AI IDE Platform Market Research (Group Overview)

1. Introduction

AI IDE is the most fiercely contested arena in this round of AI engineering, and the one that best verifies that "model capability is not the same as engineering usability." When GitHub Copilot appeared in its "autocomplete" form in 2021, the industry debated whether AI could write code; by 2026 the debate had become whether AI-written code can be trusted, audited, regression-tested, and held accountable. This shift is precisely the direct reason that AI Harness (the AI control layer / agent runtime framework) emerged as its own independent layer.

By the unified definition in this project's parameter card, AI Harness sits between the foundation model and the real business system, converting the model's uncertainty into engineering-level predictability. AI IDE is the vertically integrated form of Harness in the coding scenario: it packages six concerns — context engineering, tool execution, task orchestration, state memory, evaluation & observation, and permission governance — into the interface that developers open every day.

The key judgment is this: the gap between AI IDEs is no longer mainly about "which model is plugged in," but about the completeness of the six Harness layers. The same model and the same API key can differ by double-digit percentage points across different harnesses — this is not marketing, but Terminal-Bench's own definition of itself: "Terminal-Bench does not score a model in isolation. Every result is a model plus an agent harness... the leaderboard is a ranking of systems, not of models."

1.1. Research Scope and Subjects

This group selected 20 representative platforms, covering six organizational forms:

Batch 1 finalized six documents — Cursor, Claude Code, GitHub Copilot, Windsurf, Codex CLI, and Trae; batch 2 finalized seven — CodeBuddy, Zed, Cline, Roo Code, Tongyi Lingma, Kiro, and Gemini CLI; batch 3 (v1.1 increment) finalized seven — Qoder, WorkBuddy, DeepSeek, Antigravity, OpenCode, Junie, and Augment Code. All 20 platforms in this group have their own dedicated document; the complete list is in Section 7 Document Navigation, and this README's comparison matrix stays in sync with the body of each volume.

Selection criteria: these 20 subjects cover every significant tier difference in this group's Harness six-layer comparison matrix, and each offers enough publicly verifiable information to support a standalone document.

1.2. Position of AI IDE within the AI Harness System

AI IDE is a high-frequency, short-cycle, strong-feedback sample for testing the six-layer Harness model. Compared with enterprise-grade agent platforms, it has four characteristics:

1. Extremely short feedback loop: code has compilers, tests, and CI, so a single change can be judged right or wrong within minutes. AI IDE is therefore the scenario in which L5 (evaluation & observation) is easiest to make real among the six layers — and currently the only agent scenario that has deployed "programmatic scoring" at scale.
2. Direct side effects and high irreversible risk: an agent that writes a wrong line of code could drop a database, leak a secret key, or commit credentials into a public repository. L6 (governance & security) is therefore not a bonus but an admission requirement in AI IDE.
3. Context is naturally scarce: a mid-sized repository holds hundreds of thousands of lines of code, which no context window, however large, can fit. This makes L1 (context engineering) the first competitive dimension of AI IDE, concretely expressed as the trio of rules files, codebase indexing, and context compaction.
4. The developer is the acceptance tester: users have the professional ability to judge the output, which both lowers the barrier to building L5 and makes cognitive-bias problems such as "divergence between self-assessment and measurement" easier to expose and quantify.

This group's documents therefore do not evaluate platforms as "tools"; instead, each platform is dissected as a set of already-deployed Harness implementations: examining where it made real investment in a layer, where it left a gap, and who fills that gap.

1.3. Dissection Method

图 1-1｜AI IDE 定位与 Harness 六层架构

数据来源：基于本文分析绘制的示意图。

The same dissection framework is applied uniformly to all platforms to ensure cross-platform comparability:

2. Glossary

This table is the common terminology baseline for the entire group; each individual platform document's glossary section builds on this baseline by adding platform-specific concepts.

3. Market Overview

3.1. Adoption Is Saturated; Trust Has Not Caught Up

Three sets of first-hand or authoritative survey data sketch out the basic shape of the current market:

The meaning of this data is: market education is complete, but trust-building has not yet begun. While adoption rose from 76% to 84%, distrust rose from 31% to 46% — both rising in tandem, showing that AI IDE's value is acknowledged, but its output has not yet gained engineering-level predictability. The Stack Overflow finding that "66% say it's almost right but not quite" is especially critical: it describes not a capability problem but a verifiability problem. This is exactly what Harness's L5 layer must solve.

