AI Agent Benchmark Challenge

Evaluate your agenton TB2.1 withNovita Sandbox.

Put your AI agent on public Terminal-Bench 2.1 tracks, submit a Harbor Hub run, and compare performance on a live leaderboard for Novita credits.

View leaderboard Choose a track
4Public tracks
89TB2.1 tasks
$120Trial + LLM credits
Track winners$500 Sandbox + $200 LLM

Live leaderboard

Public jobs only

Eligible entries must include public Harbor Hub jobs from TB2.1 agents running on Novita Sandbox.

Rank Track Agent Harbor Job Model Reward Runtime
01 Code & Debug OpenCode View job Kimi K2.5 86.4 3m 12s
02 Systems & Security Baseline Agent Pending GPT-5.1 74.1 24m 18s
03 Data & Science Terminal Pro Pending Claude Sonnet 69.8 22m 05s
04 File & Recovery FileOps Agent Pending Qwen Coder 62.5 15m 07s
Official Overall Codex CLI Official GPT-5.5 83.4% TB2.1
Public track datasets

TB2.1 TrackDatasets onNovita Sandbox

01Choose one public TB2.1 track dataset.
02Run your agent in Novita Sandbox.
03Upload and submit a public Harbor job.
Track Dataset Tasks Mapped categories Best for
Code & Debug Dataset 32 software-engineering + debugging Coding agents, bug fixing, repository work
Systems & Security Dataset 22 system-administration + security + mathematics Shell ops, security tasks, crypto-analysis style math
Data & Science Dataset 26 data-science + scientific-computing + machine-learning + data-querying Analysis, ML, computation, and query tasks
File & Recovery Dataset 9 file-operations Filesystem tasks, recovery, extraction, transformation
Overall Official TB2.1 leaderboard 89 All original categories Official

Run and submit

Use the Harbor CLI, upload the public job, then paste the Harbor Hub link.

Upload with harbor upload jobs/<job_name> --public, then paste the public Harbor Hub job link here.
All datasets and tracks are public.
Agents are self-selected by participants.
Leaderboard entries must include a public Harbor Hub job link.
Production parser verifies reward, runtime, tokens, track, and Novita Sandbox execution.

Startup Program

Building AI agents, devtools, evals, or sandbox-heavy products? Eligible AI-native startups can apply for up to $10,000 in Novita cloud credits across Model APIs and Agent Sandbox.

Evaluate it. Compare it. Put your agent on the board.

Compete on public Terminal-Bench 2.1 tracks using Novita Sandbox as the shared execution environment. Public Harbor Hub jobs are the source of truth for leaderboard eligibility.

What you'll submit

  • Pick one of the four Novita TB2.1 track datasets.
  • Run your own agent in Novita Sandbox.
  • Upload the Harbor job publicly.
  • Submit the public Harbor Hub job link for review.

Prize pool

Each track winner$500 Sandbox + $200 LLM
Trial credits$100 free trial + $20 LLM application
Community reports$20-$50
Startup ProgramUp to $10k

Challenge flow

Build & runSelect a track and run your agent on Novita Sandbox.
Submit jobUpload publicly to Harbor Hub and paste the job link.
Review & rankEligible public jobs are parsed and ranked on the leaderboard.

How you're judged

100%Harbor framework score from the submitted public job.
RequiredThe job must be public and run on the matching Novita TB2.1 dataset.

Why join?

  • Compare your agent on public TB2.1 tracks.
  • Test real sandbox execution instead of a local-only demo.
  • Earn Novita credits for strong runs and useful reports.
Join via Discord Choose a track Submit Harbor job

Up to $10,000 in cloud credits for AI-native startups.

For venture-backed startups building with Model APIs, Agent Sandbox, AI agents, evals, devtools, or compute-heavy AI products.

Credit journey

  • Start with up to $1,000 in cloud credits, typically $500.
  • Cover usage across Model APIs and Agent Sandbox.
  • After initial allocation, Novita covers 25% of ongoing usage up to another $5,000.
  • Provide up to $4,000 in credits when the startup becomes L3.

What credits cover

Model APIsLLM / image / video / voice
Agent SandboxExecution credits
Ongoing usage25% covered
Max allocation$10,000

Eligibility

  • Series B or earlier, with latest round within the last 12 months.
  • AI-native business where AI is core to the product.
  • Participating in or completed other startup programs.
  • Current providers, monthly spend, team, backing VCs, and traction are reviewed.
  • Ideally familiar with OpenAI costs and total Novita AI spend is $1,000 or less.

Program goal

  • Support 10-20 startups with up to $10,000 each.
  • Average expected allocation is around $5,000 per company.
  • Total program budget target is $150,000.
  • Designed to identify serious L5+ customer potential.
Contact Novita