Doubao Shifts to B2B: Can Its Low-Price AI Coding Strategy Break Through Fierce Competition?

Lin Ming, an Agent implementation engineer at a major internet company, was once a loyal user of Cursor over a year ago. Back then, he relied on AI tools mainly for code completion or parameter adjustments, with large models embedded directly into the editor for easy access. Today, his toolkit has evolved to include Claude Code, Codex, and his company’s self-developed AI programming platform—a shift that reflects not just tool preferences but the broader technical iteration in AI coding: from "assisted completion" to "full-process takeover."

According to Lin, early AI tools like Cursor acted more like co-pilots, requiring programmers to lead the development process. In contrast, tools such as Claude Code can now independently break down requirements, call APIs, run tests, and deliver results. This capability leap allows developers to outsource entire tasks, but it also leads to a surge in token consumption. An engineer from a big tech firm revealed that their company once offered nearly unlimited API quotas but later tightened them due to cost pressures, with many developers choosing to subscribe to tools at their own expense—an indication of the market’s growing maturity.

The commercialization path for AI coding tools is becoming clearer. Anthropic, with Claude Code, was among the first to capture the market: its annualized revenue soared from $2.5 billion in February 2026 to $47 billion in May, with enterprise clients accounting for over half of its customer base and the number of large clients doubling in two months. OpenAI, while dominating the C-end market with ChatGPT, saw Codex’s weekly active users grow sevenfold in six months, exceeding 5 million.

In the domestic Chinese market, Zhipu’s GLM5.2 model competes with Claude Opus 4.8 at only one-seventh the price. Despite generating 724 million yuan in revenue and incurring a net loss of 4.718 billion yuan in 2025, its trillion-Hong Kong-dollar market capitalization reflects market expectations for a "Chinese version of Anthropic."

ByteDance’s recent launch of Doubao 2.1 Pro marks a strategic shift toward the B2B sector. The model enhances coding and Agent capabilities, as demonstrated in live comparisons with benchmarks like Claude Opus. Tan Dai, president of Volcano Engine, noted that Doubao is no longer focused solely on user scale but is targeting the developer ecosystem and real workflows. This shift comes amid dual pressures: C-end revenue relies on e-commerce commissions, with daily income below 1 million yuan, while daily computing power costs reach tens of millions of yuan. On the enterprise side, although AI code generation rates have hit 90%, demand throughput has only increased by 60%, and code quality still has gaps in areas like UI and reliability.

ByteDance’s B2B layout covers the entire chain: front-end with the TRAE editor and plugins to capture entry points, middle layer via CLI and enterprise editions to embed into R&D processes, and bottom layer supported by Volcano Ark for computing power. Its pricing strategy延续了 C-end practices: the Coding Plan costs only 9.9 yuan for the first month, and the Pro package is 200 yuan per month—one-fifth the price of similar overseas tools. The API input price is as low as 1.2 yuan per million tokens, with a comprehensive usage cost 62.7% lower than the industry average. Crucially, Doubao supports mainstream development environments like Claude Code and Cursor, allowing developers to switch models without migrating workflows—this "low price + compatibility" strategy aims for rapid market penetration.

However, the competitive logic of the B2B market differs significantly from the C-end. Google’s case is a warning: despite having developer ecosystems like Google Cloud and Android, its Gemini Code Assist, priced as low as $19 per user per month annually, still failed to win over developers. This shows that ecosystem and price are just entry tickets; product stability, task completion rate, and developer trust are the core barriers. Hong Dingkun, ByteDance’s technical vice president, pointed out that AI coding needs to build a complete system covering context engineering, architecture constraints, and test delivery to move from "code generation" to "demand closure."

Developers have very practical criteria for tool selection. Lin Ming categorizes models into three tiers: Claude and ChatGPT for core tasks, DeepSeek for daily development, and other tools as backups. This "on-demand allocation" model is common in the industry—tech bloggers note that developer loyalty is extremely low, with tool switching requiring only a change of input box. As model capabilities converge, pricing strategy becomes a key variable. Zhipu once limited sales of its Coding Plan due to unexpected demand and deducted tokens at a higher rate during peak periods, exposing the conflict between subscription models and computing power costs.

The current track shows polarization: Anthropic and OpenAI take the high-end route, positioning their models as "AI employees" and emphasizing enterprise trust and product mindset. DeepSeek and Zhipu focus on model bases, competing for API calls and private deployment shares through open source, low prices, and long contexts. ByteDance’s position is in between—it aspires to build a closed Agent system while relying on low prices to enter the market. This "middle path" may face dual challenges: struggling to compete with Claude Code in developer mindset and facing pressure from DeepSeek in cost-effectiveness competition. As model iteration slows, computing power costs will become a decisive factor, and this close-quarters battle is far from over.

Sources

• Sohu: "Doubao Shifts to B2B: Can Its Low-Price AI Coding Strategy Break Through Fierce Competition?" (https://m.sohu.com/a/1044553734_362225)