3.2. The Divergence Between Self-Assessment and Measurement

The randomized controlled trial published by METR on 2025-07-10 provides this group's single most important counter-intuitive piece of evidence: senior open-source developers actually completed tasks 19% slower after using AI tools, even though those developers self-assessed that they were about 20% faster. The roughly 39-percentage-point gap between the two figures is a baseline that must be cited repeatedly across all of this group's documents.

Plausible mechanisms behind the divergence include: developers underestimate verification and rework costs; the tacit knowledge of senior developers who are familiar with the repository cannot be reused by the AI; and AI output needs a full review pass before it can be merged, a cost that self-assessment ignores. Whatever the mechanism, the conclusion is consistent: "feels faster" cannot be the acceptance criterion for an AI IDE. This also explains why DORA 2025 concluded that "AI adoption correlates positively with delivery throughput but negatively with delivery stability" — teams can ship more changes, but not necessarily safer ones.

This group therefore recommends following one hard principle in selection and acceptance: any efficiency conclusion must have a controlled measurement; subjective self-assessment is not accepted.

3.3. Capability Benchmarks: Score Differences for the Same Model, Different Harnesses

The following scores from public benchmarks can be cited directly, each with its evaluation date and configuration:

Note: SWE-bench Verified was officially deprecated by OpenAI on 2026-02-23 (because ≥59.4% of defect tests reproduce a gold patch across all frontier models); the scores cited are the historical figures before deprecation. See 02-行业赋能/03-软件工程组/06-benchmark.md for details.

It must be emphasized: these scores are for the "model + Harness" system, not the model's bare scores. Terminal-Bench's official statement is explicit about this, and public reports also show the same model differing by several percentage points across scaffolds (such figures are lower-quality sources and can only be cited after being marked [To be verified]). Therefore, when using benchmark scores as a basis for platform selection, you must ask three things: which Harness was used, what sandbox configuration, and whether retries were allowed.

3.4. Security Incidents and the Risk Surface

The higher an AI IDE's autonomy, the larger the blast radius of security incidents. Publicly disclosed landmark incidents:

• Amazon Q incident (2025-08-11): a malicious prompt injection attempted to induce the agent to delete AWS resources. This was the first large-scale public exposure of the structural weakness whereby "processed external content can directly become instructions."
• Agentic browser risk surface: success rates on e-commerce checkout-style tasks have reached 87% (a product category that formed after Operator's launch on 2025-01-23 and Atlas's launch on 2025-10-21), but when an agent can operate a real browser holding real sessions and payment context, the consequence of prompt injection escalates from "mistaken code edits" to "financial loss."

Three engineering requirements follow from this: file-system isolation and network isolation are both indispensable (with network isolation alone, a compromised agent can exfiltrate SSH keys; with file-system isolation alone, a compromised agent can easily escape and gain network access); credentials must be invisible by default (paths such as environment variables and `~/.aws/credentials`, `~/.ssh` should be explicitly deniable or masked); destructive commands must be hard-coded to be blocked, without relying on model restraint.

The positive evidence to reference is Anthropic's officially disclosed sandbox results: in internal use, sandboxing reduced permission prompts by 84% while improving security. This proves that governance and autonomy are not a trade-off but a positive-sum relationship — constraint is what makes speed possible.

3.5. Commercialization Models: Three Billing Logics

The three logics are converging toward a "subscription + usage" hybrid. For buyers, the real questions are not the unit price but three others: does overage stop the service or bill afterward? Does code completion count against the quota? Is there an organization-level consolidated billing and audit export?

4. Cross-Platform Comparison Matrix

4.1. Basic Attributes Comparison

Note: prices are adjusted frequently by vendors; the table above lists the public prices verifiable at the time of writing. Items sourced from third-party aggregations are marked [To be verified], and the pricing details for each platform follow the corresponding volume. Devin (Cognition) is a cloud-based autonomous software engineer rather than an AI IDE, so it is outside this group's research scope; see 03-市场研究/02-AI-Agents组/10-devin.md for details.

4.2. Harness Six-Layer Maturity Comparison

Rating legend: ★★★ has a clear, configurable, auditable mechanism; ★★ has a mechanism but it is coarse-grained or not auditable; ★ relies only on model restraint or requires the user to build it themselves.

