Here is the current state of requesting low detail on that two-tile-size image that would otherwise use 1+4 base tiles instead of 1. Generally fixed; some remaining expectations not met:
| model |
chat completions |
responses |
| gpt-5.2-2025-12-11 |
273 |
327 |
| gpt-5.2-pro-2025-12-11 |
— |
327 |
| gpt-5.2-chat-latest |
292 |
327 |
| gpt-5.1-codex-max |
— |
70 |
| gpt-5.1-codex |
— |
70 |
| gpt-5.1-codex-mini |
— |
272 |
| gpt-5.1-2025-11-13 |
70 |
70 |
| gpt-5.1-chat-latest |
70 |
70 |
| gpt-5-pro-2025-10-06 |
— |
70 |
| gpt-5-codex |
— |
70 |
| gpt-5-2025-08-07 |
70 |
70 |
| gpt-5-mini-2025-08-07 |
273 |
272 |
| gpt-5-nano-2025-08-07 |
272 |
272 |
| gpt-5-chat-latest |
70 |
70 |
| o3-pro-2025-06-10 |
— |
85 |
| o3-2025-04-16 |
75 |
75 |
| o4-mini-2025-04-16 |
272 |
272 |
| o1-pro-2025-03-19 |
— |
75 |
| o1-2024-12-17 |
75 |
75 |
| gpt-4.1-2025-04-14 |
85 |
85 |
| gpt-4.1-mini-2025-04-14 |
272 |
272 |
| gpt-4.1-nano-2025-04-14 |
272 |
272 |
| gpt-4o-2024-11-20 |
85 |
85 |
| gpt-4o-2024-08-06 |
85 |
85 |
| gpt-4o-mini-2024-07-18 |
85 |
85 |
| gpt-4o-2024-05-13 |
85 |
85 |
| gpt-4-turbo-2024-04-09 |
85 |
85 |
Remaining billing anomalies
- Overbilling: gpt-5.2 sees a transition to “patches”, where Chat Completions sees the correct expected patches count. On Responses, however, it appears the “mini” billing multiplier of 1.2x is still being applied, despite this being a full-price (and even higher price) model;
- Overbilling: gpt-5.2-chat has anomalous extra billing per image on Chat Completions, which cannot be reconciled to any extra patches column or row. It is not overhead, it is per image.
- Overbilling: o3-pro is “o3” - why billing 85 instead of 75 still?
- Overbilling: a single extra token overhead on just CC with gpt-5.2 and gpt-5-mini? Not an error in my multiplier deobfuscation code, as Responses returns the exact expectation.
Anomaly validation of per-image overbilling
Running ten image inputs in a user message verifies these are per-image billing issues
| model |
chat completions |
responses |
| gpt-5.2-2025-12-11 |
2730 |
3270 |
| gpt-5.2-chat-latest |
2920 |
3270 |
| gpt-5-mini-2025-08-07 |
2730 |
2720 |
Detailed usage log of anomalous model calls
— Testing gpt-5.2-2025-12-11 (Chat Completions)
gpt-5.2-2025-12-11 Image usage: 2730 Image prompt tokens: 2730 Total Usage: 2737
pong
| input tokens: 2737 |
output tokens: 4 |
| uncached: 177 |
non-reasoning: 4 |
| cached: 2560 |
reasoning: 0 |
— Testing gpt-5.2-2025-12-11 (Responses)
gpt-5.2-2025-12-11 Image usage: 3270 Image prompt tokens: 3270 Total Usage: 3277
pong
(Also, the image you attached appears to be a solid magenta block with no visible details.)
| input tokens: 3277 |
output tokens: 26 |
| uncached: 717 |
non-reasoning: 26 |
| cached: 2560 |
reasoning: 0 |
— Testing gpt-5.2-chat-latest (Chat Completions)
gpt-5.2-chat-latest Image usage: 2920 Image prompt tokens: 2920 Total Usage: 2927
pong 
I’m here.
(Also, the image you sent appears to be a solid brigh
| input tokens: 2927 |
output tokens: 34 |
| uncached: 239 |
non-reasoning: 34 |
| cached: 2688 |
reasoning: 0 |
— Testing gpt-5.2-chat-latest (Responses)
gpt-5.2-chat-latest Image usage: 3270 Image prompt tokens: 3270 Total Usage: 3277
pong
I’m here.
