A coding agent that gets frontier scores out of DeepSeek
AgenC Core is an open source coding agent for your terminal. With DeepSeek V4.1 Flash it scored 76.2% on Terminal-Bench 2.1 for $34. The agent that matched that score on a frontier model cost $198.
The numbers
- Terminal-Bench 2.1: 76.2% (339 of 445 trials) for $34. That ties mini-SWE-agent with Muse Spark 1.1 ($198) and beats Codex with GPT-5.6 Luna max ($241) on the Harbor Hub board. See the run.
- The only DeepSeek run on that board as of October 14, 2026.
- Terminal-Bench 2.0: 71.5% for $23.84 across all 445 trials, about 5 cents a trial. See the run.
- Every trial is public, with full trajectories. When the agent looked up an answer online, we counted the trial as a failure.
Why it costs so little
- Light mode keeps the prompt lean and its opening stable, so DeepSeek serves most of it from cache. On the Terminal-Bench 2.0 run, 97.9% of input tokens came from cache.
- When a reasoning model runs out of output room, AgenC recovers and keeps going instead of returning an empty answer.
- A token budget per task (eco, balanced or max) caps what any one task can spend.
Same model, against Pi
On the 30 hardest Terminal-Bench 2.0 tasks, both agents on DeepSeek Flash with the same settings, AgenC solved 18 and Pi solved 10. Eight tasks were solved only by AgenC, and none only by Pi. AgenC used about 2.8 times the tokens to get there, because it keeps working where Pi stops.
Set it up
Install AgenC Core, give it your DeepSeek API key, and start it with the DeepSeek provider:
curl -fsSL https://get.agenc.ag/install.sh | sh export DEEPSEEK_API_KEY=your-key agenc --provider deepseek
You can also start agenc and pick DeepSeek with /provider. AgenC works with any model provider, so you can switch later. The Terminal-Bench 2.1 run used DeepSeek V4.1 Flash at reasoning effort max, in light mode (agenc --light, still experimental).
Open source under the MIT license. Source code ยท Verify your download