Industry Landscape: From Model-Capability Competition to Harness-Completeness Competition

1. Introduction: Four Tracks and One Common Variable

1.1. Research Scope and Analytical Framework

图 1-1｜Harness 六层能力模型：竞争重心从模型能力到完整度

数据来源：基于本文分析绘制的示意图。

This chapter, based on systematic research across four content and tool tracks, presents a panoramic view of the industry landscape from the AI Harness perspective:

1. AI IDE: agent development environments centered on coding scenarios (Cursor, Claude Code, GitHub Copilot, Codex CLI, Trae, and others);
2. AI Agents Platforms: agent frameworks, low-code platforms, and vertical autonomous agents (LangGraph, OpenAI Agents SDK, Claude Agent SDK, Dify, Coze, Devin, Manus, and others);
3. AI Image and Content Generation: text-to-image, image editing, and design tools (Midjourney, 即梦, 可灵, Nano Banana, Leonardo, ComfyUI, and others);
4. AI Comic-Drama and Novels: long-horizon content production (即梦 AI comic drama, 可灵, Vidu, 白日梦; 阅文妙笔, 番茄, NovelAI, Sudowrite, and others).

The analytical framework follows the Six-Layer Capability Model from Chapter 3 of this whitepaper: instead of treating platforms as "tools" and listing their features, each platform is dissected as a Harness implementation that has already been deployed in production — seeing where it made genuine investment, where it left gaps, and who fills those gaps. All vendor operating data are annotated as self-reported or with a source grade; where measurement scopes conflict, they are presented side by side and marked [To be verified].

1.2. Core Judgment of the Whole Chapter

The four tracks differ enormously in product form, customer base, and business model, yet between 2025 and 2026 they exhibit the same structural shift: the center of gravity of competition is moving from "which model is plugged in" to "the completeness of the Harness six layers". This chapter will argue: this shift has independent evidence in every track; L4 (Memory & State Layer, reproducibility and asset persistence) and L5 (Evaluation & Observability Layer) are shortfalls common to the whole industry; and the Chinese and overseas markets will follow different convergence paths along three dimensions — compliance, deployment form, and open-source strategy.

2. AI IDE Track

2.1. Market Overview: Adoption Saturated, Trust Has Not Kept Pace

The AI IDE is the most fiercely contested track, and the one that best verifies that "model capability does not equal engineering usability". Three sets of authoritative survey data sketch out the basic shape of the market:

The implication of these numbers is: market education is complete, but trust-building has not yet begun. The simultaneous rise in adoption and distrust shows that the value of the tools is acknowledged, while their output has not yet attained engineering predictability. "66% think it is almost right but not entirely right" describes not a capability problem but a verifiability problem — precisely what the L5 layer of the Harness is meant to solve.

2.2. Representative Platform Comparison Matrix

The tone-setting evidence for this track: GitHub officially stated in 2026 that "the model provides the raw intelligence, while the harness determines how effectively that intelligence is applied" (the harness shapes how effectively that intelligence is applied), and published the controlled-comparison methodology of its agentic harness; the official Terminal-Bench also made clear that its leaderboard "ranks systems, not models". The dual confirmation by the official leaderboard and top vendors elevates "Harness completeness determines product gaps" from an industry intuition to a citable engineering conclusion (methodological details in Chapter 5 of this whitepaper).

2.3. Competitive Focus

The competitive focus of the AI IDE track has clearly migrated:

1. L1 Context Engineering is the first competitive dimension. A medium-sized repository holds hundreds of thousands of lines of code, and no context window is large enough to hold them all. The three-piece set of rules file hierarchy (loaded close by), semantic codebase indexing, and context compaction determines how complex a requirement a single task can carry.
2. L6 Governance is an entry requirement, not a bonus item. The higher the autonomy, the larger the blast radius of an incident — the Amazon Q prompt-injection incident of 2025-08-11 (malicious instructions attempting to induce the agent to delete AWS resources) was the first large-scale public exposure of the structural weakness that "external content being processed can directly become instructions". Filesystem isolation and network isolation are both indispensable, credentials are invisible by default, and destructive commands require hard-coded interception.
3. Positive evidence shows that governance and autonomy are positive-sum: Anthropic officially disclosed that its sandboxing cut internal permission prompts by 84% while improving safety — constraints make the speed of scaling possible.

