China’s decision to grant the first conditional permits for Level‑3 (L3) automated driving at the end of December 2025 has shifted a technical debate into the commercial and regulatory mainstream. The Ministry of Industry and Information Technology’s authorisations for Changan’s DeepBlue SL03 and BAIC’s Arcfox Alpha S crystallised what many in the sector had anticipated: L3 is now a live commercial pathway, not just a laboratory milestone.
Built atop a mass-market foundation of advanced driver assistance, the L3 moment has been fuelled by rapid L2 adoption. Domestic data show that by the first three quarters of 2025 roughly two‑thirds of new passenger cars shipped with combined driving‑assist features — a 21.2% year‑on‑year rise and a penetration rate around 64%. That large installed base is prompting Chinese manufacturers and suppliers to press for the next leap in vehicle autonomy.
But the industry is divided about the route forward. One camp treats L3 as a necessary, legally framed bridge between driver‑in‑charge L2 systems and fully autonomous L4 operations. Proponents argue that limited operational design domains (ODDs) allow firms to trial real‑world incident handling, refine insurance and liability models, and nudge consumers’ expectations. OEMs such as BAIC and Changan, and firms planning L3 production like Lantu, Chery and GAC, are pursuing a “regulation‑first, scenario‑limited” strategy to commercialise L3 on highways and controlled expressways.
An opposing school of thought is willing to leapfrog L3 entirely. Tesla’s product taxonomy (AP, EAP, supervised FSD) and its Robotaxi trials are cited as models for bypassing a formal L3 phase, and Xpeng’s leadership has publicly forecast that next‑generation fully automated systems will emerge in 2026, skipping the interim label. These actors prefer to treat autonomy as a capability that scales across use cases rather than a rigid SAE level to be rung sequentially.
The tension between the two approaches is sharpened by how regulators are framing certification. International standards used to validate L3 systems — including tests derived from the EU’s ECE R157 — already demand fail‑safe behaviours that look suspiciously like L4 responsibility logic. For instance, when a human driver fails to respond to a takeover request, certified systems must execute a minimum‑risk manoeuvre, such as a safe stop. Academics at Tongji University warn that this effectively forces L3 applicants to reach near‑L4 technical performance to obtain approval, even if they operate commercially within a narrower ODD.
That narrowness is consequential. The Arcfox L3 permission is limited to single‑lane segments of highways and urban expressways with a top speed cap of 80 km/h and an initial B2B orientation. Such constraints blunt consumer appeal and raise questions about how much real‑world value early L3 products will deliver beyond specific fleet applications.
Hybrid technical narratives complicate the picture further. Several industry executives frame L3 not as a discrete rung but as an emergent user experience produced by widespread deployment of high‑end L2++ functions (notably city NOA) plus targeted L4 operation in confined zones such as Robotaxi corridors. In short, users may perceive the mobility experience as largely autonomous even if the legal classification varies by situation.
Market forces are pressing both narratives. City NOA installations in China surged through 2025, reaching more than 3.1 million passenger cars from January to November. Third‑party autonomy stacks — notably Momenta and Huawei’s HI solution — now account for leading shares of that market, with monthly fitment volumes entering the tens of thousands. Whether manufacturers pursue a staged L3 deployment or a direct L4 push, their success will turn on reconciling regulatory thresholds, insurance and liability frameworks, cost curves and, above all, demonstrable safety and public trust.
The immediate result is not a uniform industry pivot but a multipronged race. Some players will use the L3 license gate to accumulate operational data and harden system behaviours; others will place bigger bets on software‑centric scaling and Robotaxi validation to shortcut regulatory orthodoxies. Either way, the coming 18–36 months should reveal whether L3 becomes a durable commercial stepping stone or a regulatory artefact that the market ultimately outflanks.