folkbench.com continuously tests AI API relays and publishes their availability, latency and pricing. This repository holds the code for Folkbench Switch, the companion desktop client that writes the route you pick into Claude Code, Codex and other tools.

---

Sponsors

These relays support Folkbench's ongoing benchmarks. Become a sponsor →

	Modelflare Operated by US-based Havenbyte LLC. Supports major models (GPT, Claude, Gemini, Grok) and standard OpenAI-compatible protocols, alongside 40+ image and video generation models. Features ordered fallback across multiple models per single API Key, secondary pool groups, and comprehensive request audit logs tracking status, latency, tokens, and actual expenditures. Cache compensation guarantee: Select OpenAI pools guarantee daily cache hit rates of 65% (Price-optimized), 75% (Stable), and 85% (Premium); qualifying requests falling below these thresholds receive compensatory credits. First-time top-ups unlock discounted tiers: a single US$20 deposit unlocks the 0.015-multiplier GPT discount pool (GPT-5.6-Sol at approx. ¥0.51 input / ¥3.01 output per 1M tokens, with no usage cap); a single US$50 deposit unlocks the 0.03-multiplier Claude discount pool.
	CheapAI Focused squarely on budget efficiency, particularly through its discounted pools. Features an innovative dynamic multiplier pool that adjusts rates in real time based on pool capacity—dropping rates when resources are plentiful and balancing cost versus availability during peak periods. Offers discount tiers for cost-sensitive developers alongside Stable, Pro, and direct Official Key pools to accommodate varying workload criticalities.
	IKunCode Optimized for seamless integration into AI coding CLIs (Claude Code, Codex, Gemini CLI). Provides comprehensive support channels including email, QQ groups, and Telegram with attentive human assistance for beginners.
	FluxLane (智流) An AI API gateway for developers and teams: full-capability, stable access to global models, supporting GPT, Claude, Gemini, DeepSeek, GLM, and other mainstream models as well as major global model-vendor protocols. The platform provides policy-based routing and multi-route failover, making long-term integration and cost accounting easier for teams in China. For enterprise and university procurement in China, it can sign contracts, accept wire transfers, and provide invoices, purchase orders, and related materials (subject to commercial confirmation). Publishes model capability evaluation reports, with multipliers as low as 0.1×, and new registrants receive a $2 trial credit.

---

Relay Rankings

So many relays — which are stable, which are cheap, which leave your data alone? Success rate, latency, price and data-security level, all on one board. Overall · stability · value.

Popular Models

If you need benchmark data for a specific model across different providers, explore the per-model leaderboards:

Model	Live Benchmark Data
GPT 6 Sol	View benchmarks across providers
GPT 6 Astra	View benchmarks across providers
Claude Fable 5.1	View benchmarks across providers
Grok 4.7	View benchmarks across providers
DeepSeek V4.1 Flash	View benchmarks across providers

Folkbench Switch

A beginner-friendly setup tool for AI coding assistants. Set up Claude Code, Codex, Gemini CLI and more in one click — simple and easy. Download for macOS / Windows.

macOS: Apple Silicon / Intel (.dmg); Windows: x64 (.exe). Get installers and SHA256 checksums from Releases.

Switch is Folkbench’s open-source companion desktop app (built with Tauri and Rust). Its core purpose: switch Base URLs and API Keys between multiple proxy providers and official endpoints with one click—leaving tedious manual config edits behind.

🎯 Key Features

• 🔄 One-Click Provider Switching: Save multiple Base URLs and API Keys, apply configurations to target tools in a single click, and switch or restore at will.
• 🛡️ Lossless & State Preservation: Preview configuration diffs before applying. When switching Codex to a third-party relay, it automatically stashes your official ChatGPT OAuth session securely and restores it seamlessly when reverting.
• 📈 Integrated Benchmark Metrics: View live availability and latency from Folkbench probes directly within the client before switching endpoints.
• 🔒 Local Storage & Privacy: All API Keys and configurations remain strictly on your local disk; zero telemetry or cloud syncing. Fully offline-capable.
• 📊 Local Usage Estimator: Tracks local token consumption by session and estimates costs based on reference pricing.

🧰 Supported Tools

Switch adapts to the configuration schemes and directory conventions of each tool, touching only network parameters and preserving all other custom user settings:

Tool	Config File / Path	Notes
Claude Code	~/.claude/settings.json	Injects Base URL and intelligently handles Bearer token vs. API Key auth modes
Codex	~/.codex/config.toml	Injects custom model providers while safely stashing official credentials for lossless restoration
Gemini CLI	~/.gemini/.env & settings.json	Synchronizes environment variables and authentication schemes
Aider	~/.aider.conf.yml	Writes model names and API endpoint options
OpenCode	~/.config/opencode/	Injects provider mappings and authentication keys
Kimi CLI	~/.kimi-code / ~/.kimi	Automatically handles legacy and current directory structures
DeepSeek Harness	~/.dsh/	Configured for official CLI environments
Others	MiniMax Code, Qwen Code, Grok Build, Hermes, Pi, etc.	Continuously maintained and expanded

---

Community

File bugs and feature requests in GitHub Issues, or join Telegram or the QQ group (203286875).

License

Switch code under apps/switch/ is licensed under Apache-2.0. Blog posts, images and brand assets are excluded; see LICENSE.md.