# Superbrain - Most Efficient AI Coding Agent Built for Teams > Superbrain is the most efficient AI coding agent built for teams. It works across massive codebases with persistent context, using 68% fewer tokens and 47% lower cost than Claude Code on a typical bug, measured on SWE-bench Lite. ## What is Superbrain? Superbrain is the most efficient AI coding agent, built for engineering teams. Its proprietary TokenFold retrieval architecture maps and compresses an entire codebase, so it works across massive repositories and monorepos with persistent context, using 68% fewer tokens and 47% lower cost than Claude Code on a typical bug, measured on ten django bugs from SWE-bench Lite. Made by Open Gigantic. ## Key facts - Category: AI coding agent (developer tool). - Efficiency: 68% fewer tokens and 47% lower cost than Claude Code on a typical bug (median over the eight bugs both agents fixed, SWE-bench Lite django subset, both on Claude Sonnet 5, network open on both sides). Full data at /benchmarks. - Context: persistent, infinite context across large codebases and monorepos via TokenFold. - Workflow: issue to implementation to tests to pull request, with no change applied without approval. - Models: on plans with included usage Superbrain runs on Kilo.Flux, its router, which picks the model per request from a pool spanning Claude, GPT-5.6, Gemini, DeepSeek, Kimi, GLM and Grok and bills one published rate whichever answers; the Forge plan runs every model on the customer's own API keys. Details at /product/flux-kilo-1. - Pricing: a free plan with included usage; Prevue, a paid plan with more usage and extra usage on demand; Forge, which runs on the customer's own API keys with no usage limits. Cancel any time. Prices per region (India in INR with GST, elsewhere in USD) at /pricing. - Platforms: Windows and macOS (Apple Silicon) today, downloaded directly from /download; Linux is planned. - Maker: Open Gigantic, a research-first deep-tech company. ## Full text - https://onesuperbrain.com/llms-full.txt: the questions and answers, every plan and price by region, the Flux rate card and pool, and every benchmark run down to the issue, in one document. ## Pages - [Superbrain - Most Efficient AI Coding Agent Built for Teams](https://onesuperbrain.com): What Superbrain is, how TokenFold cuts tokens 68% and cost 47% on a typical bug versus Claude Code, measured on SWE-bench Lite, features, benchmarks, and FAQ. For engineers and engineering teams evaluating AI coding agents. - [Benchmarks vs Claude Code, Codex and OpenCode on SWE-bench Lite](https://onesuperbrain.com/benchmarks): Head to head benchmark results for Superbrain against Claude Code on ten django issues from SWE-bench Lite: 68% fewer tokens and 47% lower cost at the median, 8 of 10 resolved against 8 of 10, with per-issue data, failure analysis and stated limitations. For engineers evaluating AI coding agents. - [Research on Token-Efficient AI Agents](https://onesuperbrain.com/research): Open Gigantic's arXiv research on token-efficient agents: ObjectGraph (.og) file format (up to 95.3% token reduction) and Phase-Scheduled Multi-Agent Systems. For AI researchers and engineers. - [Download Superbrain for Windows and macOS](https://onesuperbrain.com/download): Where to download Superbrain: Windows and macOS (Apple Silicon) are available now, free during the beta, downloaded directly (no access code); Linux is planned. For developers who want to install and run Superbrain. - [Pricing: Free, Prevue and Forge Plans](https://onesuperbrain.com/pricing): Superbrain's plans, prices and what each includes: a free plan with included usage, Prevue with a larger monthly balance and extra usage on demand, and Forge, which runs every model on the customer's own API keys with no usage limits. Prices are quoted per region (India in INR with GST, elsewhere in USD). For anyone comparing what Superbrain costs. - [Introducing Kilo.Flux](https://onesuperbrain.com/product/flux-kilo-1): The release note for Kilo.Flux, Superbrain's model router: what it does (classifies each request as Swift, Standard, Deep or Frontier work and routes it to the best model for it), the model pool it routes across (Claude Haiku, Sonnet 5, Opus 4.8 and Opus 5; GPT-5.6 Luna, Terra and Sol; Gemini 3.8 Flash; DeepSeek V4 and V4.1 Flash; Kimi K3; GLM 5.3 Flash; Grok 4.6), how it caches and recovers, its published rate per million tokens (input, output, cached input) charged whichever model answers, and that it is included on every Superbrain plan with usage. For anyone asking what Flux is or how Superbrain picks models. - [Contact the Superbrain Team](https://onesuperbrain.com/contact): How to contact the Superbrain / Open Gigantic team (sales, partnerships, support, press) and the ways they can reach you back. For anyone wanting to talk to a human at Superbrain. - [Cookies](https://onesuperbrain.com/cookies): Superbrain's cookie policy: which cookies the site sets (consent record, anonymous visitor id, sign-in session, Razorpay for payments, preferences, Google Analytics, Product Hunt badge), what each is for, how long it lasts, what is recorded with analytics consent, and how to change the choice. For anyone asking what Superbrain tracks. - [Changelog and Release Notes](https://onesuperbrain.com/changelog): Release notes for Superbrain: the latest features, fixes, and improvements shipped during the beta. For existing users tracking updates. ## Research - [ObjectGraph: From Document Injection to Knowledge Traversal — A Native File Format for the Agentic Era](https://arxiv.org/abs/2604.27820v1) — 2604.27820, 12 pages. - [Phase-Scheduled Multi-Agent Systems for Token-Efficient Coordination](https://arxiv.org/abs/2604.17400v1) — 2604.17400, 8 pages.