# o3 ![Organization](https://img.shields.io/badge/Org-OpenAI-412991) ![Context](https://img.shields.io/badge/Context-200K%20tokens-1f6feb) ![Pricing](https://img.shields.io/badge/Price-%2410%20%2F%20%2440%20per%201M-2da44e) ![License](https://img.shields.io/badge/License-Proprietary-d29922) ![Modality](https://img.shields.io/badge/Modality-Text%20%2B%20Vision-8957e5) ![Released](https://img.shields.io/badge/Released-Apr%202025-555) > **Quick answer:** o3 is OpenAI's flagship reasoning model, released April 16, 2025. It uses chain-of-thought reasoning with configurable effort levels (low/medium/high) to trade latency for accuracy. It scores 88% on ARC-AGI, 75.8% on LiveCodeBench (high), and 85.9% on Arena Hard v2. Priced at $10 input / $40 output per 1M tokens with a 200K-token context window. ## At a Glance **Where o3 leads** - 88% ARC-AGI — strong abstract visual reasoning - 85.9% Arena Hard v2 — top-tier instruction following and general capability - 75.8% LiveCodeBench (high effort) — excellent competitive programming performance - 20.32% Humanity's Last Exam — frontier-class expert knowledge - Configurable reasoning effort: spend more compute for harder tasks - Full multimodal vision support **Where it lags** - High cost at $10/$40 per 1M tokens — 9× o4-mini's price - 200K context window (vs 1M+ for Gemini 2.5 Pro) - Closed API model with no open weights - High-effort setting significantly increases latency and token cost **Best for:** Frontier-difficulty reasoning, scientific analysis, competitive programming, and tasks where accuracy matters more than cost or speed. ## What o3 Is o3 is OpenAI's main reasoning model, part of the "o-series" that applies chain-of-thought reasoning before producing a final answer. Unlike the GPT series, o-series models trade raw speed for accuracy: they internally generate reasoning tokens before responding, with the depth of reasoning controlled by the effort level setting (low, medium, or high). The model represents a significant shift from GPT-4o: rather than optimising for fast, fluent generation, o3 prioritises correctness on hard multi-step problems. This makes it substantially stronger than GPT-4o on tasks involving mathematics, science, and complex coding, while being slower and more expensive per response. o3 was released alongside o4-mini — a compact, cost-efficient reasoning variant. The two models share the same reasoning paradigm but differ significantly in capability ceiling and price. o3 is the appropriate choice when task difficulty is high and the cost premium is justified by the outcome quality. ## Specifications | Field | Value | |---|---| | Organization | OpenAI | | Parameters | Undisclosed | | Context window | 200,000 tokens | | Max output | 100,000 tokens | | Architecture | Chain-of-thought reasoning (o-series) | | License | Proprietary (API only) | | Release date | April 16, 2025 | | Knowledge cutoff | June 2024 (estimated) | | Modality | Text + Vision (multimodal) | ## Pricing | | Input (per 1M tokens) | Output (per 1M tokens) | |---|---|---| | OpenAI API | $10.00 | $40.00 | Reasoning tokens (internal chain-of-thought) are billed as output tokens. Cached input receives a 75% discount. Pricing per [OpenAI pricing page](https://openai.com/api/pricing/) — verify current rates before production use. ## Context Window o3 has a **200,000-token context window** — roughly **150 pages** of text in a single request. This supports large codebases, long documents, and multi-turn conversations but is substantially smaller than long-context models like Gemini 2.5 Pro (1M tokens). ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [ARC-AGI](/benchmarks/arc-agi) | 88.0% | Benchgen evaluation | 2025-07 | | [Arena Hard v2](/benchmarks/arena-hard-v2) | 85.9% | Benchgen evaluation | 2025-07 | | [LiveCodeBench](/benchmarks/livecodebench) | 75.8% (high) | Benchgen evaluation | 2025-07 | | [Humanity's Last Exam](/benchmarks/humanitys-last-exam) | 20.32% | Benchgen evaluation | 2025-07 | | [SHADE-Arena](/benchmarks/shade-arena) | 1.6 overall success | [Anthropic research post](https://www.anthropic.com/research/shade-arena-sabotage-monitoring) | 2025-06 | *Scores above are from Benchgen evaluations or attributed third-party sources. Figures depend on harness, effort setting, and tools — see the source for methodology.* ## o3 vs Alternatives | Model | Context | ARC-AGI | LiveCodeBench | Price (in/out per 1M) | |---|---|---|---|---| | **o3** | 200K | 88.0% | 75.8% | $10 / $40 | | [o4-mini](/models/openai/o4-mini) | 200K | — | 74.2% | $1.10 / $4.40 | | [GPT-5.6 Sol](/models/openai/gpt-5-6-sol) | 1.1M | 97.5% | — | $5 / $30 | | [Gemini 2.5 Pro](/models/google/gemini-2-5-pro) | 1M | — | 73.6% | $1.25 / $10 | | [DeepSeek R1](/models/deepseek/r1) | 128K | — | — | ~$0.55 / $2.19 | o3's strongest case over o4-mini: higher ARC-AGI score and a larger reasoning ceiling on frontier-difficulty tasks. o4-mini closes the gap on LiveCodeBench (74.2% vs 75.8%) at 9× lower cost, making it the default for most coding tasks. ## How o3 Performs on Real Agent Tasks o3's SHADE-Arena score of 1.6 overall success (a safety evaluation for sabotage detection) reflects the benchmark's design — low scores indicate the model doesn't engage in sabotage behaviour, which is the expected result for a well-aligned model. Its Arena Hard v2 score of 85.9% and Humanity's Last Exam score of 20.32% position o3 near the frontier of models available at its April 2025 launch. For agent builders, o3's high-effort mode is the primary differentiator: on tasks where a single correct answer has outsized value (debugging a subtle race condition, solving a novel algorithm problem, reviewing a complex legal clause), the reasoning depth of o3-high can outperform faster models despite the cost premium. ## Use o3 via API ```python from openai import OpenAI client = OpenAI(api_key="YOUR_API_KEY") response = client.chat.completions.create( model="o3", messages=[{"role": "user", "content": "Solve this algorithmic problem step by step..."}], reasoning_effort="high", ) print(response.choices[0].message.content) ``` ## Related Models & Benchmarks - Compare: [o4-mini](/models/openai/o4-mini) · [GPT-5.6 Sol](/models/openai/gpt-5-6-sol) · [Gemini 2.5 Pro](/models/google/gemini-2-5-pro) - Benchmarks: [ARC-AGI](/benchmarks/arc-agi) · [LiveCodeBench](/benchmarks/livecodebench) · [Arena Hard v2](/benchmarks/arena-hard-v2) · [Humanity's Last Exam](/benchmarks/humanitys-last-exam) ## Frequently Asked Questions

