L
LLLOS.ai
Career
L
◆ AI Learning · Situational

Why did moving from a 200K to a 1M context window make Claude Code sessions more expensive, not just longer?

The old 200K context limit forced a compaction that reset accumulated context. The 1M window removed that forced reset — nothing stops the meter from running indefinitely.

By Pawan Nayar · from the LLOS AI Learning record

The rule

The old 200K context limit forced a compaction that reset accumulated context.

The 1M window removed that forced reset — nothing stops the meter from running indefinitely.

When Claude Code hit the 200K context limit, it compacted the history — summarising and discarding old turns. This was friction, but it was also a cost governor. The accumulated input could never exceed 200K tokens. The 1M window means a session can run for hours without hitting that ceiling. Without a forced compaction, the context grows unbounded. A developer who would have been stopped and reset at 200K now runs to 600K before noticing, and every turn in that range bills at the full inflated input size.

A developer runs a 4-hour debugging session. Under 200K, it would have been compacted twice and cost ~$8. Under 1M with no manual /compact, the same session costs ~$40.

Words you'll need

Resolved from this lesson's own record — real definitions, not invented for this page.

Tokenization

Splitting text into tokens (often subword pieces) for model processing

Prompt Engineering

Crafting questions to get better AI responses

Long-Context Models

Models that remember more previous text

Transformer Architecture

A model design for understanding language and context

Questions people ask

More from the same record, pulled by shared topic — not written new for this page.

How can a user make sure an AI model actually reads and uses a specific file, like a 'gold file'?

When a user tells an AI model to use a 'gold file', the AI model often just hears 'important reference'. The AI model then gives a confident answer.

Why does a Claude Code session get expensive as the conversation grows, even when individual turns are short?

An agentic framework like Claude Code holds conversation state, but the API has no concept of already-paid-for. Each turn packages the full history and sends it as input.

What information is lost when an AI model's session ends, and how should a developer update a documentation file?

When an AI model's session ends, all past conversations are gone. Decisions made, problems found, and solutions tried are lost.

When is using Claude Code the wrong instrument, even when the topic is technical?

Claude Code carries a session overhead: system prompt, tool definitions, and accumulated file context all bill as input on every turn. A pure conversation — explaining why output was truncated, reasoning about token accounting — needs none of that.

What to remember

The old 200K context limit forced a compaction that reset accumulated context. The 1M window removed that forced reset — nothing stops the meter from running indefinitely.