AI Coding Assistants: Completion, Agents, and the Credit Trap

The AI coding category has converged on the same interface — an editor, a chat panel, an inline suggestion — which makes the tools look interchangeable. They are not, because underneath the identical UI there are two different products with different cost curves, and almost nobody works out which one they are buying before the bill arrives.

Completion versus agency

Completion predicts the next few lines from the surrounding context. It is cheap, fast, and it improves typing speed. The cost per suggestion is tiny, the value is real but bounded, and a free tier can usually sustain it.

Agentic editing reads a request, plans changes across several files, runs commands, checks the result, and iterates. It is expensive, slow, and it can do work you would otherwise do yourself. The cost per task is orders of magnitude higher, because a single task is dozens of model calls.

The trap is that vendors sell the second and price the first. A plan advertised as "$15 a month for AI" often includes a small allowance of premium requests, and a single agent task can consume ten or twenty of them. A developer who spends the afternoon refactoring with an agent can exhaust a monthly allowance in two hours.

This is not a gotcha, exactly — it is a genuine cost that GPU inference imposes. But it is presented badly, and the result is that people budget for autocomplete and get an agent bill.

Work out which one you need

Ask what you actually want the tool to do:

• Type faster, remember boilerplate, write tests from a signature. You want completion. The free tier of several tools covers this indefinitely, and the paid plans are about team features rather than capability.
• Change a function signature across forty files, migrate a dependency, implement a described feature. You want an agent. This is the case worth paying for, and the case where the credit meter bites.
• Understand an unfamiliar codebase. This is a third thing — retrieval and explanation rather than generation — and it is where the tools differ most in quality. Test it specifically, because a good completion model is often poor at codebase-wide context.

Most developers want all three in different proportions, which is why the tools are converging. But the cost driver is overwhelmingly the agent, so the question that matters is: how many agent tasks per month do you run?

Multiply by roughly ten to twenty premium requests per task. That is your real monthly requirement, and it is usually several times the headline allowance.

The four cost axes that differ

Premium request allowance. The main meter. Look at the number, then apply the multiplier above.

What counts as premium. Some tools count only the frontier models as premium and give unlimited access to smaller ones. That is a meaningful distinction: a smaller model handles most completion and simple edits, and only the hard multi-file reasoning needs the expensive one. A tool with unlimited small-model access and a small premium allowance can be cheaper in practice than one with a flat allowance.

Context indexing cost. Some tools charge for re-indexing a repository, or for the storage of the index. If you work on several large repositories, this adds up quietly.

Cache behaviour. Prompt caching reduces cost substantially on repeated context. Tools that cache aggressively can offer more usable capacity at the same price, and almost none of them advertise it.

The self-hosted option, and what it actually costs

Several tools in this category are open source and self-hostable, and the appeal is obvious: no per-seat fee, no credit meter, bring your own API key or run a local model.

The arithmetic is less simple than it looks:

• Local models need hardware. A model good enough for multi-file reasoning does not run well on a laptop, and the electricity and hardware cost is a real subscription paid upfront.
• API keys mean you pay per token directly, which for heavy agent use is often more than a flat subscription, because you are paying retail with no volume discount.
• Self-hosting costs attention. Updates, index management, and the afternoon you lose when it breaks are not free, and they are not on any invoice.

Self-hosting wins at high volume with a dedicated ops person. It loses for an individual developer, where the $15 flat fee is cheaper than the tokens.

Where the free tiers genuinely work

One or two tools offer unlimited autocomplete and chat to individuals, permanently, with no credit meter. The quality is below the frontier models and the agent capability is limited or absent, but for completion and simple edits it is competitive.

That is a real option and it is worth testing before paying. The honest framing: if you have never used an AI coding assistant, start on a free tier and find out whether you use it enough to justify a subscription. Most developers who try it do. Some discover they only wanted completion, and those people should not pay.

What to test, in order

1. A multi-file refactor you actually need. Not a demo — the real task on your real codebase. This tells you whether the agent works and how many requests it consumed.
2. The same task on the small model, if the tool offers one. If it succeeds, your effective capacity is much higher than the headline number.
3. Codebase question answering. Point it at an unfamiliar part of your repository and ask a question whose answer you know. This is where quality varies most.
4. The credit counter after all three. This is the number that decides the plan.

Do the four tests inside the trial window, in one sitting, and write down the request count. It is the only reliable way to compare, because the allowances are not comparable on paper.

The compressed answer

Completion is free or nearly free and you should use it. Agents are worth paying for and will cost more than the plan suggests, so budget by task count rather than by seat price. Test on a real refactor, count the requests, and pick the plan that covers that number with headroom.

For tool-by-tool detail, the AI coding assistants reviews record what each plan meters and whether the small-model tier is unlimited. Cursor vs Windsurf and Cursor vs Codeium compare the credit economics of those pairs directly, which is the axis where they actually differ.