Maifushi shows how AI apps sell finished workflows

Maifushi fits a broader Chinese AI pattern: the money is moving away from plain access and toward products that can sit inside a repeatable business workflow. 36kr reported that iFlytek's open platform revenue reached 3.705 billion yuan in the first half of 2026, up 36.01% year on year, while its revenue share rose to 31.87% and surpassed smart education as the company's largest revenue source. The same report said iFlytek's large-model API and MaaS platform service revenue grew about 70% year on year.

MiniMax shows the same pull from another direction.

ifanr said MiniMax's revenue structure shifted from about 70% consumer-side revenue and about 30% enterprise-side revenue one year earlier to business-side dominance. Its business-side revenue, mainly from open platform and other enterprise services, reached $73.9 million in the first half of 2026, up more than 700% year on year, and its share rose from 30.3% to 63.4%. The synthesis is practical: AI apps sell best when they own the work after the prompt.

From model access to owned workflows

The common move is from selling access to taking custody of work. MiniMax shows it from the model side. According to ifanr, its revenue mix moved from about 70% consumer-side revenue and about 30% enterprise-side revenue one year earlier to a structure dominated by business-side revenue. Business-side revenue, mainly from open platform and other enterprise services, reached $73.9 million in the first half of 2026, up more than 700% year on year. Its share rose from 30.3% to 63.4%.

That is not a small product tweak, because it changes what the buyer is paying for.

iFlytek shows the same pressure from a different starting point. According to 36kr, its open platform revenue reached 3.705 billion yuan in the first half of 2026, up 36.01% year on year. The open platform's revenue share rose to 31.87%, passing smart education to become iFlytek's largest revenue source. Its large-model API and MaaS platform service revenue grew about 70% year on year, which suggests buyers are paying for deployable capacity rather than a standalone AI story.

Meitu sits closer to the application layer. Woshipm reported that, as of June, Meitu's AI productivity application ARR reached 620 million yuan, up 47.8% year on year. Kaipai began as a teleprompter feature inside Meitu Beauty Camera, then expanded into internet-style templates, AI covers, smart mixed editing, one-click video generation, and a vertical Agent for AI marketing videos. The important shift is the packaging: the product absorbs more steps around producing marketing video, instead of leaving the user with a blank model interface.

Maifushi makes the commercial destination explicit. By 2025, according to 36kr, AI Agent had become Maifushi's core product, covering AI marketing, AI sales, AI customer service, and AI foreign trade. Its AI application business recorded operating revenue of 1.13 billion yuan in the first half of 2026, up 123.7% year on year. 36kr also reported the company's expected shift from providing software tools to delivering business results, with charging moving from subscriptions toward result-based charging. That is the convergence: model labs, SaaS vendors, and app makers are trying to own the workflow tightly enough that variable AI cost can be passed on as usage, tokens, or outcomes.

Token growth is not the same as profit

MiniMax shows the seductive side of token growth. According to ifanr, its full-platform token consumption in July reached 20 times the January level, while global enterprise customers and developers exceeded 2 million, more than 10 times the level at the end of 2025. Gross profit also moved sharply, rising 464.8% year on year to $20.81 million in the first half of 2026.

But the loss line explains why tokens are not a business model by themselves. MiniMax's net loss decreased from about $400 million to about $360 million in the first half of 2026. That is progress, yet it still means usage growth has to outrun inference cost, sales cost, and product support. The company's own target for M3.1 makes the point: it aims to cut inference cost to about one-third of M3's initial launch cost when M3.1 is released.

iFlytek gives the same lesson from the platform side.

36kr reported that iFlytek's revenue reached 11.623 billion yuan in the first half of 2026, up 6.52% year on year, while its net loss attributable to shareholders was 204 million yuan, narrowed by 14.68%. Its large-model API and MaaS platform service revenue grew about 70%, and the open platform gross margin rose by 3.63 percentage points to 20.21%. By the end of June 2026, the platform had more than 11.5 million AI developers, including more than 3.2 million large-model developers.

Those are valuable indicators: developers create distribution, MaaS revenue shows demand, and margin expansion suggests better unit economics. The warning is that enterprise AI value can disappear inside the customer's process. Woshipm's author argues commercialization is diluted by process dilution, cost dilution, and competitive dilution. If a step is 5% of a process and becomes 20% more efficient, the whole process improves by at most 1%. McKinsey's 2025 survey found 39% saw some EBIT impact, but most credited AI with less than 5% profit impact. Gartner's early 2026 summary said at least half of generative AI projects were abandoned after proof of concept by the end of 2025.

