Cursor Composer 2.5 and the in-house model economics — when the IDE owns the model, the price war begins

Cursor's Composer 2.5 launch is the moment the AI coding tool market's pricing structure becomes unstable. An in-house model that matches Opus 4.7 and GPT-5.5 at one-tenth the per-token cost changes Cursor's unit economics — and changes the competitive position of every AI coding tool that still depends on frontier-lab API spend. GitHub Copilot's June 1 shift to AI Credits flex billing is the parallel signal: the all-you-can-use AI coding subscription is essentially over.

The unit economics shift is the news. Composer 2.5 at $0.50 input / $2.50 output per million tokens, with capability matching Opus 4.7 and GPT-5.5 on coding benchmarks, gives Cursor a per-token cost roughly an order of magnitude below the comparable frontier-lab API rates. For Cursor as a business, this means the $20/month Pro tier can support meaningfully more usage per user before margin erodes. For Cursor as a competitor, this means the marginal cost of any incremental Cursor user is now structurally lower than the marginal cost of the corresponding GitHub Copilot or Claude Code user (both of which depend on per-token API spend with Anthropic and OpenAI).

The Build in Parallel and PR review features compound the strategic position. Build in Parallel dispatches multiple competing implementations of a task and merges the best — the kind of feature that's expensive when each implementation runs through a frontier-lab API call and is essentially free when each implementation runs through Cursor's own model. PR review is Cursor-native code review against the user's branch — the workflow that GitHub Copilot's Workspace and Claude Code's PR-creation features address. With Composer 2.5 plus these features, Cursor's product surface is now competitive on every dimension where Claude Code and GitHub Copilot used to have advantages.

The GitHub Copilot move to AI Credits flex billing on June 1 is the parallel signal. The same $10 and $39 sticker prices, but usage now consumes credits from a pool against allocations. The shift is GitHub's acknowledgment that frontier-model usage costs are no longer absorbable as a flat subscription. For light users, the AI Credits pricing is neutral or slightly favorable. For heavy users — the senior developers whose workflows drive the highest model spend — the effective monthly bill can run two to three times the prior sticker price. That's the cohort most likely to switch to Cursor's flat-rate-with-higher-usage-envelope offer.

The competitive implication for Anthropic and OpenAI is mixed. On the one hand, Composer 2.5 is one less customer for Anthropic and OpenAI's frontier-tier API spend — Cursor's volume previously routed through both labs is now routed through Cursor's own model. On the other hand, the developer-cohort revenue compounding that justified Anthropic's $900B valuation still works because Claude Code remains the most-preferred tool for the senior developer cohort (46% in JetBrains' survey versus 9% for Copilot, with Cursor not in the top two for that cohort). Cursor wins the mainstream developer tier; Claude Code holds the senior tier; GitHub Copilot defends its installed base on the AI Credits pricing model. Three different competitive positions in three different developer segments.

The broader pattern is that vertically integrated AI products are increasingly defaulting to in-house models. Cursor's Composer 2.5 is the visible example; less visibly, Replit, JetBrains, Vercel, and the dedicated agentic coding startups (Kiro, Aider, Continue, Cline) are all evaluating in-house model strategies. The economic logic is the same one Apple followed with custom silicon: when you control the application, controlling the underlying compute substrate becomes strategically essential to defend margins. The frontier labs that supply the API now compete against an emerging ecosystem of in-house competitors whose models are good enough for most use cases — and the frontier labs' value proposition has to be capability superiority sufficient to justify the cost premium.

The line: the model is becoming a commodity input. The IDE is becoming the product.

Nipralo — Best AI Coding Tools 2026 → · Lushbinary — AI Coding Agents Comparison 2026 → · SitePoint — Claude Code vs Cursor vs Copilot 2026 →