The ratings for rows 07–20 have been backfilled from the six-layer capability summaries (sections 5.7 / 5.8) in each volume's "Harness Design" chapter; see the corresponding volume for the basis of each layer judgment. When a volume revises its ratings, this table must be updated in sync. Platforms 14–20 (Qoder, WorkBuddy, DeepSeek, Antigravity, OpenCode, Junie, Augment Code) were written by supplementing and verifying against the R18 retrieval report and official pages, with information current as of 2026-09-12.

5. Selection Decision Tree

The first step of selection is not comparing features, but answering three prerequisite questions.

Question 1: Can the data leave the local machine?

• No (regulated industries, sensitive source code) → prioritize solutions with an organization-level privacy mode, zero data retention, content exclusion, and audit export: Cursor Enterprise (team-level privacy mode, codebase and model access control, audit logs), GitHub Copilot Business / Enterprise (organization policies, content exclusion, audit logs, IP indemnity), Windsurf (SSO / RBAC and compliance certifications). Open-source options (Codex CLI, Zed, Cline, Roo Code, Gemini CLI) can achieve fully local execution, but the governance layer must be built by yourself.
• Yes → move on to Question 2.

Question 2: Which interface hosts your main workflow?

• Terminal-first, dominated by long-running tasks and batch jobs → Claude Code or Codex CLI. Both bill by subscription quota with predictable cost, suited to multi-hour tasks; if you need open-source audibility, choose Codex CLI (Apache 2.0).
• Editor-first, needing item-by-item review and visual diffs → Cursor (the strongest Tab and inline-diff experience) or Windsurf (plan mode + parallel sessions, suited to delegating whole sections of work).
• You already do all your R&D collaboration on GitHub → GitHub Copilot, which distributes through the platform and keeps governance and audit closest to existing workflows.
• Budget-sensitive, or needing browser-based development / domestic model ecosystems → Trae (international version starts at $10; there is also a domestic version).

Question 3: How much autonomy do you need?

One general recommendation: no matter which vendor you choose, first run in read-only mode against one or two real repositories for a week, distilling "the paths to forbid, the credentials to deny, the linters to enforce" into in-repository rules files before opening up autonomy. Teams that skip this step often find themselves coming back to do the same thing after their first incident.

6. Core Assertions of This Group

Assertion 1: Competition in AI IDE has shifted from model integration to Harness completeness. The score differences of the same model across scaffolds have been repeatedly observed; Terminal-Bench officially states that its leaderboard ranks systems, not models. When selecting, you should therefore prioritize rules hierarchy, sandbox granularity, approval policy, and trace observability rather than the length of the model list.

Assertion 2: L1 and L6 are the two layers that determine the ceiling and the floor. L1 (context engineering) determines how complex a requirement a single task can carry — rules hierarchy, indexing quality, and compaction strategy together form the ceiling. L6 (governance & security) determines how high you dare to set autonomy — sandbox, approvals, and audit together form the floor. The four layers in between (tools, orchestration, memory, evaluation) determine efficiency, not feasibility.

Assertion 3: Self-assessment cannot serve as an acceptance basis. The METR randomized controlled trial showed that senior developers are actually 19% slower after using AI, while self-assessing as 20% faster. Any team introducing an AI IDE should set up controlled measurement: select a set of repeatable tasks and record completion time, rework rate, and defect rate before and after introduction, rather than distributing questionnaires.

Assertion 4: Governance and autonomy are positive-sum, not a trade-off. Anthropic's sandbox reduced permission prompts by 84% while being safer; after OpenAI mechanically enforced architectural constraints with linters and structural tests, growing team size actually increased throughput. Constraint is not the enemy of speed, but the precondition that makes speed at scale possible.

Assertion 5: The cost model is shifting from "seats" to "consumption," and buyers need new management capabilities. GitHub switched Copilot to AI Credits billing from 2026-06-01 (1 credit = $0.01); Cursor uses a credit pool + usage-based billing; Windsurf switched to daily / weekly quota refreshes in 2026-03. This means "how much one person spends per month" is no longer a fixed number — organizations need consumption-rate observation and budget alerts.

7. Document Navigation

Suggested reading order: first read this README to establish the comparison framework, then read each volume in the order of the table above; if you care only about governance and security, read 02 and 03 first; if you care only about context engineering, read 01 and 06 first.

