AI Coding Tool Pricing: Type A vs Type B (2026)

The Type A / Type B Split: Why the Split Matters

We split AI coding tool pricing into two types so buyers can predict cost shape. Scorers must cite which type of cost behaviour applies to each platform. The old anchor averaged two very different cost pathologies into one score. The new anchor does not.

The practical implication: if your iteration workflow scales with project size rather than edit size, a Type A platform will bankrupt you at the same rate it would have at month one. A Type B platform at least lets you respond — smaller contexts, tighter diffs, better caching, flat-cap tiers.

Cost Reality: Marketing Price vs What Users Actually Pay

Headline subscription prices hide the consumption layer. Every tool in this table advertises a $20–50 entry tier; real-world invoices tell a different story. The right-most column applies the pricing-type classification.

Real-world figures triangulated across platform pricing pages, The Register coverage of the Replit effort-based pricing complaints, Trustpilot Manus cases, and community-reported HN / Reddit threads during the Cursor June 2025 and Lovable July 2025 repricing events.

Four Pricing Failure Patterns (Taxonomy)

Beyond the Type A / Type B split, the 2025–2026 AI coding market produced four distinct usage-pricing failure modes. Each has at least one dated incident behind it. Bolt is cleanly structural (#0 in the list below); the other four patterns all fall inside Type B but produce the same user outcome — a bill several multiples of what the marketing page implied.

Why Bolt Is Uniquely Type A (and Cursor Isn't)

Every “Bolt is expensive” take in 2025 missed the architectural reason. The token problem is not about Bolt's plan tiering — it is about the WebContainer runtime itself. WebContainer is Node running inside the browser via WebAssembly, with no persistent server-side index of project state. When the agent takes a turn, the planner needs context about the full project to reason about what to edit. Without a persistent index, that context has to be reconstructed from the file tree each turn.

Cursor, Claude Code, Aider, Cline — all maintain persistent indices. Edit a single component and only that component's context (plus a small dependency graph) is included. The cost scales with the edit, not the project. The June 2025 Cursor repricing (a June 2025 incident that exhausted $7,000 annual subscriptions in a day) remains a useful illustration: Cursor users were hitting unit-economics pathology — which Cursor adjusted via plan restructure — not structural pathology. That is the difference between something a vendor can fix and something a vendor cannot.

The diagnostic: if your per-prompt token cost is roughly constant across a project's lifecycle, you're on Type B. If your per-prompt cost tracks project size — if editing one line in a 200-file project costs 10x what it cost in a 20-file project — you're on Type A. In our scoring, Bolt's Pricing score is pinned at 1 regardless of plan choice, and its Rate Limits marker is downgraded to 2 based on the 55M-tokens-in-8-days pattern.

Framework note: earlier versions of the Vibedex anchor did not distinguish architectural from economic pricing pathology. Scorers must cite which applies, because the advice to a buyer is fundamentally different: mitigate Type B via prompt engineering and flat-cap tiers; avoid Type A entirely if you expect the project to grow.

The Single Question Every Buyer Should Ask

The base rate for pricing change inside 12 months in this category is roughly 1. Treat any paid-tier commitment as a 90-day decision, not an annual one. Per Vibedex our scoring, rising-hard platforms (Claude Code, Cursor, Lovable) carry a 60-day expiry — explicitly because these vendors ship pricing changes faster than quarterly benchmarks can track.

Best Pricing Transparency: Three Picks

If your constraint is bill predictability rather than absolute capability, three tools stand out in the set. All three win by removing the vendor's degree of freedom to reprice you invisibly.

Decision Framework: Project-Scope vs Edit-Scope

The single most useful question for picking an AI coding tool on price is: do your iterations scale with project size, or with edit size?

App-builder platforms (Lovable, Replit, Manus, Base44) conflate these because the agent ranges across the whole project by design. Any app-builder with a Type A runtime is acutely dangerous; any app-builder with volatile Type B pricing (Replit) is moderately dangerous. A deliberate Lovable user rotating out after V1 ships is fine; a founder who keeps iterating inside Lovable / Replit / Bolt for 3–6 months will hit one of the four failure patterns above.

Related Reading

Sources & References

All external sources were verified as of April 2026. Ratings and metrics reflect the most recent data available at time of review.

1. Cursor — Pricing (official)(cursor.com)
2. Lovable — Pricing & Plans (official)(lovable.dev)
3. Replit — Pricing (official)(replit.com)
4. Bolt.new (official)(bolt.new)
5. Manus (official)(manus.im)
6. Aider — AI pair programming in your terminal(aider.chat)
7. Cline — Autonomous coding agent (official)(cline.bot)
8. The Register — Replit customers infuriated by surprise cost overruns (Sept 2025)(theregister.com)
9. Answer.AI — Thoughts on a Month of Devin (Husain, Flath, Whitaker, 8 Jan 2025)(answer.ai)
10. Manus blog — Manus joins Meta(manus.im)
11. Mindgard — Manus Rubra Full Browser Remote Control (Rich Smith, 1 Dec 2025)(mindgard.ai)
12. TechCrunch — Lovable raises $330M at $6.6B valuation (Dec 2025)(techcrunch.com)

FAQ