OpenAI o3 Guide: The Tool-Using Reasoner Before Its Retirement

o3 is the model a lot of production reasoning workloads were quietly built on in 2025, and it is now on a clock. OpenAI's deprecations page lists the o3-2025-04-16 snapshot for shutdown on 11 December 2026, with gpt-5.6-sol as the recommended replacement. ChatGPT already dropped it in August.

That makes this less a buying guide than an exit guide. If you still have o3 in a config file, the useful questions are what it was actually good at, which of its quirks your prompts depend on, and how to prove a replacement is not a regression before December.

What o3 changed in the o-series

OpenAI previewed o3 in December 2024 and released it on 16 April 2025, alongside o4-mini. The system card published that day describes the shift plainly: o3 uses tools inside its chain of thought. It can crop or transform an image, search the web or run Python while it is still reasoning, rather than only reasoning first and calling a tool afterwards.

That sounds like a detail and was not one. Earlier o-series models thought in a sealed box and then answered. o3 could look something up halfway through a proof, or zoom into a screenshot it had misread. For visual reasoning in particular this was the step change, because the model could inspect an image more than once instead of committing to a first glance.

The API specification is conventional for its generation: a 200K-token context window, up to 100K output tokens, a knowledge cutoff of 1 June 2024, text and image input, text output. It is available on Chat Completions, Responses and Batch. In the Responses API it supports function calling, file search, code interpreter, web search, image generation and remote MCP servers.

The numbers OpenAI published

The system card's headline SWE-bench Verified figure is 71%, but it belongs to a helpful-only o3 variant rather than the shipped model, measured on a fixed 477-task subset through an internal tool scaffold and averaged over four attempts per task. On OpenAI's internal pull-request replication eval, the o3 launch candidate scored 44% against 39% for o4-mini. Those are vendor numbers on vendor scaffolds, so treat them as an upper bound for what your harness will see.

The more instructive table is the hallucination one. On PersonQA, a set of factual questions about people, o3 was more accurate than o1 (0.59 against 0.47) but also hallucinated twice as often (0.33 against 0.16). OpenAI's explanation is that o3 makes more claims overall, so it gets more right and more wrong. On SimpleQA the gap was small: 0.49 accuracy and a 0.51 hallucination rate, against 0.47 and 0.44 for o1.

The practical reading: o3 is confident and prolific. For reasoning over material you give it, that is what you want. For recalling facts from its own weights, check its work, or give it web search and make it cite.

Price history: the 80% cut

o3 launched expensive. On 10 June 2025, the same day o3-pro arrived, OpenAI announced it was cutting the price of o3 by 80%, to $2 per million input tokens and $8 per million output. OpenAI's model page still lists those rates, with cached input at $0.50.

OpenRouter lists the same model at $2 input and $8 output at the time of writing, plus an openai/o3:batch variant at half that: $1 and $4. The batch entry is not a different model. It is the same o3 behind OpenAI's asynchronous Batch API, which trades a completion window measured in hours for the discount. If any o3 work you do is offline, such as nightly classification or report generation, the batch maths is worth running before you migrate, because the replacement's batch pricing may land differently.

Remember that reasoning tokens bill as output tokens. A request that returns a 300-token answer can easily spend several thousand on hidden reasoning, so the $8 output rate is the one that dominates the invoice.

The retirement dates, in order

• 11 June 2026 — OpenAI notifies developers that the o3 and o3-pro snapshots are deprecated.
• 26 August 2026 — o3 retired from ChatGPT. This did not affect the API.
• 11 December 2026 — API shutdown for o3-2025-04-16. Recommended replacement: gpt-5.6-sol.

Third-party hosts that serve o3 set their own schedules, but the upstream snapshot going away is the date to plan against. There is only one o3 snapshot, so there is no older or newer dated version to pin to as a stopgap.

Migrating an o3 workload before December

Swapping the model string is the easy part. The work is establishing that nothing you rely on silently changed. A sensible order of operations:

1. Inventory the behaviour, not just the calls. Note which prompts depend on o3's tool use mid-reasoning, image input, or structured outputs. Those are the paths most likely to behave differently.
2. Build a frozen eval set now. Take twenty to fifty real requests with known-good outputs while o3 still answers. After December you lose the baseline. Building a 20-task eval covers the setup.
3. Compare cost per completed task. A replacement with a different reasoning-token appetite can cost more or less than its list price suggests. Measure total output tokens, not just the rate.
4. Check parameter compatibility. o-series models reject temperature, top_p and max_tokens, and use max_completion_tokens on Chat Completions. Newer models may restore some of these or change defaults such as reasoning effort.
5. Run both in parallel for a week. Shadow traffic surfaces the failure modes an eval set misses.

If you already use the smaller sibling, the o4-mini guide covers when that was the better default, and the o3-pro guide covers the high-compute variant that retires on the same day.