Vertical apps survive by owning the handoff

The pessimistic case is real. The woshipm author argues that model momentum is at its strongest level in several years, while the application market has turned extremely cold. Some investment institutions told the author they no longer look at AI applications, or that the track is dead. Satya Nadella's Agent-era warning gives that fear a cleaner form: if agents can operate across software, independent business applications may collapse.

But the evidence does not say every application is equally exposed.

The safer cases are the ones that own the handoff from intent to finished deliverable. Adobe's quarterly revenue exceeded 6.6 billion dollars in the second quarter of fiscal 2026. Grammarly, after transforming into an AI writing platform, had annual revenue last year above 700 million dollars. Harvey reached an 11 billion dollar valuation in 2026 by going deep into legal processes such as due diligence, contract review, and litigation. These products are not selling a chat box. They sit inside production loops where users need a draft, a review, a file, or a decision-ready work product.

Meitu shows the same pattern in consumer and creator work. In 2025, it posted positive net profit growth for the seventh consecutive year. Its 2026 first-half forecast expects adjusted net profit attributable to shareholders to rise 36% to 40% year on year. Core image and design product revenue rose 30.9% year on year to 1.8 billion yuan, with paid subscription users above 18.44 million. As of June, its AI productivity application ARR reached 620 million yuan, up 47.8% year on year.

Kaipai makes the lesson concrete. It grew out of Meitu Beauty Camera's teleprompter feature and now wraps teleprompter use, internet-style templates, AI covers, smart mixed editing, one-click video generation, and a vertical Agent for AI marketing videos. Meitu disclosed that the internet-style template feature has been used 122 million times cumulatively.

That is why "models beat apps" is too broad. As woshipm's author argues, defensibility shifts away from model capability itself and toward stable workflows, segmented-scenario experience, and vertical-field data fine-tuning.

Result-based pricing needs buyer math

Maifushi's numbers show why result-based pricing is attractive, but they do not remove the buyer's homework. According to 36kr, Maifushi recorded 1.96 billion yuan in operating revenue in the first half of 2026, while its AI application business produced 1.13 billion yuan. That business had 27269 paying users, and average monthly revenue per user rose 80.0% year on year.

The sell is no longer just software access.

36kr says Maifushi is moving from providing software tools toward delivering business results, with charging shifting from subscriptions toward result-based charging. For buyers, that changes the unit of analysis. The relevant price is not the seat fee. It is the cost per completed workflow that survives retries, review, integration work, and human correction.

Start with process share. Woshipm's author gives the blunt arithmetic: if one workflow step accounts for 5% of an overall process and its efficiency improves by 20%, the theoretical maximum improvement for the full process is 1%. That is the trap in paying for isolated AI output. A model-generated draft, ticket reply, or sales message may look cheap, while the surrounding handoff keeps most of the cost in place.

Then compare against the real alternative. Woshipm notes that in some regions a basic customer service worker earns only 2,000 to 3,000 yuan a month. If AI still requires frequent review, exception handling, and system cleanup, result-based pricing must beat that labor baseline after the failed attempts are counted.

The vendor's margin is a clue, not a guarantee. Maifushi's AI application business generated 864 million yuan in gross profit and had a 76.6% gross margin in the first half of 2026, according to 36kr. Excluding the agent all-in-one machine business, the margin was 84.9%. Buyers should ask which part of that margin comes from owning the workflow, and which part comes from shifting hidden work back to the customer.

How Chinese AI companies are moving from access or subscriptions toward workflow monetization

How to judge whether an AI product is worth paying for

• Use workflow AI when the vendor owns a repeatable business process, not just a model call. Favor products embedded in concrete workflows such as Maifushi's AI marketing, AI sales, AI customer service, and AI foreign trade scenarios. The sources report that Maifushi's AI application business reached 1.13 billion yuan in operating revenue in the first half of 2026, up 123.7% year on year, with a gross margin of 76.6%.
• Be cautious when AI improves only a small slice of a larger process. Do not overpay for tools that automate a narrow step unless that step controls enough of the overall outcome. One source gives the example that if a workflow step accounts for 5% of an overall process and its efficiency improves by 20%, the theoretical maximum improvement for the entire process is 1%.
• Use platform or API access when token consumption is growing and unit cost matters. For builders and enterprise teams, evaluate model platforms on consumption scale and inference-cost discipline. MiniMax disclosed that its full-platform token consumption in July was 20 times its January level, and that it aims to reduce M3.1's inference cost to about one-third of M3's initial launch cost when M3.1 is released.
• Avoid treating consumer subscription growth as proof of durable economics. Consumer AI apps need retention and inference-cost control, not just trial volume. RevenueCat's statistics on more than 115,000 subscription apps found that AI apps had a trial start rate of 8.5%, compared with 5.6% for ordinary apps, but by the 12th month AI apps had a monthly subscription retention rate of 6.1%, compared with 9.5% for ordinary apps.
• Consider vertical incumbents when AI is layered into an existing product habit. AI features appear more defensible when they reinforce an existing workflow and user base. Meitu's core image and design product revenue increased by 30.9% year on year to 1.8 billion yuan, its paid subscription users exceeded 18.44 million, and as of June its AI productivity application ARR reached 620 million yuan, up 47.8% year on year.