What is o3?

o3 is OpenAI's flagship chain-of-thought reasoning model, released April 16, 2025. It applies internal reasoning before producing a final answer, with configurable effort levels (low/medium/high) that trade latency for accuracy on hard multi-step tasks.

What is o3's context window?

o3 supports a 200,000-token context window — roughly 150 pages of text — with a maximum output of 100,000 tokens.

How much does o3 cost?

o3 is priced at $10.00 per 1M input tokens and $40.00 per 1M output tokens. Internal reasoning tokens are billed as output tokens. Cached inputs receive a 75% discount.

Is o3 open source?

No. o3 is a proprietary, API-only model from OpenAI. There are no open weights.

What is the difference between o3 and o4-mini?

o3 is the higher-capability, higher-cost reasoning model ($10/$40 per 1M). o4-mini is the compact, cost-efficient variant ($1.10/$4.40 per 1M). o3 outperforms o4-mini on the hardest reasoning tasks; o4-mini is competitive on coding benchmarks at a fraction of the cost.

What is o3's knowledge cutoff?

o3's training knowledge cutoff is estimated at June 2024.

--- *Specs and scores sourced from OpenAI's official o3 announcement (April 16, 2025) and Benchgen evaluations; third-party benchmark scores attributed inline. Pricing cited to the OpenAI pricing page. Last updated 2026-07-23.*

o3 — Training jobs

Training jobs