April Updates: #1 on SWE-bench Lite; New coin-based usage system; Simple UI to connect your LLMs (BYOK)
April was big. Here’s what we shipped at Refact.ai:
- 1st spot on SWE-bench Lite leaderboard
- New coin-based usage system
- Better UI for external models through BYOK
- A must-see webinar on how AI Agents work
- … plus a bunch of other updates and improvements.
We’ve summed it all up for you — scroll down to read our April Changelog!
Refact.ai Agent is a new SOTA on SWE-bench Lite
Refact.ai Agent is now #1 on SWE-bench Lite: 59.7% score, 179 tasks solved — fully autonomously!
Our approach is based on:
- AI Agent that solves tasks end-to-end without human input [prompt]
- Claude 3.7 Sonnet as a core model
deep\_analysis()tool, powered by o4-mini, for reasoning on harder problems- A suite of tools AI Agent calls when needed
- Focused iteration to generate 1 correct final result.
We built Refact.ai for devs like you — so this milestone is as much yours as ours. Thanks for being part of the community and inspiring us to keep improving our AI Agent!
New usage system: Coins instead of requests limits
Refact.ai switched to a usage-based pricing with coins — a fairer system where you only pay for what the Agent does.
- 1.000 coins = $1
- Quick tasks = fewer coins, complex tasks = more
- No more requests caps, or full-request burns on one-line fixes.
In your plugin, the rollout to the coins system is automatic.
How many coins do you get?
Free plan users: A one-time pack of 5.000 coins to start.
Pro plan users: 10.000 coins each month + 10.000 coins as a bonus. Unused coins roll over month to month. Plus, you can do coin top-ups.
📌 Please read our Docs and blog post for more details. Also, we’re here to answer your questions in Discord’s #help-and-debug and #feedback channels
Better BYOK interface with access to more AI models
Connect your own API keys or local models to Refact.ai Agent in seconds with the updated BYOK (Bring your own key):
- Simple UI under Settings → Configure providers
- Works with AI models from OpenRouter, Anthropic, OpenAI, LM Studio, Ollama, or custom endpoints
- Optional model settings (context window, limits).
connected, your model appears in the dropdown — ready to work with AI Agent.
💡 Tip: If you have subscriptions to some AI models, try to connect them via BYOK and make the most of your stack.
Q: Will BYOK use my Refact.ai coins?
A: No. BYOK usage is billed by your model’s provider, your Refact.ai coin balance stays untouched.
More updates in Refact.ai
✨ New models: GPT-4.1, Gemini 2.5 Pro; GPT-4.1-mini and GPT-4.1-nano (work only in chat) are now available.
🗂 Improved file attaching: New ‘Attach’ button allows adding multiple files to agent conversations instead of just one. Open files to add them to the context in a few clicks — and remove as needed.
⚪️ Follow-up controls: Added user toggle controls for AI-suggested follow-up questions in the chat interface.
Enterprise/OSS updates
🔑 Configurable third-party APIs: Set up any chat model supported by litellm (available from JetBrains 6.3.2, VSCode 6.3.5).
👨🦳 Model deprecation: starcoder2, deepseek-coder, refact. The always-up-to-date models lineup is in the Known Models list on GitHub.
How to build AI Development Agent? Watch the webinar recording
Want to know how autonomous coding actually works?
In our latest webinar, Nick Frolov (Head of Product @Refact.ai) explains the core tech behind AI agents — function calling, context windows, token limits, RAG, AST, memory, etc. — and how it all fits together in Refact.ai Agent.
Catch the recording on YouTube and follow Refact.ai on X and LinkedIn for the upcoming announcements of live events!
New content to empower your AI programming
Real devs, real results. Explore how Refact.ai helps users like you build fully functional systems, handling most of the programming behind.
1️⃣ How to create an MCP Server using Refact.ai Agent [read]
2️⃣ How I built an automated SEO Content Generator in 60 minutes [read]
🎬 Software Testing with AI Agent: QA Global Summit’25 by Geekle [watch]
Try the Latest Updates in Refact.ai
Your VSCode or JetBrains plugins are already live with all the new features. Just open your IDE — and lets vibecode something great!