2.4. Business Models: Three Billing Logics

The three logics are converging toward a "subscription + usage" hybrid. The migration of billing models means buyers need new management capabilities: organization-level consumption-rate observation and budget alerts. The GitHub Copilot case also shows a product-line consolidation trend — its agentic harness, as the single shared component of the Copilot SDK, simultaneously drives multiple experiences including CLI, App, and code review; the positioning of "the Harness as a platform asset" is becoming ever clearer.

3. AI Agents Platform Track

3.1. Market Overview: A Three-Tier Supply Structure

The supply of AI Agents platforms has a clear three-tier structure, and the three tiers follow different Harness-completeness logics:

On growth signals (all vendor-side or third-party figures, to be cited with caution): Devin's ARR, per third-party compilation, grew from USD 37 million (2025-05) to USD 492 million (2026-05), and the valuation figure rose from USD 10.2 billion to USD 26 billion — the source is third-party aggregation (low confidence), [To be verified], and this whitepaper does not use it as a basis for industry-size estimation; the operating milestones officially disclosed by Devin (processing 147 trillion tokens and driving 80 million virtual computers in the 8 months since launch) are likewise self-reported. The certain fact on the framework side is: core frameworks are generally open source and free (LangGraph, CrewAI, and AutoGen are all MIT), and revenue is shifting to the managed and observability layers (LangSmith free tier 5,000 traces/month, Developer tier USD 39/user/month).

3.2. Representative Platform Comparison Matrix

3.3. Competitive Focus and Business Models

The competitive focus of this track shows two directions:

1. Competition at the framework layer is about the "control plane". Once orchestration primitives (loops, sub-agents, tool calls) converged, differentiation shifted to state management (LangGraph's checkpointing), context governance (Claude Agent SDK's compaction), and observability (LangSmith). This is consistent with the rising weight of L3 and L4 in the six-layer model.
2. Competition at the platform and vertical tiers is about L6 and cost metering. The core question in enterprise procurement has shifted from "can we orchestrate" to "how do we manage permissions, how do we export audit logs, how do we cap costs". Devin bills by ACU (active work time), Manus by credits, Dify / Coze by runs and tokens — the three forms arrive at the same destination: translating unpredictable token consumption into budgetable units of workload, which is essentially the productization of L5 observability capability.

The structural fact of the business models: the framework layer does not make money; what makes money is the runtime, observability, and governance. This distribution corroborates the judgment of this whitepaper — value is moving from the model-access layer to the Harness layer.

Same-Day Increment (2026-09-12; corrected 2026-09-13): The OpenAI Agents API entered public beta on 2026-09-10 (grade A, per the OpenAI official Changelog; the 2026-09-12 snapshot had recorded it as 09-11 per media reports), putting the orchestration capability of the existing Agents SDK under management — together with AWS Bedrock AgentCore GA (2026-06) this forms two moves in the same direction, marking that the competitive focus of this track is shifting from "are the orchestration primitives good to use" to "is the managed runtime governable". For details, see the research library: 03-市场研究/02-AI-Agents组/21-openai-agents-api.md.

4. AI Image and Content Generation Track

4.1. Market Overview

The image track is the most commercially mature of the four. Take 美图公司 as an example (HKEX announcement figure, very high confidence): total revenue for H1 2026 was RMB 2.21 billion (+22.1% year on year), of which imaging and design product revenue was RMB 1.77 billion (+30.9%, 80% of the total); adjusted net profit attributable to the parent was RMB 650 million (+39.5%); MAU was 282 million, with over 18.44 million paid subscribers (+19.7%) and a subscription penetration rate of 6.5%; productivity-app MAU was 33 million (+43.5%), with 2.35 million paid subscribers and ARR of about RMB 620 million. 美图's observability design is especially noteworthy: it uses "the quarter-over-quarter growth in AI compute-point consumption" (over 46% in both Q1 and Q2) as a proxy indicator for value hit-rate — measuring value hit-rate by consumption depth is a rarely mature practice of L5 observability in the content track.

