Every AI coding tool's pricing page assumes you already know what a "credit" is, how it's different from a "token," and why some tools want your own API key while others don't. Here's what the jargon actually means, in plain English.
Pricing & Usage Terms
Credits — A tool's own made-up unit of usage, priced and sized differently by every vendor. One "credit" might cost a fraction of a cent in raw compute, or it might represent a full prompt-and-response cycle. Credits are not portable or comparable between tools — 1,000 credits on one platform tells you nothing about 1,000 credits on another. Always check what a credit actually buys (one message? one file edit? one full app generation attempt?) before comparing prices across tools.
Tokens — The actual unit AI models process text in — roughly three-quarters of a word in English. This is the underlying currency that "credits" are usually built on top of. Tools that bill you directly in tokens, rather than their own credit system, are more transparent about true cost, but the numbers get large fast — a single prompt can easily use thousands of tokens.
BYOK (Bring Your Own Key) — A pricing model where you supply your own API key from an LLM provider (Anthropic, OpenAI, etc.) and pay that provider directly for usage, instead of paying the tool a markup. The tool itself is often free or cheap; your real cost is the API bill. See BYOK vs Bundled Models for the full breakdown.
Bundled models — The opposite of BYOK: the tool's subscription price includes model access, and you never see or manage an API key. Simpler, but you're paying whatever markup the vendor builds in, and you can't easily swap models.
Seats — Per-user pricing, common on team and enterprise plans. A "seat" is one paid user account, usually billed monthly regardless of how much that person actually uses the tool — which is why some enterprise tools instead price by usage or a flat team fee rather than per seat.
Rate limits — A cap on how much you can use a tool in a given window (requests per hour, tasks per day), separate from your credit or token balance. You can have credits left and still be blocked by a rate limit — common on free tiers and lower paid tiers as an anti-abuse measure.
How the AI Actually Works
Agent / Agent mode — A mode where the AI doesn't just suggest one edit and wait — it plans a sequence of actions (read files, make edits, run commands, check the result) and executes them with limited or no confirmation at each step. See Agentic Coding Explained for the full spectrum from simple autocomplete to fully autonomous agents.
Context window — How much text (code, conversation history, files) a model can "see" at once, measured in tokens. A larger context window means an agent can reason about more of your codebase without losing track of earlier parts of the conversation — genuinely important for large-codebase work, less important for small, focused tasks.
Autocomplete / Tab completion — The simplest level of AI coding help: the model predicts and suggests the next few lines as you type, and you accept or reject with a keystroke. No planning, no multi-step action — just very good autocomplete.
Chat mode — You ask a question or describe a change in a conversational panel, and the AI responds with an explanation, a suggestion, or a proposed diff — but doesn't act on your files without you applying the change yourself.
Multi-file editing / Composer — The ability to make coordinated changes across several files in one operation, rather than one file at a time — table stakes for any AI IDE, notably absent from simple autocomplete extensions.
Hallucination — When a model confidently generates something that's wrong — a function that doesn't exist, an API that was never real, a "fix" that doesn't address the actual bug. Not unique to coding tools, but especially costly here because wrong code can look plausible and still compile.
Integration & Architecture Terms
MCP (Model Context Protocol) — An open standard, originated by Anthropic, that lets AI tools connect to external data sources and other tools in a consistent way — for example, letting Claude Code or Cursor pull a design directly from Google Stitch instead of you manually exporting and re-uploading it. Increasingly used as a common connector for linking AI coding agents to design tools, databases, and other services without a custom one-off integration for every pairing.
Worktree — A git feature that lets you check out multiple branches of the same repository into separate folders simultaneously. Increasingly relevant for AI coding because several newer agents use isolated worktrees to run multiple sub-agents on the same codebase without their changes colliding. See Multi-Agent Orchestration for how this plays out in practice.
SWE-bench / Terminal-Bench — Standardized benchmarks used to measure how well an AI coding agent performs on real-world software engineering tasks pulled from actual GitHub issues. Useful for rough comparison between tools, but treat any single benchmark percentage with some skepticism — real-world performance on your specific codebase and stack can differ meaningfully from a benchmark score.
Product & Business Terms You'll See on Tool Pages
Vibe coding — Describing what you want in natural language and letting an AI generate the app, rather than writing the code yourself. See What Is Vibe Coding? for the full definition and origin.
Fullstack builder — A tool (Lovable, Bolt, Replit) that generates a complete application — frontend, backend, and often a database — from a prompt, aimed at going from idea to deployed app with minimal manual coding.
AI IDE — A code editor built from the ground up with AI as a core feature (Cursor, Devin Desktop), as opposed to an AI extension bolted onto an existing editor. See AI IDEs vs Coding Agents vs AI Extensions for the full category breakdown.
Vendor lock-in — How hard it is to leave a tool and take your work with you. Tools that export clean, standard code (React, plain SQL) have low lock-in; tools with proprietary formats or restricted exports have high lock-in. Covered in depth in The Hidden Costs of Vibe Coding.
Credit-based pricing — Billing tied to usage (credits, as above) rather than a flat unlimited subscription — the dominant model for fullstack builders, meaning your actual monthly cost depends heavily on how much you iterate, not just which plan you pick.
Related Guides
- What Is Vibe Coding? — start here if you're completely new
- Agentic Coding Explained — the five levels of AI autonomy, in depth
- BYOK vs Bundled Models — the full cost and flexibility tradeoff
- The Hidden Costs of Vibe Coding — what "credits" and lock-in actually cost you
- Multi-Agent Orchestration — what happens when you run several agents at once