Context and tokens, what the model sees
Why the model forgets after 50 messages and what you can do about it.
What is a token?
A token is a word-unit for the model. Roughly: one word equals one to two tokens in English, a bit more in German. "Hallucinating" is around four tokens.
What is the context window?
The maximum amount of tokens a model can "see" at once. Like how much text fits on a desk. Push more in and the oldest part falls off.
Typical sizes (April 2026 state, this changes fast):
- GPT-5: 128k standard, 200k via API
- Claude Opus: 1M tokens in 1M-context mode
- Gemini 2.5 Pro: 2M tokens
200k tokens are roughly 400 pages of text. 1M is around 2,000 pages. The numbers keep growing, what's the maximum today is probably standard in six months.
Why this matters
In a long chat the oldest part of the conversation eventually gets dropped from context. The model "forgets" what was discussed at the start. That isn't a bug, that's the architecture.
For you that means:
- If you need a lot of context, repeat the most important points
- Summarize very long documents before pasting them in
- For really complex topics: start a new session with clean context
Context cache
Modern models cache repeating prompts. If you reuse the same text block (system prompt, document) often, it gets cached, which saves cost and time. We explain that in Level 2.
Effort levels and token budget
In Claude Code since April 2026 there's a second dial alongside the model: the effort level. Lower effort = fewer tokens for internal reasoning = cheaper and faster, but less thorough. Higher effort = more reasoning tokens = better on complex tasks, but more expensive.
The five levels since Opus 4.7 (April 16, 2026):
- low: fast answers, barely any thinking. Good for simple lookups, easy refactors, boilerplate code.
- medium: standard reasoning. Default for most use cases.
- high: thorough reasoning. For architecture decisions, debug sessions you don't want to dig through yourself, tough migrations.
- xhigh: introduced with Opus 4.7. Between high and max. For problems where you genuinely need deep analysis but want to avoid max-level cost. On the current top Opus variant, xhigh is the default.
- max: maximum reasoning. For truly tricky sessions where you need one really good answer.
What does it cost? On the current Opus variant reasoning tokens are priced like output tokens (25 USD per million tokens). An xhigh session can easily burn 30 to 60 percent more tokens than a medium session. On Max Plan it's flat-rate, on API you pay per token.
Practical tip: set default to medium or high, use xhigh or max deliberately for the hard tasks of the day. In ~/.bashrc as CLAUDE_CODE_EFFORT_LEVEL=high or in ~/.claude/settings.json, plus per-session override via flag when needed. If you're not on max with Opus yet but the model answers too shallowly on tough tasks, xhigh is often the better middle ground.
What changes from Level 4
From Level 4 on you'll learn memory systems that go beyond context. That's the way out of the "model forgets" problem.