The structural change on the supply side is equally significant: Google is injecting top-tier image capability into a general-purpose API via Nano Banana (Gemini-family image model) (first generation 1K/2K single image USD 0.039, Pro series 1K/2K USD 0.134, compiled from screenshots of the official pricing page), while domestic 通义万相 is waging price competition with API pricing of RMB 0.2 per image (Alibaba Cloud official) — basic image generation is being rapidly commoditized, and profit is migrating to workflows, asset management, and industry solutions (the Harness layer).

4.2. Representative Platform Comparison Matrix

4.3. Competitive Focus and Business Models

Two structural observations on this track:

1. The watershed of the degree of Harness-ification is at L3 and L4. 美图设计室 (multi-agent collaboration + creative asset reuse + compute-point metering), ComfyUI (workflows as JSON graphs, version-controllable, SDK-able), and Leonardo (explicit capacity constraints + Webhook + cost estimation) represent the third-generation form; Midjourney and 妙鸭 represent "strong model, weak Harness" — leading model quality but thin engineering carrying capacity, and 妙鸭 has already exited for this reason (see Section 4.4).
2. Business models move from one-time payment to a subscription + consumption hybrid. Midjourney is pure subscription, Runway pure credits, Leonardo dual-track subscription + PAYG, 美图 subscription + compute points. Consumption-based pricing presupposes precise usage metering and alerting — another instance of an L5 capability becoming commercial infrastructure.

4.4. A Typical Failure Case: 妙鸭相机

妙鸭相机 is the only negative sample in this track with a complete lifecycle, and a textbook case of "strong model, weak Harness":

• Explosion: launched 2023-07-17, topped the App Store overall chart in August 2023, daily actives broke 600,000, with over 4,000 users queued at peak to generate photos; the technical base was the "提香" model self-developed by 阿里大文娱.
• Engineering shortfall: the L2 / L3 / L5 layers were almost blank — every interaction was an atomic operation (upload, wait, pick a template, produce the photo), with no intermediate state to intervene in and no parameters to adjust; the user's only quality-correction entry point was a "make it more like me" similarity fine-tune. L6 saw an early user-agreement trust incident (the official apology and revision).
• Commercial shortfall: the digital avatar was a one-time RMB 9.9 (promotional price) buyout + 10 finished photos — typical traffic-driver pricing, lacking follow-on consumption scenarios and a subscription anchor; compared with similar identity products (可灵 caps subjects at 30 to 500 by membership tier) it lacked asset-operations depth.
• Tech-route lock-in: the training-based identity pipeline (≥20 photos per user, full per-user training, hours of queuing) did not complete its migration after zero-shot identity-injection solutions (InstantID, PuLID) matured in 2024–2025, and the cost structure was locked in. From the Harness perspective: the training-based route made identity assets into "private state inside the product" — non-migratable, non-composable, non-auditable; the zero-shot route made identity assets into "context that can circulate across pipelines", gaining real asset liquidity.
• Ending: the team formally disbanded at the end of September 2025 (the authoritative source for this fact is, to date, only the statement of a third-party tool site), and the product continues only at a minimal level of operation.

The failure proposition of 妙鸭 is worth writing into the project review of every content product: given that generation quality is already good enough, what determines a product's life or death is not the model but the Harness that carries it.

5. AI Comic-Drama and Novel Track

5.1. AI Comic-Drama: Overcapacity and Scarce Certainty

AI comic drama is the scene in the AIGC content industry where engineering has advanced fastest and where the contradiction that "model capability does not equal production usability" was exposed earliest. Industry data for H1 2026 (DataEye report and 中国网络视听协会 data, via secondary relay; the full original report was not cross-verified):

Capacity is piling up at a rate of one title every 36 seconds, the hit rate is under 0.5%, and unit prices have dropped by 90% — the industry has moved from "racing for capacity" to "racing for certainty". And certainty is precisely what models cannot provide: the same character looks different in episode 1 and episode 47, the same scene has lighting jump between two shots, and plot state is lost when writing continues across sessions. Not one of these problems can be solved by swapping in a stronger model; they can only be solved by the Harness's state management (L4), orchestration (L3), and regression evaluation (L5).

