o1-pro: Full Spec Sheet (June 2026)

o1-pro is OpenAI's highest-effort reasoning configuration. The base o1 model (released September 2024) introduced 'reasoning models' — models that produce a long internal chain of thought before writing the visible answer. o1-pro takes the same base model and runs it with substantially more reasoning compute per query: more tokens of internal reasoning, longer wall-clock per response, higher quality on the hardest tasks at correspondingly higher cost.

Pricing-wise, o1-pro is in a tier by itself. $150 input / $600 output per 1M is roughly 60× more expensive than GPT-5 ($1.25/$10) on input and 60× more expensive on output. A 1,000-in / 500-out call costs `0.001 × $150 + 0.0005 × $600 = $0.15 + $0.30 = $0.45` per call — and that's before reasoning tokens.

Reasoning tokens are the hidden cost. o1-pro can burn 10,000-50,000+ reasoning tokens on a complex problem before writing the visible answer. Those tokens bill at the output rate ($600/M) but are not returned to you. A query with 30,000 reasoning tokens + 500 visible output tokens bills `30,500 × $600/M = $18.30` for the output portion alone. A single hard query can cost $5-20.

OpenAI launched o1-pro at this price point because the unit economics of pro-tier reasoning compute are extreme — the model effectively runs the same prompt many times internally to find the best answer. As reasoning-model architecture matures, prices have fallen (o3 at $2/$8 per 1M is dramatically cheaper than o1-pro for many of the same workloads). Expect o1-pro's price to drop or the model to be deprecated in favor of higher-quality successors over 2026-2027.

Base case: a 2,000-token problem with 20,000 reasoning tokens + 1,000 visible output. `(0.002 × $150) + (0.021 × $600) = $0.30 + $12.60 = $12.90` per call. Compare with GPT-5 `reasoning_effort: high` on the same problem: `(0.002 × $1.25) + (0.020 × $10) = $0.0025 + $0.20 = $0.2025`. o1-pro is ~64× more expensive on identical token volumes.

When o1-pro pencils: a single high-stakes decision that one wrong answer would cost >$100 to fix downstream. Legal contract analysis where a missed clause costs $50K. Medical diagnosis decision support where a wrong call costs a treatment cycle. Algorithmic-trading strategy synthesis where a bad backtest costs the day's P&L.

When o1-pro does NOT pencil: anything that runs at production volume. Anything where GPT-5 `reasoning_effort: high` or o3 closes 90% of the quality gap at <2% of the price. Anything where the cost of error is bounded (a chat response, a content draft, a code suggestion).

Most teams who experiment with o1-pro discover that the marginal quality gain over GPT-5 + high reasoning effort or o3 is not worth the 30-60× price premium for their actual workloads. The honest use case for o1-pro is narrow and shrinking as cheaper reasoning models close the gap. Worked $: o1 reasoning cost calculator.

**Streaming is not supported.** o1-pro returns the response as a single block once internal reasoning completes. For a hard problem with 30K reasoning tokens, that's 60-180 seconds of wall-clock before the first byte arrives. Plan UX accordingly — a 'thinking…' placeholder visible to the user, async job submission patterns, or both.

**Prompt caching is not supported.** Unlike GPT-5 (where caching automatically activates on stable prefixes), o1-pro charges full input rate on every call regardless of prefix stability. There is no cost optimization via caching available.

**Temperature, top_p, presence_penalty, frequency_penalty are ignored.** Reasoning models on OpenAI ignore the sampling parameters that work on GPT-5 / GPT-4 — the model controls its own internal sampling strategy as part of reasoning. The API accepts the parameters for compatibility but does not apply them.

**No system messages in the traditional sense.** o1-pro uses 'developer' messages (a new role introduced for reasoning models) for instructions that previously went in `system`. The Responses API handles this distinction automatically; chat completions pre-o1 patterns need adaptation.

**Only the Responses API.** o1-pro is not exposed on `/v1/chat/completions`. Must use `/v1/responses`. Migration from GPT-4o chat code requires API surface changes, not just a model-ID swap.

Reasoning tokens on o1-pro bill at the output rate ($600/M) but are not returned to you. They are visible only in the response's `usage.output_tokens_details.reasoning_tokens` field after the fact — you cannot see them mid-stream (since streaming isn't supported anyway).

Typical reasoning token counts on o1-pro per query class: simple Q&A 500-2,000 reasoning tokens, moderate analysis 3,000-10,000, hard reasoning 15,000-40,000, math/proof tasks can hit 50K+. The variance is high — the same prompt run twice can produce 5K vs 25K reasoning tokens depending on how the internal search converges.

Cap with `max_output_tokens` to prevent runaway cost on the long tail. Setting `max_output_tokens: 20000` caps the combined reasoning + visible output at 20K tokens. The model truncates reasoning rather than producing a malformed visible answer. For production cost predictability, treat max_output_tokens as a hard guardrail.

**Pick o1-pro** only when: a single decision is worth >$100 of cost-of-error, AND you've benchmarked it against o3 and GPT-5 `reasoning_effort: high` and o1-pro wins on the eval, AND the wall-clock latency (60-180 seconds) is acceptable for the use case. This is a narrow band; most teams discover they don't need o1-pro after running the comparison.

**Pick o3** ($2/$8 per 1M) for production reasoning at scale. o3 closes most of o1-pro's quality gap at ~1% of the price and supports streaming, prompt caching, and the full standard API surface. See o3 spec sheet for the side-by-side.

**Pick GPT-5 with `reasoning_effort: high`** when: you want reasoning depth in a model that also supports streaming, prompt caching, and the unified GPT-5 feature set. ~1% of o1-pro's price for typically 80-95% of o1-pro's quality on most tasks.

Honest assessment: o1-pro's economic case is shrinking. As of June 2026, it is the right call for fewer workloads than it was at launch in December 2024. Re-evaluate quarterly against the latest reasoning-model menu.

o1-pro supports text + image input. Pass images as URLs or base64-encoded data inside a user (or developer) message's content array. Vision reasoning on o1-pro is among the strongest of any production model — multi-image reasoning, chart analysis, diagram interpretation, complex visual logic puzzles. The price premium that doesn't pencil for text-only tasks is more defensible for vision-reasoning tasks where the cost-of-error is real.

Function calling is supported via the Responses API's standard tools mechanism. Parallel tool calls are supported. Structured outputs via the `text.format` parameter (JSON Schema enforcement) work as on GPT-5.

The Responses API supports stateful conversations (`previous_response_id` parameter) so you can chain o1-pro calls without re-sending the full conversation history. Given o1-pro's input price ($150/M), this is meaningful — replaying a 5,000-token conversation history bills $0.75 per turn at input alone. Use stateful mode to avoid that.

Every number on this page was verified against OpenAI's live documentation on 2026-06-20. Sources: platform.openai.com/docs/models/o1-pro for context, modalities, parameter support; openai.com/api/pricing for input/output prices.

o1-pro pricing has been stable since launch in December 2024. Expect downward revisions or a successor model in 2026-2027 — the unit economics of pro-tier reasoning will improve as the architecture matures, and OpenAI's pricing typically follows compute-cost reductions.

Methodology: when a number could not be cross-confirmed against an official OpenAI page on the verification date, it was omitted from this card rather than guessed.