Dirac topped the Terminal-Bench-2 leaderboard for
gemini-3-flash-previewwith a 65.2% score!
It is a well studied phenomenon that any given model's reasoning ability degrades with the context length. If we can keep context tightly curated, we improve both accuracy and cost while making larger changes tractable in a single task.
Dirac is an open-source coding agent built with this in mind. It reduces API costs by 64.8% on average while producing better and faster work. Using hash-anchored parallel edits, AST manipulation, and a suite of advanced optimizations. Oh, and no MCP.
Our goal: Optimize for bang-for-the-buck on tooling with bare minimum prompting instead of going blindly minimalistic.
Dirac is benchmarked against other leading open-source agents on complex, real-world refactoring tasks. Dirac consistently achieves 100% accuracy at a fraction of the cost. These evals are run on public github repos and should be reproducible by anyone.
🏆 TerminalBench 2.0 Leaderboard: Dirac recently topped the Terminal-Bench-2 leaderboard with a 65.2% score using
gemini-3-flash-preview. This outperforms both Google's official baseline (47.6%) and the top closed-source agent Junie CLI (64.3%). This was achieved without any benchmark-specific info or anyAGENTS.mdfiles being inserted.
Note on the cost table below: A bug was discovered in Cline, the parent repo, after running these evals (issue #10314). We have submitted a PR #10315 to fix this. This bug caused the evals for Dirac and Cline to slightly underreport the numbers ($0.03 vs $0.05 per million token cache read). Although there won't be a large difference, we will update the evals soon.
All tasks for all models used gemini-3-flash-preview with thinking set to high
| Task (Repo) | Files* | Cline | Kilo | Ohmypi | Opencode | Pimono | Roo | Dirac |
|---|---|---|---|---|---|---|---|---|
| Task1 (transformers) | 8 | 🟢 (diff) [$0.37] | 🔴 (diff) [N/A] | 🟡 (diff) [$0.24] | 🟢 (diff) [$0.20] | 🟢 (diff) [$0.34] | 🟢 (diff) [$0.49] | 🟢 (diff) [$0.13] |
| Task2 (vscode) | 21 | 🟢 (diff) [$0.67] | 🟡 (diff) [$0.78] | 🟢 (diff) [$0.63] | 🟢 (diff) [$0.40] | 🟢 (diff) [$0.48] | 🟡 (diff) [$0.58] | 🟢 (diff) [$0.23] |
| Task3 (vscode) | 12 | 🟡 (diff) [$0.42] | 🟢 (diff) [$0.70] | 🟢 (diff) [$0.64] | 🟢 (diff) [$0.32] | 🟢 (diff) [$0.25] | 🟡 (diff) [$0.45] | 🟢 (diff) [$0.16] |
| Task4 (django) | 14 | 🟢 (diff) [$0.36] | 🟢 (diff) [$0.42] | 🟡 (diff) [$0.32] | 🟢 (diff) [$0.24] | 🟡 (diff) [$0.24] | 🟢 (diff) [$0.17] | 🟢 (diff) [$0.08] |
| Task5 (vscode) | 3 | 🔴 (diff) [N/A] | 🟢 (diff) [$0.71] | 🟢 (diff) [$0.43] | 🟢 (diff) [$0.53] | 🟢 (diff) [$0.50] | 🟢 (diff) [$0.36] | 🟢 (diff) [$0.17] |
| Task6 (transformers) | 25 | 🟢 (diff) [$0.87] | 🟡 (diff) [$1.51] | 🟢 (diff) [$0.94] | 🟢 (diff) [$0.90] | 🟢 (diff) [$0.52] | 🟢 (diff) [$1.44] | 🟢 (diff) [$0.34] |
| Task7 (vscode) | 13 | 🟡 (diff) [$0.51] | 🟢 (diff) [$0.77] | 🟢 (diff) [$0.74] | 🟢 (diff) [$0.67] | 🟡 (diff) [$0.45] | 🟢 (diff) [$1.05] | 🟢 (diff) [$0.25] |
| Task8 (transformers) | 3 | 🟢 (diff) [$0.25] | 🟢 (diff) [$0.19] | 🟢 (diff) [$0.17] | 🟢 (diff) [$0.26] | 🟢 (diff) [$0.23] | 🟢 (diff) [$0.29] | 🟢 (diff) [$0.12] |
| Total Correct | 5/8 | 5/8 | 6/8 | 8/8 | 6/8 | 6/8 | 8/8 | |
| Avg Cost | $0.49 | $0.73 | $0.51 | $0.44 | $0.38 | $0.60 | $0.18 |
🟢 Success | 🟡 Incomplete | 🔴 Failure
Cost Comparison: Dirac is 64.8% cheaper than the competition (a 2.8x cost reduction).
* Expected number of files to be modified/created to complete the task.
See evals/README.md for detailed task descriptions and methodology.
- Hash-Anchored Edits: Dirac uses stable line hashes to target edits with extreme precision, avoiding the "lost in translation" issues of traditional line-number based editing.
- AST-Native Precision: Built-in understanding of language syntax (TypeScript, Python, C++, etc.) allows Dirac to perform structural manipulations like function extraction or class refactoring with 100% accuracy.
- Multi-File Batching: Dirac can process and edit multiple files in a single LLM roundtrip, significantly reducing latency and API costs.
- High-Bandwidth Context: Optimized context curation keeps the agent lean and fast, ensuring the LLM always has the most relevant information without wasting tokens.
- Autonomous Tool Use: Dirac can read/write files, execute terminal commands, use a headless browser, and more - all while keeping you in control with an approval-based workflow.
- Skills & AGENTS.md: Customize Dirac's behavior with project-specific instructions using
AGENTS.mdfiles. It also seamlessly picks up Claude's skills by automatically reading from.ai,.claude, and.agentsdirectories. - Native Tool Calling Only: To ensure maximum reliability and performance, Dirac exclusively supports models with native tool calling enabled. (Note: MCP is not supported).
Install Dirac from the VS Code Marketplace.
Install the Dirac CLI globally using npm:
- Authenticate:
- Run your first task:
dirac "Analyze the architecture of this project"
You can provide API keys via environment variables to skip the dirac auth step. This is ideal for CI/CD or non-persistent environments:
ANTHROPIC_API_KEYOPENAI_API_KEYOPENROUTER_API_KEYGEMINI_API_KEYGROQ_API_KEYMISTRAL_API_KEYXAI_API_KEY(x.ai)HF_TOKEN(HuggingFace)- ... and others (see
src/shared/storage/env-config.tsfor the full list).
dirac "prompt": Start an interactive task.dirac -p "prompt": Run in Plan Mode to see the strategy before executing.dirac -y "prompt": Yolo Mode (auto-approve all actions, great for simple fixes).git diff | dirac "Review these changes": Pipe context directly into Dirac.dirac history: View and resume previous tasks.
- Open the Dirac sidebar in VS Code.
- Configure your preferred AI provider (Anthropic, OpenAI, OpenRouter, etc.).
- Start a new task by describing what you want to build or fix.
- Watch Dirac go!
Dirac is open source and licensed under the Apache License 2.0.
Dirac is a fork of the excellent Cline project. We are grateful to the Cline team and contributors for their foundational work.
Built with ❤️ by Max Trivedi at Dirac Delta Labs