(Also, the image appears to be a solid bright magenta color with no visible details.)
| input tokens: 3277 |
output tokens: 33 |
| uncached: 589 |
non-reasoning: 33 |
| cached: 2688 |
reasoning: 0 |
— Testing gpt-5-mini-2025-08-07 (Chat Completions)
gpt-5-mini-2025-08-07 Image usage: 3280 Image prompt tokens: 2730 Total Usage: 3287
pong — I can see the images. How can I help with them?
| input tokens: 3287 |
output tokens: 24 |
| uncached: 599 |
non-reasoning: 24 |
| cached: 2688 |
reasoning: 0 |
— Testing gpt-5-mini-2025-08-07 (Responses)
gpt-5-mini-2025-08-07 Image usage: 3270 Image prompt tokens: 2720 Total Usage: 3277
Pong — I see the image you uploaded. How can I help with it?
| input tokens: 3277 |
output tokens: 23 |
| uncached: 589 |
non-reasoning: 23 |
| cached: 2688 |
reasoning: 0 |
Markdown report:
| model |
chat completions |
responses |
| gpt-5.2-2025-12-11 |
2730 |
3270 |
| gpt-5.2-chat-latest |
2920 |
3270 |
| gpt-5-mini-2025-08-07 |
2730 |
2720 |
Documentation
- Images and vision needs an update to the patches calculation table to include gpt-5.2
- The description of detail:low there should directly note the variable tile cost per model, and that with patches vision models detail has no effect
Caching overbilling for images
Despite sending the same input to the same model throughout tests, there is much more uncached context being billed than the increments of 128 token blocks that is documented.
Then, this additional overbilling despite a cache hit is particularly worse using Responses in conjunction with vision:
- a jump from uncached: 239 to 589 on gpt-5.2-chat-latest
- a jump from uncached: 177 to 717 on gpt-5.2
The overbilling for alleged cache miss can be directly attributed to the billing formula of a “multiplier” being applied outside of cache billing:
717 - 177 = 540
- 10 images to gpt-5.2 on Chat Completions: 2737 input tokens
- 10 images to gpt-5.2 on Responses: 2737 input tokens
- A difference of 540 tokens from each image overbilled by 54 tokens (and each overbilled by +1)
Conclusion: On Responses, OpenAI is not just overbilling the multiplier, but also excluding this overbilling amount from the cache discount. Formulaic failure.
Vision model truth table
This Python dict with limited fields is used for my scripting. It informs of endpoint compatibility, vision method and cost multiplier (with other gates merely notes). “plus” is the token overhead of any request (while “msg”, per-message overhead, seems unwavering).
MODEL_CAPABILITIES = {
# GPT-5.2 Family (reasoning.effort "none" will now tolerate temperature/top_p on non-chat)
"gpt-5.2-2025-12-11": {"cc": 1, "responses": 1, "msg": 4, "plus": 2, "vision": "patch", "mult": 1.0, "verbosity": 1, "min_effort": "none", "alias": "gpt-5.2"},
"gpt-5.2-pro-2025-12-11": {"cc": 0, "responses": 1, "msg": 4, "plus": 2, "vision": "patch", "mult": 1.0, "verbosity": 1, "min_effort": "none", "alias": "gpt-5.2-pro"},
# "chat" in gpt-5.x: Not supporting sampling parameters even at "none", the opposite
"gpt-5.2-chat-latest": {"cc": 1, "responses": 1, "msg": 4, "plus": 2, "vision": "patch", "mult": 1.0, "verbosity": 1, "min_effort": "none"},
# GPT-5.1 Family (reasoning.effort: "none" introduced as default) - any "min_effort" will indicate a reasoning model
"gpt-5.1-codex-max": {"cc": 0, "responses": 1, "msg": 4, "plus": 2, "vision": "tile", "tile": 70, "min_effort": "low"},
"gpt-5.1-codex": {"cc": 0, "responses": 1, "msg": 4, "plus": 2, "vision": "tile", "tile": 70, "min_effort": "low"},
"gpt-5.1-codex-mini": {"cc": 0, "responses": 1, "msg": 4, "plus": 2, "vision": "patch", "mult": 1.2, "min_effort": "low"},
"gpt-5.1-2025-11-13": {"cc": 1, "responses": 1, "msg": 4, "plus": 2, "vision": "tile", "tile": 70, "verbosity": 1, "min_effort": "none", "alias": "gpt-5.1"},
