Worked examples
The simulator's four example sessions on GPT-5.5 with 24h retention, priced turn by turn by the same cost model. OpenAI says entries typically last about 30 minutes and can last up to 24 hours. We use 30 minutes.
-
Coding agent
Tool calls and edits in quick succession, with a few longer pauses. 40 turns, 20K tokens of starting context, 3K in and 600 out per turn, 1.5 minutes apart, three long breaks of 8 minutes.
Open this in the simulator- With caching
- $3.50
- Without
- $19.36
- Saved
- $15.86 82%
- Pays for itself on turn 2
- Cache stays warm the whole way
-
Long chat
A person reading and typing, with one long break. 30 turns, 3K tokens of starting context, 150 in and 500 out per turn, 3 minutes apart, one long break of 45 minutes.
Open this in the simulator- With caching
- $0.92
- Without
- $2.34
- Saved
- $1.42 61%
- Pays for itself on turn 2
- Cache goes cold 1 time
-
Big document
One large document, questioned a dozen times. 12 turns, 150K tokens of starting context, 100 in and 700 out per turn, 7 minutes apart.
Open this in the simulator- With caching
- $1.94
- Without
- $9.52
- Saved
- $7.58 80%
- Pays for itself on turn 2
- Cache stays warm the whole way
-
Unattended loop
A script calling the model nonstop, the prompt growing every turn. 200 turns, 30K tokens of starting context, 1.5K in and 300 out per turn, 30 seconds apart.
Open this in the simulator- With caching
- $37.59
- Without
- $321.99
- Saved
- $284.41 88%
- Pays for itself on turn 2
- Cache stays warm the whole way
- Long-context prices from turn 135
When caching pays for itself
For a prefix sent again and again inside the cache lifetime. The count does not depend on the prefix size, as long as it is above the minimum.
24h retention, about 30 min
There is no write charge. A read costs $0.50 per million tokens instead of $5.00, so caching never costs more than sending the prefix uncached.
A cached prefix pays for itself after 1 read within the lifetime.
Assumed lifetime OpenAI says entries typically last about 30 minutes and can last up to 24 hours. We use 30 minutes.
Caching rules for GPT-5.5
- Caching is automatic, with no write charge. A read costs 0.1x the input price.
- Only 24h retention is offered. Entries typically last about 30 minutes and can last up to 24 hours, so the simulator uses 30 minutes. Breakpoints fall every 2,048 tokens, so only whole multiples of that are cached.
- OpenAI says the minimum prompt length varies by request on models before GPT-5.6, so the simulator applies none.
- Short context is up to 272K input tokens. Above 272K, long-context prices apply: $10.00 input, $1.00 cached and $45.00 output. Assumed: the long-context rate applies to the whole request, cached and written tokens included. OpenAI's page states only the threshold.
Sources: OpenAI's pricing page, caching guide. See all prices and sources.
Similar models
Per million tokens, with the 40-turn coding agent on each model's default cache setting.
| Model | Input | Cache read | Output | Coding agent |
|---|---|---|---|---|
| GPT-5.5 | $5.00 | $0.50 | $30.00 | $3.50 |
| GPT-5.6 Sol | $4.00 | $0.40 | $20.00 | $2.72 |
| GPT-6 Astra | $10.00 | $1.00 | $50.00 | $6.81 |
| Claude Opus 5 | $5.00 | $0.50 | $25.00 | $5.23 |
| Gemini 4 Argon | $2.00 | $0.10 | $10.00 | $0.92 best case* |
* Gemini implicit caching, best case.
Questions
What is the cached input price for GPT-5.5?
Input costs $5.00 per million tokens, with no charge to write the cache, $0.50 to read it and $30.00 for output.
Is prompt caching worth it for GPT-5.5?
In the 40-turn coding agent example, caching costs $3.50 against $19.36 without, a saving of $15.86 (82%). With 24h retention there is no write charge, so caching never costs more than not caching.
How long does the GPT-5.5 cache last?
24h retention, about 30 min: OpenAI says entries typically last about 30 minutes and can last up to 24 hours. We use 30 minutes.
Does GPT-5.5 have a minimum cacheable prompt length?
OpenAI says the minimum varies by request on models before GPT-5.6, so the simulator applies none.
An estimate from published list prices, and your bill may differ. Token counts are what the example sets, not tokenized text. Not covered: batch and fast-mode prices, thinking tokens beyond the output shown, tool fees and regional uplifts.