5.2. AI Comic-Drama Representative Platform Comparison Matrix

The layering rule of this track is clear: L1 through L3 (four-modality input, intelligent storyboard, multi-shot narrative, node workflows) have by 2026 become standard equipment on leading platforms, highly homogenized; L4 has not yet homogenized — from "up to 7 reference images" (Vidu) to "about 100k tokens compressed to about 4k tokens" (海螺 H3) to "workflow JSON is the state" (ComfyUI), the implementation paths are entirely different. Whoever builds L4 solidly (character anchoring, asset library, shot state machine) will be able to push AI comic drama from "can generate" to "can continuously produce a hundred episodes".

5.3. AI Novels: The Long-Horizon State-Management Race

AI novels are the content scene where L4 and L1 bear the greatest pressure in the six-layer model: the core difficulty of long serialized fiction is "not collapsing character settings or forgetting foreshadowing after hundreds of thousands of words" (the web-novel industry names this guaranteed-to-occur fault "eating the book"). The industry has converged on three L4 paradigms:

The platform landscape shows "Chinese and overseas platforms solving the same problem differently": 阅文妙笔 enters with "妙笔通鉴" in a "second brain" positioning (specializing in foreshadowing mining and detail retrieval); its disclosed business indicators are Miaobi daily actives more than doubling, daily-average token consumption up over 90%, and weekly author usage over 75% (vendor self-reported); 番茄小说 is known for the strongest L6 (mandatory AI declaration, expansion and continuation writing banned for guaranteed-minimum books, machine judgment of low-quality content); NovelAI stands in the subscription market on Lorebook and privacy-first (XSalsa20 client-side encryption, no training); Sudowrite supports English long-form with Story Bible + reading the first 20,000 words + 25-document chapter continuity. On the community side, the GitHub open-source project Claude-Code-Novel-Writer (v4.1, 2026-08-21) provides, through multi-agent orchestration of AGENTS.md + 7 roles + 5 skills, an engineering model of "authoritative source / derived index separation" (the manuscript is the authoritative source; state tracking is regenerable).

5.4. Governance Is a Prerequisite, Not an After-the-Fact Check

L6 in both content tracks has already shifted from a "bonus item" to a "precondition for launch":

• Regulatory baseline: the 《人工智能生成合成内容标识办法》 (issued 2025-03-07, in effect 2025-09-01) requires videos to carry a prominent explicit label on the opening frame and an implicit label written into the file metadata; the mandatory national standard GB 45438—2025 quantifies it further: explicit label text height no less than 5% of the shortest edge of the frame, lasting no less than 2 seconds at normal playback speed, with metadata fields including Label / ContentProducer / ProduceID, etc.
• Enforcement instance: on 2026-04-28, 即梦 AI was lawfully investigated by the cyberspace authority for failing to effectively implement the labeling rules — even for a top-tier big-company platform, the labeling pipeline can break.
• Platform governance divergence: from 2025-09-23, 番茄 mandates authors to declare "whether AI is used"; AI continuation and expansion writing are limited to authors of non-guaranteed-minimum books (authorization tiered by contract type); 晋江文学城 permits only three AI use scenarios: text proofreading, creative-element assistance, and creative rough-outline assistance. Overseas platforms govern mainly around copyright ownership (whether work rights are claimed, whether user works are used for training) and privacy (encryption, no-training commitments).

The direct implication for engineering teams: compliance labeling must be built as an automatic node at the tail of the generation pipeline, a first-class attribute of the artifact alongside resolution and frame rate, and cannot rely on manual post-hoc addition.

6. Cross-Track Common Judgments

6.1. Competition Is Shifting from Model Capability to Harness Completeness

Aggregation of independent evidence across the four tracks:

The four tracks converge on the same conclusion in entirely different ways: once model capability crosses the "usable" threshold, the perceptible gap between products is determined mainly by the completeness of the Harness six layers. This and the evaluation findings of Chapter 5 of this whitepaper are two sides of the same coin — since an evaluation score is itself a composite product of "model × Harness", market competition naturally unfolds in the same composite way.

