This page maps the leading AI-coding products as of April 2026 — Cursor, Cognition (Devin and Windsurf), OpenAI Codex, Claude Code, Replit Agent, GitHub Copilot, and Poolside — along the dimensions of product positioning, model strategy, valuation, and the structural constraints shaping competition. It organizes the field by who builds the model versus who builds the surface, and summarizes the developments (the Windsurf acquisition saga, Cursor's scale, frontier-lab integration) that have moved positions to date.
Snapshot
Valuation and funding
| Date | Company | Figure | Note |
|---|---|---|---|
| 2025-06 | Cursor (Anysphere) | $900M raised | June 2025 round |
| 2025 | Cursor (Anysphere) | ~$10B valuation |
Products and positioning
Cursor (Anysphere) is an independent AI code editor positioned as model-agnostic. It is perceived as a power-user tool, with features including Max Mode and Background Agents, and draws on coder-community loyalty. It carries no foundation-model moat and depends on the cost of the Claude, GPT, and Gemini models it serves. Anysphere was valued at roughly $10 billion in 2025 and raised $900 million in June 2025. See Cursor (Anysphere).
Cognition (Devin and Windsurf) runs a dual-surface strategy: Devin, positioned as an autonomous engineer, and Windsurf as an editor. Cognition purchased Windsurf in July 2025 after OpenAI's $3 billion deal to acquire it collapsed, giving Cognition the Windsurf install base and making it a direct Cursor competitor. The credibility of Devin's early demos has been questioned. See Cognition AI, (Source: Raw Sources/Windsurf Acquisition Saga 2025.md).
OpenAI Codex is built on GPT-5.3-Codex (February 2026), following GPT-4.1-Codex earlier. It is distributed through a Codex app on macOS (February 2026) and through Codex within ChatGPT. Its model vertical integration and direct-to-consumer distribution are offsetting factors against the time OpenAI lost when the Windsurf acquisition failed and against the coder-community loyalty held by Anthropic's Claude Code. See GPT-5.3 Codex.
Claude Code (Anthropic) runs on Claude Opus 4.6 and Claude Sonnet 4.6, with a Claude Mythos-based version anticipated. It is distributed through a CLI, direct API, and agent skills. It is perceived for technical quality and holds coder-community loyalty, but lacks a full IDE and requires developer sophistication. The New York Times published "Five Ways People Are Using Claude Code" on January 23, 2026. See Anthropic.
Replit Agent is positioned as a cloud IDE with a full stack covering hosting, agent, and deployment, running on Claude via Google Cloud Vertex AI. Its strengths are full-loop deployment and a large free user base; it serves a different customer segment (indie developers, students) than the enterprise-focused tools. See Replit.
GitHub Copilot (Microsoft) runs on OpenAI and internal models. It holds an incumbent install base and deep IDE integration, with distribution through GitHub enterprise representing the single largest reach in the field. Its product iteration is perceived as slower than Cursor's or Windsurf's.
Poolside is building a coding-specific frontier model, led by CEO Jason Warner (formerly GitHub CTO). In June 2025 Warner stated publicly that "Most companies shouldn't build foundation models," a position Poolside is itself attempting to satisfy by building a foundation model for coding specifically. See Poolside.
Smaller and adjacent players
Codeium, formerly Windsurf, was acquired in July 2025 and absorbed by Cognition. Magic.dev, Tabnine, JetBrains AI Assistant, and AWS CodeWhisperer are smaller players or enterprise-specialized tools. Adjacent AI-developer-tooling companies include Hugging Face (model hub and transformers library, the de-facto open-weight distribution channel), Granola (AI note-taking adjacent to developer productivity), Julius AI (AI data-analysis assistant, with coding-and-data overlap), and Standard Intelligence (a small dev-tool startup).
Model strategy: agnostic versus integrated
The field divides on whether a vendor builds its own model. Model-agnostic products (Cursor, Cognition) can serve Claude, GPT, or Gemini and select the best model per task, but hold no foundation-model moat. Model-integrated products (Codex, Claude Code) operate a vertical stack but are locked to one provider's model quality.
Workflow platforms are becoming an agnostic assignment layer above the agents themselves: Atlassian's Jira updates announced by July 21, 2026 let teams assign tasks to Claude Code, GitHub Copilot, or Atlassian's own agent, with chief product and AI officer Tamar Yehoshua saying she doesn't want "all of my company's information and contacts just going to Anthropic" (Source: theinformation.com).
Structural constraints
Self-reported productivity versus measured productivity
The METR study of developer productivity found that AI coding tools made experienced open-source developers 19% slower, even as those developers self-reported being faster (METR — Measuring Impact of Early-2025 AI on Experienced Open-Source Developer Productivity). The gap between measured and self-reported productivity bears on vendor productivity claims across the market.
Microsoft–OpenAI acquisition constraint
The collapse of OpenAI's Windsurf acquisition indicated that OpenAI cannot acquire technology Microsoft considers competitive with GitHub Copilot, a constraint on OpenAI's coding-AI M&A strategy ((Source: Raw Sources/Windsurf Acquisition Saga 2025.md)).
Positions as of April 2026
Several parties improved their positions through the Windsurf saga and related moves. Google completed a reverse acquihire of Windsurf and a reverse acquihire of Character.AI, giving it a competitive AI-coding position on multiple fronts. Cognition gained scale through its absorption of Windsurf. Anthropic holds coder-community loyalty around Claude Code, which functions as a commercial advantage.
Other parties lost ground or remained absent. OpenAI's Windsurf acquisition failed, leaving Codex to play catch-up, and its internal coding-AI account relies on unverified productivity claims. GitHub Copilot's perception slipped relative to Cursor and Claude Code. Apple had no AI-coding offering as of April 2026.
Competitive pressure persisted into mid-2026. On June 25, 2026 The Information reported that Google had begun reorganizing its recently launched AI coding strike team in an effort to close the gap with Anthropic, which it described as the most lucrative AI application category (Source: theinformation.com). The same day, an OpenAI adoption study (with Columbia, Duke, and the University of Pennsylvania) reported organizational Codex adoption rising from near zero in August 2025 to about 17%, with 99.8% of OpenAI's own employee output tokens coming from Codex (Source: axios.com). See Agentic AI.
The July 2026 pricing round
During the week of July 6, 2026, four coding-capable models shipped within roughly 48 hours at sharply different prices: OpenAI priced GPT-5.6 Sol at $5/$30 per million input/output tokens (ahead of GPT-5.5 on Terminal-Bench 2.1 but trailing Claude Fable 5 on SWE-Bench Pro), Meta priced Muse Spark 1.1 at $1.25/$4.25 with a one-million-token context window, SpaceXAI and Cursor priced Grok 4.5 — trained on trillions of tokens of Cursor interaction data — at $2/$6, and Cognition released SWE-1.7, a reinforcement-learned open-weight model optimized for Devin (Source: gettheleverage.com; bloomberg.com). Bloomberg analysis of the round described cost efficiency, not raw capability, as the models' biggest immediate selling point, with OpenAI saying GPT-5.6 is designed to complete more work while consuming significantly fewer tokens (Source: bloomberg.com). See Inference Economics and Token Pricing.
Relationships
- depends-on: AI Coding Agents.
- related: Cursor (Anysphere), Cognition AI, Replit, Poolside, (Source: Raw Sources/Windsurf Acquisition Saga 2025.md), METR — Measuring Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, The Darwin Gödel Machine (Sakana AI, May 2025), GPT-5.3 Codex.