Contracts should protect workflow portability

When a vendor sells a finished workflow rather than a tool, the contract has to cover the work itself. Privacy terms should say which customer data can be used for training, which data is only processed for the task, and how logs are retained. Compliance terms should name the required jurisdiction and audit evidence. Localization should cover language, hosting region, and human support hours, because MiniMax's overseas markets contributed about 60% of its revenue in the first half of 2026, according to ifanr.

Portability is the buyer's escape hatch.

That means export rights need to include source files, generated assets, structured records, account history, and workflow metadata. Prompt portability should be explicit too: system prompts, task templates, evaluation rules, and approval steps should be downloadable in a usable format. If the vendor changes the underlying model, the customer should get notice, regression testing time, and the right to hold a stable version for contracted work.

The pricing schedule also needs tripwires. Woshipm's author points to a scale trap: when user revenue is below inference cost, more usage can deepen losses. RevenueCat's data, cited by woshipm, shows why vendors keep adjusting packaging: AI apps had an 8.5% trial start rate versus 5.6% for ordinary apps, but by the 12th month monthly subscription retention was 6.1% versus 9.5%. Annual subscription retention was 21.1% versus 30.7%.

Enterprise buyers should turn that instability into thresholds. The contract should define token caps, overage prices, result fees, minimum quality measures, and renegotiation points before rollout. Maifushi's shift from providing software tools toward delivering business results, with charging moving from subscriptions toward result-based charging, makes this unavoidable. If the product owns the workflow, the buyer must still own the exit path.

In your own stack, watch the margin math before the demo. Maifushi's AI application business reached a 76.6% gross margin in the first half of 2026, according to 36kr; without the agent all-in-one machine business, it was 84.9%. The new machine business was lower, at 54.9%. That spread is the clue: workflow software can absorb model cost better than bundled hardware when it owns the repeatable work.

Ask vendors to price the finished handoff, not the feature.

The next useful test is operating drag. Maifushi's sales expense ratio fell to 11.4%, and its administrative expense ratio fell to 4.6%, while research and development expenses were 325 million yuan. If a tool claims result-based charging, make it show the templates, integrations, review loops, and unit economics behind that claim.

For readers outside China

• Availability: The source material covers Chinese companies and Chinese-language reporting, but it does not give a practical availability guide for overseas users. MiniMax is described as having overseas markets contribute about 60% of revenue in the first half of 2026, and Meitu Xiuxiu is reported to have topped overall app rankings in 52 countries and regions. For iFlytek, Maifushi, Zhipu, and the specific enterprise services discussed, availability outside China is not disclosed in sources.
• Pricing: The sources give company revenue, ARR, gross margin, and loss figures, but not end-user price cards for the products discussed. They do say that Maifushi's charging model will evolve from subscriptions toward result-based charging, and that MiniMax's model consumption is tied to token usage. Specific subscription prices, API token prices, enterprise contract prices, and regional prices are not disclosed in sources.
• Closest Western equivalents: Anthropic, for model/API and agent infrastructure comparisons; Harvey, for vertical professional workflow AI in legal processes; HeyGen, for AI video-generation workflow comparisons; Grammarly, for an incumbent productivity app transformed into an AI writing platform; Adobe, for a creative-software incumbent embedding AI into existing workflows
• Data residency: The source material does not cover data residency, hosting regions, cross-border data transfer, enterprise compliance terms, or whether overseas customer data is processed inside or outside China. Treat those as open procurement questions rather than assuming equivalence with Western SaaS deployments.

Sources

• ifanr MiniMax 违背「祖训」，企业收入暴涨 700%，偷偷变成了「卖铲子」的 https://ifanr.com/1676874
• woshipm 模型越强，应用反而值钱了 https://woshipm.com/ai/6442894.html
• 36kr 业绩快报 | 迈富时2026年中报：经营现金流净流入达5亿元，AI原生布局驱动业绩倍增 https://36kr.com/p/3938112386858115
• woshipm 技术真，泡沫也真：AI商业化的三次稀释与规模陷阱 https://woshipm.com/ai/6432216.html
• woshipm AI 发展了 4 年，把应用发展没了？ https://woshipm.com/ai/6432360.html
• 36kr 科大讯飞中报里的AI商业化进度条 https://36kr.com/p/3957597360438656