6.2. L4 and L5 Are Industry-Wide Shared Shortfalls

Cross-aggregating the six-layer maturity of the four tracks yields a highly consistent conclusion:

The common root cause of the L4 shortfall is: the model itself is stateless, and every "consistency" problem is essentially a cross-call, cross-session, cross-cycle state-injection and persistence problem — character consistency (comic drama), setting consistency (novels), migratable identity assets (image), and recoverable sessions (IDE) are projections of the same kind of engineering problem onto different tracks. The common root cause of the L5 shortfall is: the output of content and open-domain tasks lacks an objective yardstick for judgment, and public benchmarks are absent. Therefore, L4 and L5 are both an industry-wide shortfall and, in the coming two years, the clearest differentiation window — whoever first turns "reproducibility and asset persistence" (L4) and "effect evaluation" (L5) into product capabilities rather than marketing claims will win, in their own track, a differentiated position like Vidu's subject library, 海螺's Context-IR, or 美图's compute-point observability.

7. Differences Between the Chinese and Overseas Markets

7.1. Ecosystem Structure

• Overseas: the main axis is "model vendors + subscription + API economy". The allocation logic is product strength and global market coverage (PixVerse covers 175 countries; 可灵's roughly 70% overseas revenue share is its growth engine), with pricing mainly USD subscriptions and usage billing.
• China: the main axis is "big-company channel closed loop + free strategy + revenue-sharing ecosystem". 即梦 connects to 抖音, 可灵 connects to 快手, with model + tools + distribution integrated; some domestic tools enter on a fully free strategy (e.g., 字节 Trae, and 腾讯元宝's public statement that it is "currently fully free to use, with no plans to charge yet"); on the content side a distinctive revenue-sharing system has formed (comic drama shared by "duration × unit price × type coefficient × copyright coefficient") along with guaranteed-minimum incentives (the 2026 full-year guaranteed-minimum budget for live-action short dramas at the 抖音 group exceeds RMB 1.5 billion).

7.2. Compliance Environment

• China: compliance is an explicit hard constraint with enforcement instances already on record — the 《人工智能生成合成内容标识办法》 and GB 45438—2025 form a quantified baseline, and 即梦 being investigated on 2026-04-28 is the landmark event; content platforms additionally face filing thresholds (in January 2026 the filing threshold for key micro-short-dramas was raised from RMB 1 million to RMB 3 million) and subject-matter review.
• Overseas: the transparency obligations of the EU AI Act apply from 2026-08-02 to labeling of AI-generated text directed at the public (an open-source license does not constitute an exemption); the United States is driven mainly by platform self-regulation and litigation. Overall, China follows a "pre-labeling + platform verification" model, the EU a "transparency obligation" model, and the United States an "after-the-fact accountability" model.

7.3. Deployment Form

• China: Xinchuang (信创) and domestic-chip adaptation are real demands — on the first day of MiniMax H3's open-source release, adaptation was completed for 华为昇腾, 摩尔线程, 沐曦, 海光, 昆仑芯, 天数智芯, 壁仞, and others; private deployment carries high weight in enterprise procurement (Dify's private-deployment capability is its core selling point).
• Overseas: cloud first, SaaS first; local deployment appears mainly in regulated industries and the open-source community (ComfyUI's fully local offline capability lets it satisfy both kinds of need at once).

7.4. Open-Source Weight Strategies

• Chinese vendors are more aggressive about open-sourcing model weights: MiniMax H3 (open weights for a video model), 通义 Qwen family (coding and image), 智谱 GLM family (NovelAI's Xialong is fine-tuned from GLM-4.6) — open weights have become a lever for acquiring the developer ecosystem and domestic-chip adaptation.
• Overseas top vendors handle it in tiers: OpenAI open-sourced Codex CLI under Apache 2.0 (CLI layer open source, model closed source); the framework layer is generally MIT (LangGraph, CrewAI, AutoGen); frontier model weights are essentially closed source. Overall, China shows an "open weights in exchange for ecosystem" strategy, and the overseas shows an "open-source tools and frameworks in exchange for standards, closed weights to protect profit" strategy.

7.5. Overview of Differences

It must be emphasized: the differences are differences of degree, not of essence. The two markets are fully aligned on the two structural judgments that "the center of gravity of competition is shifting from model capability to Harness completeness" (Section 6.1) and "L4 / L5 are shared shortfalls" (Section 6.2) — the differences show up in the speed of convergence and the constraints, not in the direction.

