Long-Context Price Cliffs: When Your Prompt Costs 2x to 5x More
Big context windows are useful, but some providers charge more once a prompt passes a threshold. A request that crosses the line can cost two to five times more.
Known thresholds (November 2026)
| Model | Input $ / 1M | Output $ / 1M | Note |
|---|---|---|---|
| Claude Haiku 5.5 | $0.10 | $0.50 | up to 100,000 tokens, then $0.50 / $2.50 |
| GPT-6 Sol | $2 | $10 | up to 272,000 tokens, then $4 / $15 for the entire request |
| Gemini 3.1 Pro | $2 | $12 | up to 200,000 tokens, then $4 / $18 |
- Grok 4.5 and Grok 4.3: higher rates above 200,000 tokens
- Claude Sonnet 5.5, Opus 5.5 and Fable 5.1: one rate across the full 1 million token window
Why it matters
Note that on GPT-6 Sol the higher rate applies to the whole request, not just the extra tokens. A prompt of 280,000 tokens costs more than twice as much as one of 270,000.
How to avoid the cliff
- Trim or summarize history before sending
- Retrieve only relevant passages instead of whole documents
- Split large jobs into several requests under the threshold
- Choose a model with flat pricing if you often send very long prompts
What a cliff looks like in dollars
On GPT-6 Sol the higher rate applies to the whole request. A prompt of 270,000 tokens costs about $0.54 in input at $2 per million. A prompt of 280,000 tokens, just 10,000 tokens more, costs about $1.12 at $4 per million. Adding 4% more text more than doubles the input cost.
Claude Haiku 5.5 has a similar step at 100,000 tokens. If the higher rate of $0.50 per million applies to the full prompt, a 110,000-token prompt costs about $0.055 in input, against $0.01 for a 100,000-token prompt. Check each provider's documentation to see whether the higher rate covers the full request or only the extra tokens.
Long context is not always needed
Sending an entire document set is convenient, but it is often wasteful. Models tend to use information at the start and end of a long prompt better than the middle, and every extra token adds cost and latency. Before choosing a long-context setup, test whether retrieval, which sends only the relevant passages, gives equal quality.
Strategies to stay under the threshold
- Retrieve, do not dump: search your documents and send the top passages.
- Summarize history: replace old chat turns with a short summary.
- Chunk the job: process a large document in sections, then combine the results.
- Trim formatting: remove boilerplate, repeated headers and unnecessary markup.
- Pick the right model: where long prompts are normal, use one with flat pricing across the window.
Combine with caching
If you do need a very long, stable context, for example a large codebase or manual that many requests share, caching can cut the repeated cost sharply. Check how cache pricing interacts with the long-prompt rates before you rely on it. See prompt caching explained.
Keep a threshold table
Write down the thresholds for each model you use, and enforce them in code with a token counter. A simple check before each call can warn you when a prompt is about to cross a cliff, or automatically trim it.
Frequently asked questions
Which models have flat pricing across the whole window?
At the time of writing, Claude Sonnet 5.5, Opus 5.5 and Fable 5.1 bill the full one million token window at one rate.
Does the higher rate apply to output too?
On several models it does. GPT-6 Sol goes from $10 to $15 per million output tokens above its threshold.
Where can I model the effect?
Use the cost calculator and manually enter the higher prices to see the new monthly total.
Estimate carefully
The cost calculator uses the base rates, so for long prompts check the thresholds above and adjust the prices manually. See also hidden LLM API costs.
Thresholds checked November 4, 2026.
Track spend by team and project: the AI Spend Tracker spreadsheet is launching soon.