Back to get.agenc.ag

76.2% on Terminal-Bench 2.1 for $34

AgenC Core is an open source coding agent for your terminal. Its public Terminal-Bench 2.1 run with DeepSeek V4.1 Flash scored 76.2% for $34. Every entry on the Harbor Hub leaderboard spent at least $134.

Chart of score against cost on Terminal-Bench 2.1. AgenC Core with DeepSeek V4.1 Flash sits at the top, 76.2% for $34. The leaderboard entries all cost $134 or more. — Score against the cost of a full run, cheaper runs at the top. Leaderboard data from Harbor Hub, October 14, 2026.

Score and cost, entry by entry

Each row pairs an agent with a model, so this compares combinations, not agents alone.

Agent	Model	Effort	Score	Cost
Codex	GPT-6 Astra	high	87.4%	$774
Codex	GPT-6 Astra	medium	87.0%	$619
Codex	GPT-6 Astra	low	86.7%	$463
Codex	GPT-6 Astra	max	86.7%	$1,070
Codex	GPT-6 Astra	xhigh	85.8%	$889
Claude Code	Fable 5	xhigh	83.8%	$553
Codex	GPT-5.5	xhigh	83.2%	$2,059
Terminus 2	Fable 5	high	80.5%	$439
Cursor CLI	Grok 4.5	high	79.3%	$134
Claude Code	Opus 4.8	high	78.9%	$287
Codex	GPT-5.6 Terra	max	78.4%	$421
Terminus 2	GPT-5.5	xhigh	78.0%	$494
AgenC Core	DeepSeek V4.1 Flash	max	76.2%	$34
mini-SWE-agent	Muse Spark 1.1	xhigh	76.2%	$198
Codex	GPT-5.6 Luna	max	75.7%	$241
Claude Code	Sonnet 5	high	74.6%	$288
Terminus 2	Gemini 3 Pro	high	73.9%	$224
Claude Code	Opus 4.7	max	68.9%	$600
Terminus 2	Opus 4.7	max	66.1%	$582
Gemini CLI	Gemini 3.1 Pro	high	65.8%	$236
Gemini CLI	Gemini 3 Pro	high	65.8%	$248
Terminus 2	Gemini 3.1 Pro	high	65.6%	$230
Claude Code	GLM-5.1	max	58.6%	$277

How we ran it

The official Terminal-Bench 2.1 package: 89 tasks, 5 trials each, 445 trials, with the official timeouts and resources.
When the agent fetched an answer or a solution online, we counted that trial as a failure. Late finishes count as failures too. The score before those flags was 81.3%.
$34 is the cost Harbor Hub lists for the run, the same basis as the other entries. At DeepSeek's list prices for the recorded request times it came to $35.54.
All 445 trials are public with full trajectories: the Terminal-Bench 2.1 run.

Head to head with Pi 1.0

Table of AgenC Core against Pi 1.0 with the same model and tasks. 30 hardest tasks with DeepSeek Flash: 18 solved against 10, using 2.8 times the tokens. GPT-6 Luna max with the Luna fix: 24% fewer tokens and 17% fewer calls. Warm start: 107 ms against 137 ms. — Same model, same tasks, alternating run order, independent audit. One run per task.

Try it

AgenC Core works with any model provider. To run it the way it ran here, see AgenC Core with DeepSeek.

Install AgenC Core

Open source under the MIT license. Source code · Verify your download