8. Summary

By 2026, the AI IDE market has completed the shift from "can it be used" to "do we dare to delegate authority." The 84% adoption rate shows the tool itself has been accepted; the 46% distrust rate, the 66% "almost right but not quite," and the −19% measured efficiency in the METR trial together show that the engineering-bearing layer has not caught up.

This is the value of dissecting AI IDE as a Harness: it replaces the unactionable conclusion that "the model isn't smart enough" with engineering items that can be fixed one by one — "L1 rules hierarchy not built," "L2 sandbox not enabled," "L5 lacks a mechanical criterion," "L6 lacks audit export."

The common conclusion across this group's twenty volumes is: the decisive factor in selection is not the model list, but six-layer completeness and the team's own engineering discipline. The same tool is a completely different product in a team that has set up rules files, linters, regression sets, and approval policies versus a team that goes straight to full-auto mode.

Information Gap Statement

1. The R08 dedicated retrieval report was not written up: this README is based on the verified hard data issued by the principal, the A-grade first-hand materials in the R01 overview retrieval report, and direct checks of official vendor pages. Any pricing and version numbers sourced from third-party aggregations are marked [To be verified].
2. Platforms 07–13 (CodeBuddy, Zed, Cline, Roo Code, Tongyi Lingma, Kiro, Gemini CLI): their basic attributes and six-layer ratings have been backfilled into the Section 4.1 and 4.2 matrices from each volume's main text; when any volume is later revised, this table must be updated in sync.
3. Windsurf and Trae pricing conflicts across sources (Windsurf Pro has both $15 and $20 figures; Teams has both $30 and $40 figures); this table prefers official vendor figures, marked [To be verified].
4. Claude Code user numbers and the specific pricing of Codex and Gemini CLI have not been authoritatively disclosed; marked [To be filled].
5. Some evaluation dates for Terminal-Bench 2.0 and SWE-bench Pro have not been confirmed on official pages; marked [To be verified].
6. This group found no first-hand source supporting "enterprise deployment effect data"; all case-style effect figures are marked [To be verified] or were not adopted.
7. Platforms 14–20 (v1.1 increment batch): Qoder's organizational-tier pricing, WorkBuddy's individual/enterprise pricing and model list, DeepSeek's original benchmark reports, Antigravity's pricing and Changelog details, OpenCode's MAU basis and Zen pricing, Junie's current quota values, and Augment Code's pricing and version timeline have each been itemized in the "Information Gap Statement" of the corresponding volume, marked [To be verified] / [To be filled]; the conflicting account of the successor relationship between Kiro and Amazon Q Developer is clarified in Section 1.1 of 12-kiro.md.

9. References

1. 2025 Stack Overflow Developer Survey — Stack Overflow, 2025-07-30. https://survey.stackoverflow.co/2025/
2. Stack Overflow 2025 Developer Survey press release — Stack Overflow, 2025. https://stackoverflow.co/company/press/archive/stack-overflow-2025-developer-survey/
3. DORA 2025 State of AI-assisted Software Development — Google Cloud / DORA, 2025. https://dora.dev/
4. Effective context engineering for AI agents — Anthropic, 2025. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
5. Sandboxing: a safer and more autonomous approach — Anthropic, 2025. https://www.anthropic.com/engineering/claude-code-sandboxing
6. Harness engineering: leveraging Codex in an agent-first world — OpenAI, 2026-02-11. https://openai.com/index/harness-engineering/
7. Effective harnesses for long-running agents — Anthropic, 2025. https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
8. Harness design for long-running application development — Anthropic, 2026. https://www.anthropic.com/engineering/harness-design-long-running-apps
9. Introducing the Model Context Protocol — Anthropic, 2024-11-25. https://www.anthropic.com/news/model-context-protocol
10. Model Context Protocol official site and specification — MCP / AAIF, 2024–2026. https://modelcontextprotocol.io/
11. Terminal-Bench official site — Stanford / Laude Institute, 2025–2026. https://www.tbench.ai/
12. SWE-bench official site — Princeton / community, 2023–2026. https://www.swebench.com/
13. GitHub Copilot plans (official documentation) — GitHub, 2026. https://docs.github.com/zh/copilot/get-started/plans
14. Cursor official pricing page — Anysphere, 2026. https://cursor.com/pricing
15. TRAE official site — ByteDance, 2026. https://www.trae.ai/