In recent years, the competition in Navigation on Autopilot (NOA) has essentially revolved around one thing: whose definition sounds more advanced.
Labels like L2.5, L2.9, the number of cities covered, and "near-autonomous driving" did not need to be verified; they only needed to be believed.
Now, this path is closed. GB 47955—2026 Safety Requirements for Combined Driving Assistance Systems in Intelligent Connected Vehicles has been approved and released, and will become mandatory for new models applying for type approval starting January 1, 2027.
When compliance shifts from a bonus point to an entry ticket, the question becomes simpler: once everyone holds the same entry ticket, how can NOA still compete?
The Era of Marketing Rhetoric Ended by a Standard
Why did regulators have to step in at this specific moment?
At the 2026 World Intelligent Connected Vehicle Conference held in October, Guo Shougang, Director-General of the Equipment Industry Department I of the Ministry of Industry and Information Technology (MIIT), disclosed that as of August this year, approximately 61 million passenger vehicles in China were equipped with combined driving assistance (L2) features, of which about 15 million possessed Navigation on Autopilot (NOA) capabilities.
Earlier, in July, he publicly mentioned that since the beginning of 2026, the penetration rate of L2 combined driving assistance in passenger vehicles reached 70.5%, while the penetration rate of NOA reached 34.2%.
This implies that when 70% of new cars are equipped with these systems, "exaggerated marketing by individual companies" is no longer an isolated issue but a road safety concern.
Consider the weight of this standard. It is China's first mandatory national standard for L2 combined driving assistance.
It is not an industry initiative, nor a group standard; it is an admission requirement.
A noteworthy point in the MIIT's interpretation is that compared with UN Regulation R171, the Chinese standard imposes stricter requirements on Operational Design Domain (ODD), functional requirements, driver state monitoring, and user notification, and adds certain field test scenarios.
This is not merely alignment; it adds an extra layer above the international baseline.
In recent years, NOA marketing materials have largely centered on three things: vague tier rhetoric, the number of cities covered, and implications of "near-autonomous driving."
The mandatory standard blocks these three points individually: systems are categorized into basic single-lane, basic multi-lane, and navigation assistance tiers, so tiers are no longer marketing buzzwords; NOA allows automatic lane changes and ramp entries/exits, but prohibits lane changes over solid lines and requires early exit from autonomous lane changing near intersections; physical buttons return, driver state is monitored progressively, and the system is disabled for at least 30 minutes after multiple violations; Data Storage System for Automated Driving (DSSAD) recording becomes mandatory; and "the driver always bears driving responsibility" is written into the standard's principles.
The MIIT positions this standard as "baseline requirements ensuring the safe operation of combined driving assistance systems," setting specific requirements for basic single-lane, basic multi-lane, and navigation assistance products respectively. Above this baseline, differences in traffic efficiency, comfort, anthropomorphism, and complex scenario coverage can still span generations.
It is clear that the mandatory standard does not end experience competition; it ends claim competition.
Compliance no longer creates differentiation, but without compliance, there is no eligibility to participate. The other side of this statement is that all differentiation is forced into the layer above compliance.
And that layer is beyond the reach of standards. Information and slogans fail here; only two things remain usable: what you dare to promise, and what you dare to invest.
Competing on Responsibility: From "Willingness to Use" to "Willingness to Pay"
An entry ticket solves whether you can sell, but it cannot answer why people buy.
Zhang Xiaoyu, Executive Vice President of Changan Automobile, provided a set of figures at the TEDA Forum: the current penetration rate of Urban NOA (Navigation Assist System) is 20.9%, but the usage rate is less than 6.3%.
A survey by Soochow Securities offers another perspective. Only 31% of car owners frequently use Urban NOA, 20% explicitly stated they do not dare to use it, and over 70% rarely use it or have never activated it.
Penetration rate indicates presence in the vehicle; usage rate indicates human willingness to use it. The former is achieved through configuration lists, while the latter can only be accumulated through error-free experiences.
In May 2026, BYD announced something no mainstream automaker had done before: providing a one-year safety guarantee for the urban navigation features of its God's Eye A and B systems.
Prior to this, in February 2025, BYD made God's Eye a standard feature across all models priced above CNY 100,000 and available on most models below that price under the "Intelligent Driving for All" initiative.
More than a year later, backing up the urban navigation scenarios corresponding to God's Eye A and B changed the weight of the term "standard": moving from "the car has this feature" to "if something goes wrong using this feature, the automaker covers it."
BYD Chairman Wang Chuanfu described this as "taking on the responsibilities of L3 and L4 stages ahead of time during the L2 stage."
This is the first truly competitive aspect after the mandatory standard: responsibility.
When feature claims are constrained by standards, whoever dares to back up real performance with their balance sheet can widen the gap in the dimension of "willingness to use." This relies not on parameters, but on promises.
However, the value of a promise cannot stop at a launch event. There are two questions that must be asked.
First, has it changed user behavior?
BYD itself provided the answer. After the smart parking guarantee was launched last July, the usage rate of the parking function rose from 21% to 93%, with an accident rate approaching zero. Although this was parking rather than urban navigation, the mechanism is the same: a guarantee is not an insurance product; it eliminates trial-and-error costs for users. Insurance sells risk to users; a guarantee buys risk back onto the company.
Second, under what circumstances is compensation not paid?
"Compliant use" is the keyword. Hands-off, eyes-off, exceeding ODD, failure to take over as prompted by the system, and non-compliant use may all fall within the exemption scope. The true boundary of a guarantee lies not in the four characters "full compensation," but in the length of the exemption clauses.
