a Full Project Management board in Trello. (2 minutes)
Parallel Actions
While traditional LLMs are limited to a few reliable actions at a time, Incredible Small 1.0's Live-Code Model Architecture enables hundreds of parallel actions with minimal latency. This breakthrough eliminates the bottlenecks of token-based function calling, delivering enterprise-scale performance that transforms how AI systems handle complex, multi-step workflows.
Data-critical Tasks
Incredible Small 1.0 achieves near-perfect accuracy when working with provided or extracted data, virtually eliminating hallucinations in data-critical tasks. From precisely referencing API identifiers to transferring large datasets, the model maintains data integrity across simple and complex operations—setting a new standard for reliable AI-powered data processing.
Needle in a Haystack
Incredible Small 1.0 maintains exceptional accuracy when locating specific information within large, complex datasets. Whether finding a single transaction in millions of records or identifying patterns across sprawling databases, the model delivers precise results without the degradation typically seen in traditional models when processing large contexts.
Data as Context
Traditional AI models treat everything as text tokens, creating a bottleneck where large datasets consume hundreds of thousands of tokens, making enterprise-scale processing prohibitively expensive. Incredible's Live-Code Architecture processes data alongside standard text tokens through symbolic representation, allowing our models to fully utilize datasets when taking actions and conducting analysis at a fraction of the traditional cost—enabling AI that can work with gigabytes of information rather than just the small samples that fit within traditional context windows.
Incredible — 2 min demo: Transcript to Full Project Management
Why this Matters
Every company building with LLMs faces the same fundamental constraints. Your applications are limited by models that hallucinate with critical data, choke on large datasets, and require complex orchestration just to execute basic multi-step workflows. These aren't edge cases—they're dealbreakers that force you to build elaborate safeguards, restrict functionality, or abandon AI features entirely.
Incredible Small 1.0 is a drop-in replacement that eliminates these constraints. Simply swap your API endpoint and immediately unlock:
Reliable data processing with near-zero hallucination on your actual business data
Enterprise-scale workflows with hundreds of parallel actions instead of sequential bottlenecks
Massive cost savings—67% reduction on output tokens plus efficiency gains from processing gigabytes of data directly
True automation that completes complex tasks without breaking down or requiring constant human intervention
Whether you're a startup building your first AI feature or an enterprise scaling existing applications, you don't need to rebuild your infrastructure or retrain your team. Your existing OpenAI, Anthropic, or Google API calls work exactly the same—but now they actually work at scale.
"If you want an LLM to write code, use Claude. If you want PhD-level intelligence, use GPT. If you want AI that takes actions in products and gets work done, use Incredible," said Emil Wagman, Head of Development at Incredible.
Competitive Performance & Benchmarks
Even though no formal benchmarks exist yet for parallel execution and massive data handling capabilities, these innovations translate to strong performance on traditional benchmarks.
Unmodified TAU-bench: Measures how well AI models can use tools to complete real-world tasks like re-booking flights or shopping, simulating actual user workflows in customer support scenarios.
IFEval (Instruction Following Eval): Tests whether models can precisely follow complex, multi-constraint instructions, measuring instruction adherence.
Incredible Small 1.0 outperforms SOTA models on function calling, while matching their reasoning capabilities in knowledge benchmarks.
"Despite very limited resources, we've created models that compete with world-leading AI companies where it matters most: getting real work done," said Philip Alm, co-founder and CEO of Incredible. "We spent three years solving a different problem than everyone else. While others optimize for general intelligence with billions in funding, we optimize for how much real work models can actually complete.”
Availability and Pricing
Incredible Small 1.0 is available now in early research preview at $1.15 per million input tokens and $5.00 per million output tokens—a 67% cost reduction on output compared to Claude Sonnet and 50% reduction compared to GPT-5. When accounting for data processing efficiency improvements, the cost savings are even more significant.
Access Options:
AI Assistant: 200+ app connectors at incredible.one
API Access: Drop-in replacement for OpenAI/Anthropic/Google APIs at docs.incredible.one/quickstart
MCP Support: Full Model Context Protocol compatibility for agent developers
The company plans to release Incredible Big, its flagship model, in the coming weeks with substantial improvements across all benchmarks.
Note: All benchmarks reflect unmodified, raw model performance without task-specific fine-tuning. Results can be independently validated following instructions in the respective papers. TAU-bench (arxiv.org/abs/2406.12045), IFEval (arxiv.org/abs/2311.07911), GPQA (arxiv.org/abs/2311.12022)
Note: SWE-LLM 2024, accessible on HuggingFace (https://huggingface.co/viss-ai/LLama3-70B-SWE-LLM)