8. Three Possible Paths of Landscape Evolution

8.1. Path One: Platform Convergence

Big companies, leveraging their integrated advantage in models, channels, and compute, productize all six Harness layers, and small and medium players degrade into model suppliers or vertical plugins. Supporting signals: GitHub unifies multiple product lines with a single agentic harness; the "model + tools + distribution" closed loops of 即梦 / 可灵; 美图 integrating multiple product lines with a platform-level compute-point system. Constraint: L4 and L5 are precisely the layers hardest for big companies to go deep on — asset persistence and effect evaluation require long-term customer companionship, not traffic tactics, and vertical players still have a chance to hold their incumbent advantages.

8.2. Path Two: Vertical Specialization

Industry knowledge and the state layer become the moat, and general-purpose platforms are hard to replace. Supporting signals: the continuous-production capability of AI comic drama depends on the industry-specific asset library and storyboard state machine; the "project constitution" and foreshadowing ledger of web novels depend on domain semantics; in enterprise Agents procurement the weight of L6 (permissions, audit, privatization) keeps rising. 妙鸭's failure supports this path from the negative side — a "general-purpose hit product" without a state layer and operational depth has an extremely short lifecycle. Constraint: the vertical market size is limited, making it hard to amortize foundation-model R&D costs; in the end it mostly depends on big-company model APIs, and bargaining power is constrained by others.

8.3. Path Three: Open-Source Erosion

Open-source workflows and open-source weights squeeze closed-source SaaS from both ends: ComfyUI covers professional production with a 60,000+ node ecosystem; Codex CLI proves that closed-source vendors will proactively open-source the Harness layer to contest standards; Chinese vendors' open-weight strategy keeps lowering the bar for self-building. If open protocols such as MCP and open standards such as AGENTS.md keep penetrating (AGENTS.md has been adopted by 60,000+ open-source projects, hosted by the Agentic AI Foundation under the Linux Foundation), the six-layer capability will increasingly exist in the form of "assemblable open-source components", and closed-source platforms will be forced to narrow their profit surface toward "managed services and governance" (L5 / L6). Constraint: open-source solutions lack evaluation and governance by default, and enterprise self-building is costly — this in turn gives commercial products a space for survival built around L5 / L6.

8.4. Observation Indicators

The three paths are not mutually exclusive; the final landscape is more likely a mixture of all three. Judgments on where the landscape is heading can rest on the following observable indicators:

9. Conclusion

Based on systematic research across the four tracks, this chapter draws three conclusions:

1. Competition is shifting from "model capability" to "Harness completeness". The AI IDE track is doubly confirmed by the official leaderboard and top vendors ("ranks systems, not models"; the harness determines the efficiency of intelligence application); in the Agents track frameworks are free and revenue flows to managed services and governance; in the image track basic generation is rapidly commoditized and profit flows to workflows and asset management; in both content tracks there is overcapacity and scarce certainty. The four tracks converge on the same structural change in different ways, and its mirror on the evaluation side is the finding of Chapter 5 of this whitepaper — a score has always been a composite product of "model × Harness".
2. L4 (Memory & State) and L5 (Evaluation & Observability) are industry-wide shared shortfalls. The model is stateless, and every "consistency" is a state-management problem; content and open-domain tasks lack an objective yardstick, and public benchmarks are absent in multiple high-value directions. The complete-lifecycle failure of 妙鸭相机 (strong model, weak Harness, no state-asset liquidity, no evaluation loop) and the 0.47% hit rate of AI comic drama mark the commercial cost of these two layers' shortfalls from both the positive and negative sides; cases such as Vidu's subject library, 海螺's Context-IR, and 美图's compute-point observability mark the differentiated return of making up the shortfalls.
3. China and overseas markets will take different shapes in the mixed evolution of the three paths: China is characterized by hard compliance constraints, channel closed loops, and an open-weight strategy; the overseas by a subscription and API economy, transparency obligations, and tool open-sourcing. Whatever the shape, the ultimate decisive variable of the landscape is the same — whoever can convert the model's uncertainty into engineering predictability holds the pricing power of the next competitive cycle.

