Q0-Preview Tops Fable 5.2 Using Old Claude Opus 5.5 via Agent, Not VLA/WAM

Background: Fable 5.2 Sets New Benchmark

On November 10, Anthropic released Fable 5.2, achieving 64.9% accuracy on the RoboDojo Sim dual-arm robot simulation benchmark, a significant improvement over Fable 5.1 and well ahead of GPT-6 Astra. The model demonstrated capabilities like bottle-cap twisting, Rubik's cube manipulation, and drawing, highlighting physical action as a new frontier for foundation models.

Q0-Preview Surpasses Fable 5.2 with an Older Model

As of November 11, the RoboDojo Sim Agent leaderboard shows Q0-Preview (submitted October 30) at 68.87% task success rate and a comprehensive score of 74.2, making it the first entry to break 70. Remarkably, Q0-Preview is driven by Anthropic's previous-generation model Opus 5.5, which officially scores only 39.4% on the same benchmark. This raises a deeper question: where are the boundaries of foundation models' physical action capabilities, and how can spatial understanding and reasoning be fused with embodied know-how?

Core Philosophy: Spatial Computing Is Embodied Operation

Q0-Preview's central thesis: determining the contact location and contact pose for each manipulation stage. For a simple task like grasping an apple, after identifying the target object, the agent must decide which part of the apple to grasp, from which direction the gripper should approach, and with what orientation. Converting these judgments into a target end-effector pose and then into executable actions solves the problem.

General multimodal agents already possess strong multimodal perception and coding abilities. Q0-Preview leverages these by using single- or multi-view observations to identify target pixel locations, then applying camera intrinsic/extrinsic parameters and pixel-coordinate calculation tools to perform spatial computing—mapping image-space contact targets into the world coordinate system.

Specifically, the gripper's target pose and contact information are expressed via three spatial reference points: the midpoint of two ordered points defines the reference contact position, the line between them defines the grasping direction, and a third point supplements the approach direction. Thus, position, orientation, and contact information are unified into a computable spatial-computing problem.

The multimodal agent decides the current sub-task based on observation and instruction, submits the target position, orientation, and gripper command. Downstream motion planners like cuRobo convert these into continuous actions. After execution, the agent inspects the new observation, checks the result, and adjusts the next step, forming an "observe–compute–execute–feedback" closed loop.

Harness: From "Can Compute" to "Can Do"

Around this core idea, Q0-Preview builds a Harness that standardizes interaction between the model, observations, tools, and execution modules.

1. Define Execution Boundaries and Tool Calls

Unify the observations, robot state, and prior information available to the model. Specify thinking and tool-call budgets, and the output format for control commands. Tasks advance in short action chunks, with visual checkpoints at necessary intermediate stages to decide whether to continue or replan. The model is regulated to call existing or self-written programs for image cropping, local zoom, and other tools that assist spatial computing.

2. Accumulate General Manipulation Experience via Simulation Trial-and-Error

During development, simulation trial-and-error yields general operational rules for the model to reference. Examples: using the idle arm's wrist camera to assist observation and reduce occlusion; planning paths that balance collision avoidance with subsequent visibility; adjusting motion speed according to grasp, transport, and place phases. These skills transfer from simulation to real robots.

3. Supplement Task Requirements and Clarify Completion Conditions

For tasks with vague descriptions, Q0-Preview adds static task rules, completion conditions, and necessary operational suggestions to reduce the gap between model understanding and evaluation requirements. The agent never reads real-time rewards or environment judgment signals from the evaluator during execution.

4. Manage Context and Task Memory

Retain initial information, recent observations, and a compact task memory that records confirmed progress, current state, and future goals, controlling context-growth overhead. Explicit interfaces isolate upstream decision-making from downstream execution, enabling continuous chaining of multi-stage operations.

Demonstrated Capabilities

Q0-Preview shows impressive performance across multiple dimensions, in some tasks exceeding the smoothness of expert teleoperation.

1. Dual-Arm Synchronous Operation

Set targets for left and right arms within the same action stage, advancing parallelizable operations simultaneously.

2. Dual-Arm Collaborative Operation

Arrange division of labor around a shared goal, coordinating execution based on dependencies between supporting, handover, and other actions.

3. Precise Manipulation

Determine position and orientation through spatial computing from multi-view observations, then correct deviations using new observations after approach.

4. Adaptive Adjustment

Re-localize targets, adjust poses, or revise plans based on object state changes and action execution results.

5. Memory

Retain previously acquired information and completed steps to inform later decisions, enabling tasks that depend on historical information.

Open Challenges and Future Directions

While Q0-Preview demonstrates the potential of frontier foundation models in embodied manipulation, several hard challenges remain for real-world deployment.

Inference Efficiency and Cost

Frontier models may consume millions of tokens and hours of inference time for seemingly simple embodied operations. Such time and economic costs are still far from practical deployment requirements. Reducing redundant reasoning and tool calls while preserving manipulation capability is a must-solve problem for embodied agents.

Source of Embodied Capabilities

A more fundamental question: where does the model's embodied manipulation ability actually come from? GPT-6 Astra's entry into embodiment sparked speculation that its training included massive embodied ego-data, driving a leap in spatial understanding and manipulation. However, Q0-Preview's success with Opus 5.5 forces us to ask: which factors—embodied data, code generation, tool use, or 3D generation—contribute most to physical manipulation ability? How do the model's innate capabilities and the Harness design interact? These cannot be answered by a single leaderboard score; systematic ablation experiments are needed to guide future data construction, training objectives, and system design.

Latency and Dynamic Adaptation

Long inference latency limits current embodied agents' ability to adapt to dynamic environments. In the real world, objects and environments do not wait for the model to finish thinking; changes between observation and execution can invalidate plans. Timely response to dynamic changes, fusion of visual, force, and tactile feedback, and maintaining operational reliability over long runs amid disturbances and failures remain open research directions.

Despite these challenges, the zero-shot generalization shown by foundation models opens a new technical path for embodied intelligence. Q0-Preview's exploration suggests that combining general model capabilities with embodied domain knowledge and constraints can continuously expand robotic task boundaries. The team plans to release a technical report, faster and cleaner code, more real-robot experiments, and further discussion of current method limits and future directions.

GitHub repository:

https://github.com/XPolicyLab/XPolicyLab/tree/main/policy/Q0_Preview