EN / 中文

GPT-6 Astra Drives Real Vehicles, Spark Debate on General Large Model vs Mass-produced AD Systems

by gaishiqiche·November 16, 2026

"If you cannot transform into a large model company, I believe there is no opportunity." Recently, Zhou Guang, CEO of DeepRoute.ai, reiterated this judgment in an interview.

This is not the first time he has made such a statement.

In 2024, he predicted that the autonomous driving industry must transition to being driven by large models, otherwise it would face dimensional reduction attacks. At that time, however, most people were skeptical.

"When I spoke about this back then, people might not have believed it would happen. After GPT-6, everyone should realize: this is happening now."

What brought renewed attention to this judgment was the real-vehicle test completed by GPT-6 Astra: it was the first time that "general-purpose large models can control real vehicles to complete driving tasks" was verified in a visibly tangible way.

Zhou Guang estimates the window period is two to three years.

Before this timeframe ends, autonomous driving companies that have not completed their transition to the large model paradigm will face the risk of elimination.

Some view these remarks as a wake-up call, while others consider them excessive anxiety among peers.

A 134-Meter Slow Motion Reveals Two Coordinate Systems

In September, during a public test conducted by independent evaluation agency DrivingBench, four general-purpose large models—GPT-6 Astra, Claude Fable 5.1, Grok 4.6, and GPT-5.6 Sol—were tasked with directly controlling a 2022 Toyota Corolla via tool interfaces to complete a closed-course route: starting, navigating around cones, reversing, and parking, covering a total distance of 134.7 meters.

Ultimately, GPT-6 Astra completed the course on its second attempt in 5 minutes and 22 seconds, with an average speed of approximately 1.5 km/h; other models had completion rates ranging from 6% to 45%.

During the same period, on the robotics simulation leaderboard RoboDojo, GPT-6 Astra, which had not undergone specialized robot training, also ranked first.

Its method of completing tasks—an average speed of 1.5 km/h, closed courses, simple routes, and requiring multiple attempts—has almost no reference value in real-world road scenarios. DrivingBench itself acknowledged that these tasks were "very simple," involving no complex traffic flows, random pedestrians, or continuous driving on urban roads.

However, its symbolic significance lies in the fact that a general-purpose large model without specific training for driving controlled a real vehicle through tool interfaces to complete a full set of actions for the first time.

This had never happened before September 2026, providing visible evidence for Zhou Guang's 2024 judgment for the first time.

Placing this event against the backdrop of China's actual autonomous driving landscape, the differences between the two coordinate systems become immediately apparent.

The current capabilities of general-purpose large models: closed courses, low speeds, simple routes, and requiring multiple attempts. The current level of mass-produced autonomous driving systems in China: urban roads, average speeds of 20-30 km/h, leading players achieving MPI (Miles Per Intervention) at the hundred-kilometer level, and daily usage by millions of users for commuting.

However, another aspect of the market structure is equally noteworthy.

In the first half of 2026, Huawei, Tesla, and Momenta held the top three positions in China's high-level autonomous driving market share, with the remaining share distributed among more than ten players.

This is a market that has not yet consolidated. If the judgment that "only a few will remain in the future" holds true, it implies that industry reshuffling has just begun; conversely, no single player has achieved absolute dominance, leaving space for new entrants and those undergoing transformation.

Zhou Guang's judgment is based more on the slope of technological evolution rather than a comparison of current capabilities.

Zhou Guang explained that when GPT-2 and GPT-3 first emerged, outsiders also thought they were unimpressive, but just two years later, the situation was completely different. This pattern of evolution differs from previous ones.

This also means that while general-purpose large models start from a low baseline in driving tasks, whether their rate of progress will surpass traditional autonomous driving approaches is the core point of industry disagreement.

No Autonomous Driving Company Has Stayed Still

To determine if the term "traditional autonomous driving companies" is accurate, one must look at the technical routes publicly disclosed by various companies in 2026.

Huawei Qiankun's choice is the WEWA 2.0 architecture. At the Huawei Qiankun Technology Conference in April 2026, ADS 5 was officially released, described by officials as "upgrading comprehensively from rule-driven to AI agents oriented toward autonomous driving." The cloud world engine introduced a "multi-agent game" mechanism, and online reinforcement learning achieved "generation, learning, and verification simultaneously," improving training efficiency by 10 times; the vehicle-side world behavior model introduced safety risk field theory, generating real-time risk heat maps.

In September, Jin Yuzhi disclosed that Qiankun's annual R&D budget was increased from CNY 18 billion at the beginning of the year to over CNY 19 billion. By August, the total number of vehicles equipped with Huawei Qiankun intelligent driving exceeded 2 million units.

Momenta's public route combines "one flywheel, two legs" with world models. In April 2026, the R7 world model was released. Cao Xudong stated at the launch event: "Prediction is the cornerstone of intelligent evolution; the core of physical AI lies in deep cognition of fundamental physical laws of the world."