Statement of Information Gaps

1. Devin's ARR and valuation data (USD 37 million to USD 492 million; USD 10.2 billion to USD 26 billion) come from third-party aggregation (low confidence); all are marked [To be verified] and were not used in any industry-size estimation.
2. The authoritative source for the disbanding of the 妙鸭相机 team (end of 2025-09) is, to date, only the statement of a third-party tool site; no tech-media coverage or business-registration corroboration has been located, [To be verified].
3. The AI comic-drama market size (RMB 40 billion / USD 4 billion overseas) and hit-rate data come mainly from the DataEye report via secondary relay (百度百科, KOCPC, and others); the full original report was not obtained for cross-verification.
4. Consumer pricing across platforms has multiple conflicting figures (即梦, 可灵 domestic site, 美图设计室, Coze enterprise edition, Vidu, and others); this whitepaper accepts only verifiable official-page figures, and the rest are marked [To be verified] or not cited.
5. The release date of 可灵 3.0 and "up to 3 minutes", and 白日梦's "95%+ cross-episode similarity", etc., appear only in secondary sources, [To be verified].
6. The 阅文妙笔 business indicators (daily actives +100%, token +90%, weekly usage over 75%) and the 美图设计室 Agent Teams results are both vendor self-reported figures.
7. Most platforms in the AI novel and image tracks have no officially disclosed verifiable effectiveness indicators; third-party review data are mostly marketing pieces and were not cited.
8. This whitepaper does not adopt any self-media data on big domestic companies' AI coding efficiency as hard data (no primary source for this kind of data was found in this project's search).

10. References

1. 2025 Stack Overflow Developer Survey — Stack Overflow, 2025-07-30. https://survey.stackoverflow.co/2025/
2. State of AI-assisted Software Development 2025 — DORA / Google Cloud, 2025-11-12. https://dora.dev/research/2025/dora-report/
3. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (randomized controlled trial) — METR, 2025-07-10. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
4. Evaluating performance and efficiency of the GitHub Copilot agentic harness across models and tasks — GitHub Blog, 2026. https://github.blog/ai-and-ml/github-copilot/evaluating-performance-and-efficiency-of-the-github-copilot-agentic-harness-across-models-and-tasks/
5. Terminal-Bench official site and Leaderboard — Stanford / Laude Institute, from 2025 to 2026. https://www.tbench.ai/
6. AGENTS.md official site — Agentic AI Foundation (under the Linux Foundation). https://agents.md/
7. Sandboxing: a safer and more autonomous approach — Anthropic, 2025. https://www.anthropic.com/engineering/claude-code-sandboxing
8. MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities — MiniMax official blog. https://www.minimax.io/blog/minimax-h3
9. 美图公司 2026 interim results announcement — 香港交易所披露易 (hkexnews). https://www.hkexnews.hk/
10. 国家网信办等四部门《人工智能生成合成内容标识办法》（国信办通字〔2025〕2 号） — issued 2025-03-07, in effect 2025-09-01. https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
11. Mandatory national standard GB 45438—2025《网络安全技术 人工智能生成合成内容标识方法》 — 全国网络安全标准化技术委员会. https://www.tc260.org.cn/upload/2025-03-15/1742009439794081593.pdf
12. DataEye《2026 上半年 AI 剧 / 漫剧数据报告》 — DataEye研究院 (reprinted via secondary sources such as KOCPC). https://en.kocpc.com.tw/archives/25203
13. 澎湃新闻《1914 元制作 1 集？漫剧仍困在隐性成本中》 — 澎湃新闻, 2026. https://www.thepaper.cn/newsDetail_forward_33993493
14. 中国青年报《"妙笔通鉴""漫剧助手"发布，AI 赋能网文创作和 IP 改编》 — 中国青年报, 2025. https://new.qq.com/rain/a/20251017A08J6J00
15. GitHub — forsonny/Claude-Code-Novel-Writer (Multi-Agent Novel Writer v4.1) — 2026. https://github.com/forsonny/Claude-Code-Novel-Writer
16. Apache Software Foundation《AI Agents and AGENTS.md Files (Draft)》 — ASF legal-discuss, 2026. https://www.apache.org/legal/generative-tooling-agents.html
17. 通义万相 and Qwen-family models: 阿里云 official pricing page — 阿里云. https://help.aliyun.com/
18. Leonardo.ai official pricing and API documentation — Leonardo. https://leonardo.ai/pricing
19. Runway official pricing page — Runway. https://runwayml.com/pricing
20. Comfy official site — Comfy Org. https://comfy.org/