Navigation Explore Code Guide
Context Window: 200K vs 1M
When to switch to the extended context window and at what cost
PDF editions · Guide v3.44.1 · Updated
PDF
← All cards
Context is model-specific
| Direct API model | Context | Maximum output |
|---|---|---|
| Haiku 5.5 | 1M native | 128K |
| Haiku 4.5 (legacy) | 200K | 64K |
| Sonnet 5.5 | 1M native | 128K |
| Opus 5.5 | 1M native | 128K |
| Fable 5.1 | 1M native | 128K |
Guide model snapshot, October 27, 2026. Provider settings can expose different limits. Inspect /context for the active session.
No native-1M premium
Opus, Sonnet, and Fable 5.x use standard rates beyond 200K input tokens (Haiku 5.5 is priced higher above 100K). They do not need the old long-context beta header. Sonnet 5.5 has no 200K variant on the direct API.
Older Opus 4.6 and Sonnet 4.6 have [1m] variants with plan/provider restrictions. Do not transfer their access rules to current models.
Price a complete request
Illustrative USD totals without caching or retries:
| Input / output | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| 50K / 5K | $0.15 | $0.30 |
| 150K / 20K | $0.50 | $1.00 |
| 500K / 50K | $1.50 | $3.00 |
Use separate input, output, cache-read, and cache-write counts. A context-window size is capacity, not a bill or an accuracy score.
Control accumulation
/context shows the token breakdown. /compact summarizes the conversation; /clear starts fresh. Save decisions and verification commands before resetting context.
/autocompact 500k saves a smaller compaction window. /autocompact auto restores the model default. Native 1M sessions normally compact around 967K.
Load relevant files rather than filling the available window. Split independent tasks when separate contexts help. No fixed percentage guarantees reliable recall. Idle compaction (v2.1.299) compacts a 200K+ conversation idle about 54 minutes before its one-hour cache expires; "idleCompaction": false disables it.
Evaluate retrieval
Older MRCR results for Opus 4.6 or Sonnet 4.5 remain results for those models and benchmark conditions. They do not establish retrieval quality for Sonnet 5.5 or Opus 5.5.
Test on your own document set with answerable questions, expected citations, and distractors. Measure missed evidence as well as correct answers.
Sources: Model configuration, model specifications, pricing.
Enter your email to read the full card and get the complete PDF bundle.
All content is free and open-source. We just ask for your email.