According to prospectuses and public disclosures, Momenta has accumulated over 12 billion kilometers of real-vehicle mileage and more than 100 million segments of golden training data, conducting pre-training and post-training of world models based on real data from 800,000 mass-produced vehicles.

In late October, Cao Xudong offered a judgment in public: scaling L4 requires approximately 100 billion kilometers of data. In July 2026, Momenta went public on the Hong Kong Stock Exchange.

Tesla's path involves parallel development of data scale and end-to-end models. On May 21, 2026, supervised FSD became available in China; overseas markets shifted to a USD 99/month subscription model starting in February, while mainland China currently maintains a CNY 64,000 one-time purchase option, with the timeline for launching subscriptions unpublished.

DeepRoute.ai chose to pivot to foundation models. DeepRoute did not select the World Model route but instead developed Foundation Models. Zhou Guang explained, "There is no need for such large parameter counts for generative tasks." He used a vivid analogy: humans may have evolved for a long time to develop language, requiring a brain capacity of 1,500cc; many animals might only need 2ml, yet they still operate within world models.

He believes VLA (Vision-Language-Action) is a sub-task of foundation models, not the endpoint; what truly needs breakthroughs is the "upper limit of intelligence," where simple models fail in long-tail scenarios requiring causal reasoning. Public data shows that DeepRoute ranked second in the industry in third-party insurance registrations in August 2026.

Looking at these four routes together, one commonality is obvious: none of the leading autonomous driving players in 2026 remain stuck in traditional modular/rule-driven architectures. Whether it is Huawei's WEWA, Momenta's R7 world model, Tesla's end-to-end FSD, or DeepRoute's foundation model, all public statements point towards the direction of large models.

The real divergence is not about "whether to do it," but revolves around three questions:

Parameter count and architecture: Which is more efficient—general foundation, vertical world model, or extreme VLA route? Weight of the data flywheel: To what extent does data scale naturally emerge intelligence, or must the correct paradigm exist first? Commercialization pace: Should we scale vehicle installations first and iterate capabilities, or perfect the model first and then pursue scale?

These are open industry-level questions with no public answers currently.

The End of the Window Period: 2027 Is a Critical Year

The "two to three year window period" mentioned by Zhou Guang coincides with several confirmed industry milestones.

The first milestone is the mandatory national standard for L3. On July 30, 2026, GB 44721-2026 "Safety Requirements for Intelligent Connected Vehicle Autonomous Driving Systems," a mandatory national standard, was published and will take effect on July 1, 2027, applying to L3/L4 level vehicles. This is China's first mandatory safety standard specifically for L3/L4, meaning compliance thresholds shift from encouraging policies to hard entry requirements.

The second milestone is the scaled deployment of Robotaxis. Zhou Guang noted a phenomenon: "Scalable Robotaxi deployment is very fast, similar to Tesla's rollout speed. Some companies took 10 years to deploy one or two thousand units, while others started deploying Robotaxis like pouring dumplings into boiling water." He judged that moving MPI from hundreds to thousands, and then to tens of thousands of kilometers are key steps for L4 scaled deployment; "Today, regardless of the L4 system, there are human operators in the background, using various cloud-based safety strategies layer by layer. When reaching the ten-thousand-kilometer level, scaled deployment can basically begin."

The third milestone is the validation of payment models for autonomous driving. The 75% subscription rate for FSD overseas, the actual conversion rate of the CNY 64,000 one-time purchase option in the Chinese market, and data on optional installation rates for domestic automakers' intelligent driving features—these figures are answering a fundamental question: Are consumers willing to continuously pay for high-level autonomous driving? This answer directly impacts the business models of all autonomous driving companies.

Putting these milestones together, 2027 is clearly a critical year, a consensus within the industry.

As for the question of "how many years traditional autonomous driving companies have left," preliminary answers may only be seen by mid-2027, when L3 compliant models hit the market, Robotaxi operating cities expand, the mass production performance of next-generation models from various companies, and the performance of the next version of GPT in driving tasks collectively provide a clearer industry landscape than today.

Over the past two years, the most common topics in the autonomous driving industry have been "whose end-to-end approach is more advanced" and "whose urban NOA drives smoother." These competitions essentially represent involution within the same technical paradigm. However, Zhou Guang's latest judgment shifts the industry's focus from "who is doing it better" back to "should we switch tracks."

Interestingly, Huawei is upgrading WEWA, Momenta is developing the R7 world model, Tesla's FSD is itself an end-to-end large model, and DeepRoute is betting on foundation models. Everyone is changing, just with different postures and rhythms.

The real divergence has never been about "whether to transition to large models," but rather "using models of what size, pairing them with how much data, and pushing them to market at what pace."

As Zhou Guang stated, the steam engine era of autonomous driving has arrived.

The industry is waiting for a new answer. Before that arrives, everyone sitting at the table is already rearranging their chips.