OpenAI ships GPT-5.2 with adaptive reasoning budgets for API developers
The new flagship model decides on its own how much reasoning to spend per request, and OpenAI says median latency drops 40% without benchmark regressions. Demo content.
OpenAI has now confirmed the adaptive budget mechanism also ships to ChatGPT consumer plans next month, not only the API.
OpenAI has released GPT-5.2 to all API customers, introducing what the company calls adaptive reasoning budgets: the model now decides per request how much internal deliberation a task actually needs, instead of relying on a fixed reasoning-effort parameter.
What changed
Until now, developers had to pick a reasoning tier up front. GPT-5.2 replaces that with a single automatic mode plus an optional hard cap for cost control.
- ▸Median latency on short requests drops roughly 40%
- ▸Long-horizon agentic tasks keep full deliberation depth
- ▸A new
max_reasoning_tokenscap protects budgets - ▸Pricing per token is unchanged from GPT-5.1
"Developers should not have to guess how hard their own request is. The model knows better than they do." — OpenAI engineering post
Why it matters
Fixed reasoning tiers were a persistent source of wasted spend: teams defaulted to high effort everywhere because tuning per endpoint was tedious. Adaptive budgets move that optimization into the model itself, which is likely to pressure Anthropic and Google to answer with similar mechanisms.
Early testers report the biggest gains on classification and extraction workloads, where the model now answers almost instantly instead of overthinking trivial inputs.
▮Story updates
- 15:207 Jul 2026
OpenAI has now confirmed the adaptive budget mechanism also ships to ChatGPT consumer plans next month, not only the API.
- 14:507 Jul 2026
Initial developer reports show 30-45% latency reductions on short prompts, matching OpenAI's claims.
Source: OpenAI
