不用专门训练自动驾驶 GPT-6 Astra已经能开真车了？

Key conclusion: GPT-6 Astra, a general-purpose large model, was able to control a real Toyota Corolla through 134.7 meters in a parking lot. However, this was a closed-course, low-speed experiment, not a real driver's license test, and it does not mean the model can drive autonomously on public roads.

1. Among the four models evaluated, Astra was the only one to complete the full course; Claude Fable 5.1's best result was 45%, Grok 4.6 reached 11%, and GPT-5.6 Sol reached 6%.

2. Astra's second attempt took 5 minutes and 22 seconds, for an average speed of about 0.42 m/s (around 1.5 km/h). It drove very slowly.

Practical takeaway: A large model can adjust its driving strategy based on lessons from its first failure without retraining, but safety boundaries still require careful attention.

1. In the first round, Astra covered only 67.3 meters (about 49% of the course). After reviewing the attempt, it kept speed below 0.8 m/s and used 100% steering in 20 of 24 motion commands, ultimately finishing.

2. The model observes the environment through a camera and MCP tools, sets direction and speed, stops immediately, and controls the vehicle through openpilot; in effect, it acts as the driving brain.

3. Because the test area was closed, the speed cap was 3.5 m/s, each model ran only one set of continuous context, and the camera had blind spots, ordinary users should not equate this result with mature autonomous-driving capability.

Product development: The general-purpose large model demonstrated an ability to understand the real world and operate physical equipment, giving brands a possible path to intelligent products without necessarily relying on a specially trained model.

1. GPT-6 Astra completed a 134.7-meter real-car route by forming a closed loop from perception to execution using camera visual input, MCP tools, and the underlying control system. This shows that a large model can act as the central brain of intelligent hardware.

2. It was not retrained for driving. Instead, it adjusted speed and steering in the second attempt based on learnings from the first failure, highlighting the value of in-context learning for product experience refinement.

Consumer trends and user perception: Safety, cost, and real-world constraints will directly shape consumers' trust in AI-driven driving.

1. The model initially refused to control a real vehicle for safety reasons; after the tool was renamed DrivingBench Sandbox, refusals decreased. This demonstrates that safety boundaries and prompt-induced behavior are both product-design issues.

2. The test speed was very low, averaging only 0.42 m/s. One trial consumed about 6.6 million tokens and cost about $7.74. If brands use large-model capability as a selling point, they should avoid asking consumers to pay a high premium for a concept.

3. This result cannot be directly compared with urban NOA or Robotaxi. For brands, the more relevant takeaway is the general-purpose model's future potential for physical-world operations, not immediate mass production deployment.

Growth opportunity: General-purpose large models are beginning to operate real vehicles, signaling that AI applications are expanding from computer tasks to physical-world tasks. Sellers should watch for new services and product opportunities enabled by this capability.

1. DrivingBench allowed models such as GPT-6 Astra to control a Toyota Corolla through a camera and standard tools. Astra was the only model to finish, showing that a large model can act as a low-barrier physical-operation brain.

2. The test was billed by token consumption: completing 134.7 meters used about 6.6 million tokens and cost $7.74. This cost structure gives physical-world AI services a quantifiable measurement basis.

Risk and response: The results are still experimental and should not be treated as evidence that the ability to drive a real car is commercially deployable.

1. The test took place in a closed parking lot with no other vehicles, pedestrians, or traffic signals. Speed was capped at 3.5 m/s, and each model ran only one continuous context, so the sample size was limited.

2. The model has limitations including safety refusals, camera blind spots, and insensitivity to nearby obstacles, making direct deployment in customer delivery projects risky.

3. A useful lesson comes from Astra's behavior after its first failure: it deliberately reduced speed and increased steering angle. Sellers should first test in sandbox scenarios with a small area, low speed, safety supervisors, and human intervention, then expand gradually.

Product design and hardware requirements: This real-car test revealed the hardware chain needed to connect a large model to vehicle control. Factories building intelligent devices should pay attention to interface compatibility and sensor configuration.

1. The test vehicle accessed the CAN bus through a comma four device and an OBD-C interface, with low-level control built on openpilot. Future AI-driving-related products will need standardized data interfaces and drive-by-wire execution capabilities.

2. The model perceived the environment through a forward-facing camera, but the camera had clear blind spots that made it difficult to see nearby obstacles or the vehicle's sides and rear. This creates a clear requirement for sensor placement and surround-view coverage.

Business opportunity and digitalization implications: Large models are becoming control brains for the physical world, offering factories opportunities in intelligent hardware and automation upgrades.

