EXPLAINER

OpenAI Reasoning Models: o-series, Effort, Reasoning Tokens and Deprecations

modelfreshness: volatilekey claims checked 2026-10-09

OpenAI's reasoning models think before answering. The API bills hidden reasoning tokens as output, lets you set a reasoning effort and request a summary (never the raw reasoning), and older o-series models are being retired on published dates.

Short answer: OpenAI's reasoning models (the o-series from o1, then o3/o4-mini, later folded into the GPT-5 line) spend extra reasoning tokens before answering. Those tokens are billed as output but not exposed; you control cost with reasoning effort, can ask for a summary, and must budget max_output_tokens so thinking does not eat the answer. Several o-series models have published shutdown dates.

What is it?

A reasoning model works through the problem internally before it replies. You pay for that hidden work (as "reasoning tokens"), you can ask it to think less or more (the effort setting), and you can ask for a short summary of its thinking, but you never receive the raw thinking itself. Because model names and retirement dates change quickly, the safe habit is to read OpenAI's own model and deprecation pages on the day you build rather than trusting a blog post.

Also known as: OpenAI o1, OpenAI o3, o4-mini, o-series models, OpenAI reasoning models, reasoning tokens, reasoning effort, reasoning summary.

Technical definition

Terminology from OpenAI's documentation, checked on the date below: "reasoning tokens" are generated in addition to input and output tokens and can run from a few hundred to tens of thousands per request; the exact count appears in the response usage object (output_tokens_details.reasoning_tokens); they are billed as output tokens but the raw reasoning tokens are not exposed through the API. If generation reaches the context limit or max_output_tokens, the response can be "incomplete" with reason max_output_tokens, possibly before any visible output, so you can pay for input and reasoning without getting an answer. The reasoning guide lists effort values none, minimal, low, medium, high, xhigh and max, with model-specific support (some models do not accept none or minimal and have different defaults). A summary parameter can return a concise, detailed or auto-selected summary of the reasoning. On supported models, persisted reasoning items let a model use earlier-turn reasoning in later turns without exposing the raw text, and switching model families drops incompatible reasoning. The guide is written for the Responses API. History as documented by OpenAI and its search-visible announcements: o1-preview and o1-mini (September 2024); o3 and o4-mini (April 2025); the GPT-5 family in the API (August 2025), whose system card compares gpt-5-thinking variants against o3 and o4-mini. Deprecations page, as read: o1-preview (shutdown 2025-07-28) and o1-mini (2025-10-27) were announced on 2025-04-28; o1, o1-pro, o3-mini and o4-mini are listed with shutdown 2026-10-23; o3, o3-pro and the dated GPT-5 snapshots with shutdown 2026-12-11; the o3-deep-research and o4-mini-deep-research models were shut down 2026-07-23.

Entity type: model · Freshness class: volatile.

When should I use it?

Use reasoning models for multi-step problems where accuracy matters more than latency and cost; use faster non-reasoning models for simple tasks.

Implementation patterns

  • Start at a lower effort and raise it only if an evaluation shows gains.
  • Set a generous max_output_tokens and handle the "incomplete" status explicitly.
  • Log reasoning-token counts per request to track cost.
  • Ask for a reasoning summary only if you need to show users why — not as a source of truth.
  • Pin to a dated snapshot and track its retirement date; plan replacement testing before it.

Technologies and frameworks

Commonly associated: Responses API, reasoning.effort, reasoning summaries, usage.output_tokens_details.reasoning_tokens, max_output_tokens, persisted reasoning items. Names are examples of what exists, not recommendations, and projects change quickly.

Practical guide

  1. Open the current OpenAI model list and deprecations page and choose a dated snapshot.
  2. Run your evaluation set at two effort levels and compare quality and cost.
  3. Set output limits that leave room for reasoning plus the answer.
  4. Handle incomplete responses and log usage details.
  5. Schedule a re-evaluation before any announced shutdown date.

