Tsinghua's PilotDeck 2.0 Beats Claude Code on Agent Bench While Cutting Task Costs to One-Sixth

Verdict: Worth a spin if you run long agent workflows and the bill is the bottleneck. PilotDeck 2.0, open-sourced October 17 by Tsinghua's THUNLP lab with ModelBest and the OpenBMB team under AGPL-3.0, is an agent operating system built around three ideas that attack cost directly: physically isolated workspaces, a white-box memory you can inspect and edit, and strong-and-weak model routing — the expensive model plans, the cheap ones execute. The routing numbers from its launch coverage are the interesting part: on 7 hard benchmark tasks, a single Claude Sonnet 4.6 run scored 69.1 at $18.36, while PilotDeck's routed combo (Sonnet 4.6 planning, MiniMax-M2.7 executing) scored 70.6 at $3.15 — better result, one-sixth the cost. That's a simulation result from the team's own coverage, not my own run, but the pattern matches what anyone who has watched an agent burn Opus tokens on reformatting already suspects.

What PilotDeck 2.0 actually adds

Per the official repo and OpenBMB's launch announcement, PilotDeck is a model-agnostic agent operating system — you bring your own models, including local ones. Version 2.0's headline features, as reported in Chinese launch coverage by 量子位:

• Workspace isolation. Each project gets its own file system, memory store, and skill set — a per-project sandbox so writing a technical doc for one client doesn't leak tone and context into the next task. The retrieval range is bounded per workspace, which is also a latency win.
• White-box memory. Every memory the agent forms is visible, editable, deletable, and pinnable. A "Dream" mode consolidates and deduplicates scattered memories during idle time, with one-click rollback if the consolidation drops something important. This is the anti-black-box take on agent memory — instead of writing a prompt begging the model to forget a wrong parameter, you just fix the record.
• Always-on execution. Long-running tasks keep going after you close the window, and the agent surfaces candidate follow-up tasks on its own, dropping results as local files with a summary report.
• Native MCP support across web UI, CLI, and IM clients (WeChat Work, Feishu) with identical backend behavior.

The cost math, in one table

The numbers from 量子位's coverage of the launch, on the team's Workbuddy Bench and a 7-task hard benchmark:

Workbuddy Bench: PilotDeck running Qwen3.6-35B-A3B scored 79.1, ahead of Claude Code's reported 77.3 — notable if it holds, since Qwen3.6-35B-A3B is a small active-parameter MoE you can host yourself, not a frontier API.

The 7-task hard set: MiniMax-M2.7 alone was cheap ($1.90) but scored 37.1 — unusable. Claude Sonnet 4.6 alone scored 69.1 but cost $18.36. The routed combination scored 70.6 for $3.15. In a lighter social-media-ops test, an all-Opus-4.5 pipeline cost $12.58 per run; routing top-level planning to Opus 4.5 and text polishing down to Sonnet 4.5 cut it to $2.83, roughly a 70–77% saving depending on which pair you look at.

My read: the pattern is real but the exact multiples are workload-dependent. Routing pays when the task decomposes into a small planning core and a large execution tail — translation, formatting, retrieval, summarization. It pays little when every step needs frontier judgment, and it adds a failure mode: a weak sub-agent that quietly mangles execution while the planner never sees it. The white-box memory helps audit that, which is arguably the more durable feature than the cost split.

Who should try it, and who shouldn't

Try it if you're running recurring multi-step workflows (monitoring, content pipelines, report generation) where 80% of tokens go to mechanical execution — this is exactly the profile where strong-and-weak routing compounds. It ships for macOS, Linux, and Windows with Docker, and AGPL-3.0 means you can self-host freely as long as you keep modifications open — same license family as Grafana, not the MIT-style terms of MiniCPM-V 4.7, so read the terms before embedding it in a closed product. Skip it if your agent usage is short, single-shot Q&A — there's no cross-task memory to amortize and routing overhead buys you nothing. And if you're weighing the bill side of coding agents specifically, our Claude API heavy-user cost breakdown covers that math directly.

Caveat (Oct 18, 2026): PilotDeck 2.0's existence, licensing (AGPL-3.0), model-agnostic design and workspace concept are verified against the official GitHub repo and OpenBMB's launch announcement; all benchmark scores and cost figures come from 量子位's launch coverage of team-provided data and were not independently reproduced. Self-hosting hardware requirements are in the repo README.