1. GPT-6 Astra does not directly generate steering torque or brake pressure. Instead, it controls the vehicle through standard tool actions: observing the environment, setting direction and speed, and stopping immediately. This division between a large-model brain and low-level execution creates supporting demand across controllers, actuators, and sensors in the supply chain.

2. After its first failure, Astra reviewed the attempt in context and adjusted its speed and steering strategy on the second run without retraining parameters. Production-line digitalization can borrow this approach of continuously improving strategy from operational data.

3. The test remains a closed, low-speed demonstration. If factories plan to invest in related products, they should validate in small batches rather than treating it as mature mass-production readiness.

Industry trend and new technology: General-purpose large models are moving into control of real physical devices, an important signal in the current embodied-intelligence and physical-AI trend.

1. DrivingBench evaluated GPT-6 Astra, Claude Fable 5.1, Grok 4.6, and GPT-5.6 Sol. Only Astra completed the 134.7-meter course, showing that foundation models differ sharply on real driving tasks.

2. The technical architecture combines model understanding, MCP tool calls, and openpilot execution to enable observation of the environment, setting direction and speed, and immediate stopping. Its modular design can be extended to other robotics scenarios.

3. After failing in the first round, Astra was not retrained. Instead, it reviewed its attempt in the same context, lowered speed below 0.8 m/s, and increased steering angle in the second round. This demonstrates the value of in-context learning for physical control.

Client pain points and solutions: Safety, cost, and stability are the three core issues clients need to solve.

1. The model initially refused to execute because of safety concerns, and refusals decreased after the tool was renamed DrivingBench Sandbox. Service providers should help clients build real operational boundaries rather than relying on name changes or prompt-level workarounds.

2. The current approach averages about 0.42 m/s and consumes about 6.6 million tokens and $7.74 per run. Both efficiency and cost fall short of real-time, product-grade autonomous driving requirements.

3. Potential supporting services include closed-course benchmarking, multi-model capability assessment, sensor blind-spot compensation, safety-supervisor intervention processes, and failure attribution optimization.

Platform opportunity and approach: A third-party experiment pushed a large model originally designed for computer tasks into real-vehicle driving. This shows that model platforms should actively explore physical-world use cases.

1. When OpenAI released GPT-6 Astra, it emphasized computer operation, browser use, software engineering, and scientific research, with no mention of driving. DrivingBench, however, showed that Astra can control a real vehicle through MCP tools, pointing to a new direction for expanding the developer ecosystem.

2. Platforms can standardize actions such as observing the environment, setting direction and speed, and stopping immediately as callable tool interfaces, making it easier for automakers and developers to connect to low-level execution systems.

Operational management and risk mitigation: Physical operation carries higher risk than text tasks, so platforms need greater caution in capability access, billing, and risk control.

1. The model refused to control a real vehicle for safety reasons, but refusals decreased after the tool was named DrivingBench Sandbox. Platforms should prevent model behavior from being steered around safety limits through naming or descriptions.

2. The four models' results were clearly different: one completed the course, while the other three reached 45%, 11%, and 6%, respectively. This shows that physical-world benchmark testing is needed to select and manage model capabilities; demo performance alone is insufficient.

3. One drive consumed about 6.6 million tokens and cost about $7.74, at a speed of only 0.42 m/s. Platform designers should factor the economics of high-token-consumption tasks into quota, billing, and developer support policies.

Research proposition: This experiment verified that a general-purpose large model can form a closed loop of perception, judgment, planning, and control through a real vehicle, offering a concrete sample of foundation models entering the physical world.

1. Among the four models, GPT-6 Astra was the only one to complete the 134.7-meter course; Claude Fable 5.1, Grok 4.6, and GPT-5.6 Sol posted best results of 45%, 11%, and 6%, respectively.

2. The system was assembled using comma four, OBD-C, CAN bus, openpilot, and MCP tools. The large model observes the environment, sets direction and speed, and stops, while low-level execution is handled by conventional systems.

New questions and implications: In-context learning, safety boundaries, and limited samples are the key research themes.

1. After its first failure, Astra reviewed the attempt in the same context, lowered the speed cap to 0.8 m/s, and used 100% steering in 20 of 24 commands. This shows that physical-control strategies can be optimized without retraining.

2. The model initially refused to control a real car, and refusals decreased after the tool was renamed DrivingBench Sandbox. This suggests that superficial factors may affect how the model judges real-world consequences, raising a safety-governance question.

3. The experiment has significant limitations: a closed parking lot, no traffic participants, a speed limit of 3.5 m/s, only one continuous test per model, and camera blind spots. Its significance is not replacing autonomous driving but showing that general-purpose models are gradually acquiring physical-world operation ability, providing a reference for world-model and embodied-intelligence research.