API model: gpt-5.3-chat-latest available (aka "instant" on ChatGPT)

API
1

No API-only model specifically yet, but you get what was just released to ChatGPT, “experiment with the latest version with less feature support”:

gpt-5.3-chat-latest Supports:
temperature: no
top_p: no
Real system role: no

  • Unlike prior models of “chat”, currently supports sending reasoning_effort, internally: ‘minimal’:5, ‘low’:16, ‘medium’:64, and ‘high’:200.
    (all ineffective to get lots of thinking done for minimal tasks, but at high, up to 700/700 split of reasoning-to-seen tokens for increasing question).
  • “verbosity” does have an effect: “low” is extremely low output (internal “oververbosity” of 1)

No pricing yet, but probably follows 5.3 codex, along with no discount for running on service_tier or batch.

3 Likes
2

@_j have you done any benchmarks on speed/latency? Is it really that “instant”?

3

0-day benchmarks before anybody heard of it aren’t going to be worthwhile for characterizing what performance OpenAI will deliver in the future. This model, constantly changing with whatever is running on ChatGPT, will get an immediate “pass” once there is an API-oriented 5.3 to actually build on.

The model will, in an about-face again, accept “minimal” but not “none” reasoning effort. Earlier “none” reasoning was a lie: you could talk your way into more reasoning than the amount they were hiding from being billed, and get it in usage.

2 Likes
4

And like that: no reasoning effort choice any more for this model.

So…OpenAI lurking, reading → patching, but not for the multiple broken API surfaces otherwise.