Common failure modes

  • Setting max_output_tokens too low and getting no visible answer while still paying for reasoning.
  • Using high effort for simple extraction tasks.
  • Treating a reasoning summary as a faithful transcript of the model's internal process.
  • Copying a model name from a blog post that has since been retired.
  • Assuming prompts tuned for a non-reasoning model carry over (long step-by-step instructions can be unnecessary).

Troubleshooting

  • Empty or truncated answer — Check status and incomplete_details; raise max_output_tokens or lower effort.
  • Costs jumped after switching models — Compare reasoning_tokens in usage; lower effort or switch to a non-reasoning model for easy tasks.
  • Model not found / deprecated error — Check the deprecations page for the shutdown date and the listed replacement; re-run your evaluations on the replacement.

Contrasted with

Questions this page answers

  • What are OpenAI reasoning models?
  • What are reasoning tokens and how are they billed?
  • What does reasoning effort do in the OpenAI API?
  • Can I see the raw chain of thought from an OpenAI reasoning model?
  • When are o1, o3 and o4-mini being shut down?
  • Why did my reasoning model return an empty answer?

Typical goals: understand what reasoning tokens and effort mean for cost; know which OpenAI reasoning models are being retired; use reasoning models without losing the answer to the token limit.

Sources and verification

Status: key claims checked on 2026-10-09 against the primary sources listed below; everything else is not independently verified. The claims under "Checked against primary sources" were compared with the linked pages on that date. Pages change, so re-check anything time-sensitive before relying on it.

Checked against primary sources (2026-10-09):

  • Reasoning tokens are in addition to input and output tokens, counted in usage.output_tokens_details.reasoning_tokens, billed as output tokens, and raw reasoning tokens are not exposed via the API.
  • Incomplete responses can occur with reason max_output_tokens before any visible output.
  • Effort values listed in the guide: none, minimal, low, medium, high, xhigh, max, with model-specific limits; summary parameter values concise, detailed, auto.
  • Persisted reasoning does not expose raw reasoning; switching model families omits incompatible reasoning.
  • Deprecation rows: o1-preview (2025-07-28) and o1-mini (2025-10-27) announced 2025-04-28; o1, o1-pro, o3-mini, o4-mini shutdown 2026-10-23; o3, o3-pro and dated GPT-5 snapshots shutdown 2026-12-11; o3-deep-research and o4-mini-deep-research shut down 2026-07-23.
  • OpenAI's "Introducing OpenAI o3 and o4-mini" page is dated April 16, 2025, and its o1-preview announcement page is dated September 12, 2024.

Primary sources consulted:

Not independently verified:

  • Current model names: the guide and the deprecations page named different "current" models on the same day, so no current model is named here.
  • Claims that o3 makes fewer errors than o1 or similar performance comparisons.
  • Pricing.
  • Whether the o-series retirement dates will hold; OpenAI revises schedules.

Freshness class: volatile — Products, versions, protocols, benchmarks, leaderboards, regulations and dates. Verify against the primary source before relying on specifics; re-review every 3 months.

Time-sensitive claims to verify before relying on them:

  • Which models exist, their names and their effort support (the reasoning guide and the deprecations page used different current-model names on the check date).
  • Shutdown dates, which can be revised.
  • Pricing and billing rules.
  • Parameter names and allowed effort values.
  • Whether the o-series naming is still used for new models.

Further reading and canonical sources (not all checked):

  • OpenAI reasoning models guide — developers.openai.com/api/docs/guides/reasoning
  • OpenAI API deprecations — developers.openai.com/api/docs/deprecations
  • Introducing OpenAI o3 and o4-mini — openai.com/index/introducing-o3-and-o4-mini/
  • Introducing OpenAI o1-preview — openai.com/index/introducing-openai-o1-preview/
  • GPT-5 system card (August 2025) — cdn.openai.com/gpt-5-system-card.pdf