"gpt-5.1-chat-latest": {"cc": 1, "responses": 1, "msg": 4, "plus": 2, "vision": "tile", "tile": 70, "min_effort": "medium"},
# GPT-5 Family (reasoning.effort: "minimal" floor w default "medium", new "verbosity")
"gpt-5-pro-2025-10-06": {"cc": 0, "responses": 1, "msg": 4, "plus": 2, "vision": "tile", "tile": 70, "verbosity": 1, "min_effort": "high", "alias": "gpt-5-pro"},
"gpt-5-codex": {"cc": 0, "responses": 1, "msg": 4, "plus": 2, "vision": "tile", "tile": 70, "min_effort": "low"},
"gpt-5-2025-08-07": {"cc": 1, "responses": 1, "msg": 4, "plus": 2, "vision": "tile", "tile": 70, "verbosity": 1, "min_effort": "minimal", "alias": "gpt-5"},
"gpt-5-mini-2025-08-07": {"cc": 1, "responses": 1, "msg": 4, "plus": 2, "vision": "patch", "mult": 1.2, "verbosity": 1, "min_effort": "minimal", "alias": "gpt-5-mini"},
"gpt-5-nano-2025-08-07": {"cc": 1, "responses": 1, "msg": 4, "plus": 2, "vision": "patch", "mult": 1.5, "verbosity": 1, "min_effort": "minimal", "alias": "gpt-5-nano"},
"gpt-5-chat-latest": {"cc": 1, "responses": 1, "msg": 4, "plus": 3, "vision": "tile", "tile": 70, "sampling": 1},
# O-Series (reasoning.effort: "low" floor, no verbosity)
"o3-pro-2025-06-10": {"cc": 0, "responses": 1, "msg": 4, "plus": 2, "vision": "tile", "tile": 75, "min_effort": "low", "alias": "o3-pro"},
"o3-2025-04-16": {"cc": 1, "responses": 1, "msg": 4, "plus": 2, "vision": "tile", "tile": 75, "min_effort": "low", "alias": "o3"},
"o4-mini-2025-04-16": {"cc": 1, "responses": 1, "msg": 4, "plus": 2, "vision": "patch", "mult": 1.72, "min_effort": "low", "alias": "o4-mini"},
"o1-pro-2025-03-19": {"cc": 0, "responses": 1, "msg": 4, "plus": 2, "vision": "tile", "tile": 75, "min_effort": "low", "alias": "o1-pro"},
"o1-2024-12-17": {"cc": 1, "responses": 1, "msg": 4, "plus": 2, "vision": "tile", "tile": 75, "min_effort": "low", "alias": "o1"},
# deep-research in model requires internal RAG tool or web_search with context:medium, no user location
#"o3-deep-research-2025-06-26": {"cc": 0, "responses": 1, "vision": "tile", "tile": 75, "alias": "o3-deep-research"}, #815t-375,
#"o4-mini-deep-research-2025-06-26": {"cc": 0, "responses": 1, "vision": "patch", "mult": 1.72, "alias": "o3-deep-research"}, #909t-468
# GPT-4x Family (Standard Sampling)
"gpt-4.1-2025-04-14": {"cc": 1, "responses": 1, "msg": 4, "plus": 3, "vision": "tile", "tile": 85, "sampling": 1, "alias": "gpt-4.1"},
"gpt-4.1-mini-2025-04-14": {"cc": 1, "responses": 1, "msg": 4, "plus": 3, "vision": "patch", "mult": 1.62, "sampling": 1, "alias": "gpt-4.1-mini"},
"gpt-4.1-nano-2025-04-14": {"cc": 1, "responses": 1, "msg": 4, "plus": 3, "vision": "patch", "mult": 2.46, "sampling": 1, "alias": "gpt-4.1-nano"},
"gpt-4o-2024-11-20": {"cc": 1, "responses": 1, "msg": 4, "plus": 3, "vision": "tile", "tile": 85, "sampling": 1},
"gpt-4o-2024-08-06": {"cc": 1, "responses": 1, "msg": 4, "plus": 3, "vision": "tile", "tile": 85, "sampling": 1, "alias": "gpt-4o"},
"gpt-4o-mini-2024-07-18": {"cc": 1, "responses": 1, "msg": 4, "plus": 3, "vision": "tile", "tile": 85, "mult": 33.333, "sampling": 1, "alias": "gpt-4o-mini"},
"gpt-4o-2024-05-13": {"cc": 1, "responses": 1, "msg": 4, "plus": 3, "vision": "tile", "tile": 85, "sampling": 1},
"gpt-4-turbo-2024-04-09": {"cc": 1, "responses": 1, "msg": 4, "plus": 3, "vision": "tile", "tile": 85, "sampling": 1, "alias": "gpt-4-turbo"},
# special: "computer-use-preview" takes only screenshot tool return; any CC "search" model takes no images
#"computer-use-preview": {"cc": 0, "responses": 1, "msg": 4, "plus": 2, "vision": False, "tile": 65}, # requires truncation:auto
#"gpt-5-search-api-2025-10-14": {"cc": 1, "responses": 0, "msg": 4, "plus": 2, "vision": False, "alias": "gpt-5-search-api"},
}
