Models, Tokens, Context, and Usage Limits

Choose a model based on task difficulty, latency, budget, and verification risk. Model names, plan allowances, and reset rules change quickly. Treat numbers as dated snapshots and check provider documentation before purchasing or planning a deadline.

1. LLM, Tokenizer, Token, and Context

An LLM generates token sequences. A tokenizer converts text and code into tokens; tokens are not identical to words. Context includes prompts, files, instructions, tool output, images, logs, and conversation history. The context window is the maximum context the model can process at once.

Large context does not guarantee correct output. Irrelevant files can dilute important rules, while missing evidence forces the model to guess.

2. Why a Model Can Forget or Be Wrong

Common causes include truncated history, noisy context, conflicting documentation, stale assumptions, hidden tool output, ambiguous acceptance criteria, and a long task that changes state between steps. Reduce these risks with a short source-of-truth list, small stages, checkpoints, tests, and a handoff note.

3. Model Families and Coding Work

Providers offer fast economical models, balanced general models, and higher-capability models. The names differ, but the decision pattern is stable. Fast models are useful for search, formatting, small tests, and predictable edits. Balanced models fit most features and debugging. High-capability models are valuable for ambiguous architecture, difficult cross-file changes, security analysis, and stubborn root-cause work.

Do not use the most expensive tier for every operation. Escalate when the task or evidence justifies it.

4. An Efficient Selection Strategy

Start with the tool's recommended default for a small scoped task. Measure whether it can follow repository rules, edit the correct files, run checks, and explain its diff. Move to a stronger model when the task spans several systems, has high blast radius, repeatedly fails verification, or requires deeper reasoning. Move back to a faster model for mechanical follow-up work.

5. Model Selection Matrix

6. Usage Windows Are Not Universal

A five-hour window and a weekly limit are not universal rules. OpenAI and Anthropic may use session windows and additional plan-specific limits. Gemini products can use daily request quotas and rate limits. Z.AI documents its own usage windows and credits. OpenCode inherits the rules of the configured provider. Kimi limits depend on the current product and plan.

The actual consumption may depend on model, repository size, tool calls, task complexity, cache behavior, and parallel sessions. A displayed message count is not always a fixed number of prompts.

7. How to Read a Limit Correctly

Identify the product, plan, model, unit, rolling or fixed window, reset time, shared pools, fallback behavior, and what happens after the limit. Distinguish a subscription allowance from pay-as-you-go API billing. Record the date and official URL whenever you document a number.

8. Save Tokens, Quota, and Time

• Keep scope to one outcome.
• Point to relevant files instead of the whole repository.
• Ask for analysis before a risky edit.
• Reuse repository instructions for stable rules.
• Stop failed loops and provide the exact error.
• Run deterministic checks locally.
• Commit completed stages and write a handoff before switching agents.
• Avoid parallel agents until file ownership and worktrees are clear.

9. When a Limit Is Reached

Do not restart the project. Inspect status and diff, run available checks, commit valid completed work, and write HANDOFF.md with completed work, pending work, failing tests, current files, and the exact next step. Wait for reset or switch providers only when account and organizational policy allow it.

10. Freshness Rule

Review plan limits and model availability every one to two weeks, pricing about every two weeks, installation commands every two to four weeks, and evergreen concepts less frequently. If the page is stale, trust the linked provider documentation over cached numbers.