HomeAI models
OpenAI: o3
OpenAI
•
o3
•
Released 
April 2025

OpenAI
o3

OpenAI's April 2025 text-and-image reasoning model, faster to first token than GPT-5 (high) but scheduled for shutdown December 2026.

Modality:
Text
Image
Video
PDF
model ID
openai/o3
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

111.43
tok/s

Intelligence Index *

20.2
/ 100

Context Window *

200000
tokens

Input price

12
Anytoken

Output price

48
Anytoken
o3: Multi-Step Reasoning Over Text and Images, Before Its December 2026 Shutdown o3 is OpenAI's April 2025 reasoning model, the top of the o-series until GPT-5 succeeded it. It accepts text and image input and returns text, with a 200,000-token context window and up to 100,000 output tokens. OpenAI positions it for math, science, coding and visual reasoning. It generates quickly for a reasoning model, which suits multi-step analysis, code review and chart interpretation. The o3-2025-04-16 snapshot shuts down on December 11, 2026, so new builds should plan migration. Access o3 via API on AnyAPI.ai

Performance

Reasoning Depth at Lower Latency Than GPT-5 (High)

o3 pairs multi-step reasoning with image understanding. OpenAI reported 69.1% on SWE-bench Verified and 88.9% on AIME 2025; both are vendor-run figures. Independently, Artificial Analysis measured 137.2 tokens per second and a 6.47-second time to first token, against 107.6 tokens per second and 66.01 seconds for GPT-5 (high). Reasoning models are often too slow to start answering for interactive use. o3 is more usable for analyst-facing tools where a long wait is unacceptable. Quality-critical batch jobs may still favor slower, higher-scoring models.

Benchmarks

o3 on Artificial Analysis: Fast Generation, Mid-Pack Intelligence

Artificial Analysis scores o3 at 20 (estimated) on its Intelligence Index, below the median of 26 for reasoning models in a similar price tier; GPT-5 (high) scores 23. Output speed is 135.8 tokens per second on OpenAI's API, well above that tier's median of 77.7. Speed depends on the host: OpenAI measured 152.2 tokens per second versus 80.3 on Azure. These are third-party measurements under Artificial Analysis test conditions, not OpenAI specifications. Newer index versions may shift scores.

Output Speed

*
111.43
tok/s

Intelligence Index

*
20.2
/ 100

MMLU *

Broad world knowledge and problem-solving
85
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
83
%

HLE *

Adherence to multi-step structured instructions.
20
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
81
%

Technical Specifications

What the model supports

o3 takes text and image input and returns text only, with a 200,000-token context window and a 100,000-token output cap. Reasoning tokens are not visible through the API but occupy the context window and are billed as output, so usable answer length on hard prompts is smaller than the cap suggests. The knowledge cutoff is June 1, 2024, which matters for recent libraries or events. The snapshot is deprecated, with a December 11, 2026 shutdown.
Verified Specifications — 
o3
*
Input modalities
Text
Image
Video
PDF
output modalities
Text
Context window
200000
 tokens
Maximum output tokens
100000
Reasoning
Yes
Knowledge cutoff
April 2025
Pricing (standard)
12
 AnyTokens in
 / 
48
 AnyTokens out

Limitations & Trade-offs

Where o3 falls short

1
Scheduled shutdown. OpenAI marks o3 deprecated, and the o3-2025-04-16 snapshot shuts down on December 11, 2026, with gpt-5.6-sol as the recommended replacement. That is roughly two months away. Teams starting new long-lived products should not build on o3. Existing integrations need a migration and regression-testing plan, since behavior on reasoning-heavy prompts will differ between models.
2
Factual recall hallucination. In OpenAI's system card, o3 scored 0.59 accuracy and a 0.33 hallucination rate on PersonQA, versus 0.16 for o1. OpenAI says o3 tends to make more claims overall, producing both more accurate and more inaccurate ones. Unsourced questions about people or obscure facts are risky. Pair o3 with retrieval or verification, or prefer a model with lower measured hallucination for grounded Q&A.
3
Reasoning overhead on latency and budget. Artificial Analysis measured a 6.47-second time to first token, far from instant, and reasoning tokens occupy context and are billed as output. Short chat turns, classification and high-volume extraction pay this overhead for little benefit. Use a non-reasoning or smaller model for those, and set output limits high enough that reasoning does not truncate the visible answer.
4
Superseded on capability and freshness. Artificial Analysis scores o3 at 20 on its current index, while it lists GPT-5.5 (Low) at 31. Input is limited to text and images, output is text only, and the knowledge cutoff is June 2024. For frontier-level agentic or coding quality, or knowledge of recent libraries, a newer OpenAI model is the better choice.

Best-Fit Workloads

Where this model earns its place

01

Chart and diagram reasoning


OpenAI says o3 performs especially strongly at visual tasks like analyzing images, charts, and graphics, and the API accepts image input. It fits financial report review, technical diagram interpretation and screenshot-based troubleshooting, where the answer needs reasoning over both pixels and text. Output is text only, so it cannot generate images. Visual accuracy claims here are vendor-reported.

02

Multi-step code debugging and review


OpenAI reported 69.1% on SWE-bench Verified, a vendor-run score on a fixed 477-task subset. The 100,000-token output cap allows large patches or file rewrites in one call, though reasoning tokens share that budget. It suits code review and bug triage tools. Newer models score higher on current coding benchmarks, so check the replacement for frontier coding agents.

03

Quantitative and scientific analysis


OpenAI reported 88.9% on AIME 2025 and 83.3% on GPQA Diamond without tools, supporting use in math-heavy and science-heavy analysis such as derivations, data interpretation and technical write-ups. Verify factual claims independently, since PersonQA results show elevated hallucination on open-domain recall. Treat the scores as vendor-reported rather than independently reproduced.

04

Migration baselining


OpenAI notes that snapshots let you lock in a specific version so behavior remains consistent. With the o3 snapshot retiring in December 2026, it works as a fixed reference when evaluating replacement models on your own prompts, covering reasoning quality, latency and output length. This is a time-limited use, not a long-term deployment target.

Pricing in anytokens via AnyAPI
Input
12
₳
Output
48
₳
Cache write
—
₳
Cache read
—
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access o3 via AnyAPI.ai

Access o3 through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate o3 without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access o3 and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test o3 against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use o3 from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use o3 for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

o3 has a 200,000-token context window and a 100,000-token maximum output. These are separate limits. Hidden reasoning tokens occupy the context window and are billed as output, so long prompts plus heavy reasoning can leave less room for the visible answer than the maximum suggests.

Yes. OpenAI lists o3 input as text and image, with text as the only output. It can analyze charts, diagrams and screenshots alongside text instructions, but it cannot return images, audio or video.

Yes. OpenAI marks o3 as deprecated and lists the o3-2025-04-16 snapshot for removal from the API on December 11, 2026, with gpt-5.6-sol as the recommended replacement. Existing integrations should be migrated and regression-tested before that date.

OpenAI reported 69.1% on SWE-bench Verified, 88.9% on AIME 2025 and 83.3% on GPQA Diamond; these are vendor figures. Independently, Artificial Analysis scores o3 at 20 (estimated) on its Intelligence Index, below GPT-5 (high) at 23, with output speed around 137 tokens per second.

On OpenAI's PersonQA evaluation, yes: o3 had a 0.33 hallucination rate versus 0.16 for o1, though o3's accuracy was higher at 0.59. OpenAI attributes this to o3 making more claims overall. Use retrieval or verification for fact-heavy tasks.

* Benchmark data source: Artificial Analysis artificialanalysis.ai