There is also a more professional distinction. A guarantee does not change the legal liability attribution of L2; the driver remains the responsible party, and the automaker assumes only vicarious fulfillment of civil compensation. Accident liability determination for car owners, premium records, and subsequent recourse will not disappear because of a phrase like "full compensation." However, "driver responsibility" does not mean system immunity: the guarantee buys out the user's civil compensation burden, not the automaker's own product liability risk.
Therefore, the watershed for competing on responsibility is clear. Those who shout full compensation without mentioning exemptions are doing marketing. Those who disclose boundaries and still dare to pay are demonstrating capability.
But willingness to use is only the first hurdle. The second hurdle is willingness to pay.
At the World Intelligent Connected Vehicle Conference in October this year, NIO Founder William Li provided another verification angle. He narrowed the metrics for the second half of autonomous driving down to two verifiable indicators, one of which he directly defined as "users willing to pay."
Li's original words were that if users are not willing to pay, no matter how well autonomous driving is explained, it has not truly created user value.
However, the industry is squeezing out the space for willingness to pay.
On one hand, there is price parity. The top-tier Leapmotor A10 comes with lidar and urban navigation for CNY 86,800; Horizon Robotics uses a single Journey 6M chip to break the threshold for Urban NOA down to the CNY 100,000 level; Huawei Qiankun has expanded into the CNY 100,000–150,000 range. Hardware costs have dropped to CNY 4,500–5,000, a 43% decrease over two years.
On the other hand, there is standardization. NOA has shifted from an optional high-end feature to a free gift included in all configurations, with BYD's "Intelligent Driving for All" in February 2025 being the turning point.
Thus, an awkward fact emerges: NOA was commoditized before it matured.
When it becomes a free feature, why would users pay separately for it?
This is the true commercial proposition on the demand side.
It leaves only three paths: selling responsibility (guarantees, protections, clearly written compensation boundaries); selling continuous service (subscriptions, OTA upgrades, unlocking by capability); selling stratification (splitting Urban NOA, curb-to-curb, and L3 capabilities into different price tiers).
All three paths share a common trait. They are not on the configuration list, but in renewals and repurchases.
Competing on Roadmap and Scale: Divergence in Timelines, Not Entry Tickets
After rules are unified, directions diverge.
Xpeng Founder He Xiaopeng advocates skipping L3. He believes L3 merely transfers responsibility to the system without improving safety or optimizing experience, arguing that the true endpoint is L4, with the window opening in the second half of 2028.
BYD Group Senior Vice President and Dean of the Automotive New Technology Research Institute, Yang Dongsheng, holds the opposite view. He states that L3 safety requirements are "hundreds of times higher" than L2, requiring coverage of long-tail scenarios such as tire blowouts, slippery roads, and system failures, involving overall upgrades to motors, steering, braking, chassis, and electronic architecture. The mass production window for limited-scenario L3 is also 2028.
Two roadmaps point to the same year but give opposite answers on whether the intermediate step is worth taking.
The essence of the divergence is not technical.
L3 sells responsibility transfer within limited scenarios, which can be priced now. L4 sells freedom from human intervention, waiting until after 2028. One chooses to cash in first; the other chooses to wait first.
More telling than verbal debates are automakers' product actions within the same timeframe.
On November 13, the Qijing GT7 Ultra Comfort Edition was launched, with both trims featuring a "full-link redundant architecture oriented toward L3," priced above CNY 200,000. On the same day, the Wenjie M8 launch event also featured descriptions of an L3-oriented architecture.
The commonality of these actions is: no one waited for the L3 mandatory standard to land before acting, but no one extended L3 architecture to products in the CNY 100,000 segment.
The reason is simple. Industry estimates suggest the incremental hardware cost for L3 compared to mainstream L2 is CNY 15,000–30,000. Cars in the CNY 80,000 segment cannot bear this increment, so L3 architecture will likely appear first in vehicles priced above CNY 200,000. Affordable-side NOA must remain at L2.
So-called roadmap divergence manifests simultaneously as price-band division of labor at the product level.
Mandatory national standards for L3/L4 will not be implemented until July 2027. This means the next year and a bit is an "L3 era without L3": automakers are selling L3-oriented hardware architectures and redundancy preparations, not already compliant L3 capabilities.
To distinguish genuine preparation from rhetoric, look at three things: Is the execution layer structurally dual-redundant (dual power supplies, dual steering, dual braking, dual domain controllers, truly independent and switchable)? Are there public Minimum Risk Maneuver (MRM) test results? Is there a timeline for L3 admission application? If none of these exist and only "oriented toward L3" is marked on PPTs, it is a pricing tool.
Divergence aside, both paths share the same engineering prerequisite: regardless of choosing L3 or L4, stable delivery of end-to-end models is not exempted by roadmap choice. Algorithms must be deployed across hundreds of thousands or even millions of vehicles; engineering capability is an unavoidable hurdle.
Therefore, roadmap competition is about "when responsibility can be taken," while scale competition is about "whether data loops are fast enough."
Divergence changes timelines and narratives, not entry conditions.
Whether choosing L3 or L4, everyone must pass through the same narrow gate: whether users dare to use it and are willing to pay.
Conclusion: How Can We Still Compete?
The answer may lie in three aspects: Compete on responsibility (are exemption clauses detailed enough?); Compete on roadmap (can the architecture pass a public minimum risk maneuver test?); Compete on scale (can the data loop maintain leadership under compliance constraints?).
All three share a common trait: they are neither on parameter tables nor on compliance certificates.
The mandatory standard defines the track boundaries but does not shorten the distance to the finish line. China's autonomous driving is at a critical stage transitioning from road testing to large-scale commercialization. The upcoming competition is not about who runs fastest within the boundaries, but who can get the most people willing to board within those boundaries.
Cars costing CNY 80,000 already have it. The next question is: how many people are willing to pay again for it?