OpenAI o3 is a reasoning model designed for multi-step work involving analysis across text, code, and images. But if you’re choosing a model now, its status matters more than its historical benchmark scores: OpenAI marks o3 as deprecated in the API and says it is no longer available in ChatGPT. OpenAI’s o3 documentation names GPT-5 as its successor; its current reasoning guidance recommends starting with gpt-6-astra for most reasoning workloads and considering lower-cost models for other needs. Match the candidate to your workload and verify account availability before migrating. (o3 model documentation; o3 and o4-mini launch announcement; reasoning guidance; model catalog)
What is the ChatGPT o3 model?
OpenAI describes o3 as a reasoning model for complex tasks, including work that combines text, code, and images. Its API documentation lists text and image inputs, text output, a 200,000-token context window, and a maximum output of 100,000 tokens. These are documented specifications for o3—not a promise that the model is still available to new users or that every ChatGPT feature works through the API. (o3 model documentation)
When it launched, OpenAI highlighted math, science, coding, visual reasoning, technical writing, and instruction-following. Those descriptions help explain what o3 was built to do, but the launch post is historical; it now notes that o3 is no longer available in ChatGPT and is deprecated in the API. (o3 and o4-mini launch announcement)
Is o3 still available in ChatGPT or the API?
No longer in ChatGPT; deprecated in the API. OpenAI’s o3 model page marks the model as deprecated. Its launch announcement likewise says o3 is no longer available in ChatGPT and is deprecated in the API. (o3 model documentation; launch announcement)
For a new production integration, a deprecated model is not a sound starting point. OpenAI’s o3 page names GPT-5 as its successor, while its current reasoning guidance recommends gpt-6-astra as the starting point for most reasoning workloads. The current API model catalog also presents gpt-6-astra for complex reasoning and coding, and gpt-6-luna for cost-sensitive, high-volume workloads. Treat these as candidates to evaluate, not guaranteed drop-in replacements; confirm the model is available to your account and supports your specific workflow. (o3 model documentation; reasoning guidance; model catalog)
What did o3 cost through the API?
OpenAI’s o3 documentation lists standard API rates per 1 million tokens of $2 for input, $0.50 for cached input, and $8 for output. It separately lists Batch API rates of $1 for input, $0.25 for cached input, and $4 for output. These are API charges, not ChatGPT subscription prices. Since o3 is deprecated, treat these as documented o3 rates—not as current prices for a replacement or a reason to start a new integration. Check the current pricing page for the model you plan to use. (o3 model documentation; API pricing)
For standard o3 rates, estimate token charges as:
Estimated token charge = (uncached input tokens × $2 + cached input tokens × $0.50 + billable output tokens × $8) ÷ 1,000,000
Count cached input separately from uncached input so the same tokens are not charged twice in the estimate. For reasoning models, billable output includes internal reasoning tokens as well as the visible answer. (prompt caching guide; token counting guide)
For example, 100,000 uncached input tokens and 10,000 total billable output tokens would cost $0.28 at the listed standard rates: (100,000 × $2 + 10,000 × $8) ÷ 1,000,000. This illustration assumes no cached input, no tool charges, and that 10,000 is the total billable output-token count, including reasoning tokens. It is not a measured request or a price estimate for a current successor model.
How to choose a replacement for an o3 workload
OpenAI’s current guidance offers a practical shortlist for testing:
- Start with
gpt-6-astrafor complex reasoning work. OpenAI recommends it for most reasoning workloads and describes it as its flagship model for complex reasoning and coding. Its API documentation lists text and image input, text output, and structured outputs. Standard rates per 1 million tokens are $10 for input, $1 for cached input, $12.50 for 30-minute cache writes, $25 for 12-hour cache writes, and $50 for output. Evaluate it when the task’s quality requirements justify the higher listed token rates. (reasoning guidance; Astra model page; API pricing) - Evaluate
gpt-6-lunawhen cost or high-volume processing is the priority. OpenAI describes it as its most efficient model for focused, high-volume tasks. Its API documentation lists text and image input, text output, and structured outputs. Standard rates per 1 million tokens are $0.10 for input, $0.01 for cached input, $0.125 for 30-minute cache writes, $0.25 for 12-hour cache writes, and $0.50 for output. Test whether its quality and response time meet your task’s thresholds before routing production traffic to it. (model catalog; Luna model page; API pricing)
These are selected USD standard-processing API token rates, not ChatGPT subscription prices or a complete cost estimate for every workflow. Cache-write rates apply to tokens billed as cache writes under the applicable prompt-caching conditions; they are not an extra fee added on top of the ordinary input rate for those same tokens. Check the pricing and prompt-caching documentation for the model, processing mode, and caching setup you intend to use. (API pricing; prompt caching guide)
Both replacement model pages list text and image input with text output, and list structured outputs as supported. Confirm the specific output format and API feature behavior required by your application in the relevant model documentation. ChatGPT’s interface features do not by themselves establish API support. (Astra model page; Luna model page)
Don’t select a replacement solely because it has a high score on a benchmark or a similar model name. Compare quality, response time, throughput, and cost on the same representative tasks. A model’s position in OpenAI’s guidance is a starting point for evaluation, not proof that it will be the best fit for your application.
A practical evaluation:
- Define the task set. Gather real examples covering routine cases, difficult cases, and known failure modes. Handle sensitive data under your organization’s policies.
- Set decision thresholds before testing. Specify acceptable answer quality, maximum response time, required image or tool support, and acceptable cost per task.
- Run the same cases on each candidate. Keep prompts, inputs, and evaluation criteria consistent. For consequential tasks, have a qualified reviewer score correctness and error severity.
- Measure operational fit. Record response times and token use for your workload, and check the candidate’s documented limits, rate limits, and supported features.
- Estimate current costs. Use the selected model’s current API pricing and your observed input, cached-input, cache-write, and output usage. Account for cache writes only where your caching setup and billing mode apply them. Compare cost per successfully completed task—not just price per million tokens.
- Choose with a fallback plan. Decide how your service will respond to timeouts, rate limits, or model availability changes, and how you’ll evaluate future model updates.
These are recommended evaluation steps, not test results for o3 or any replacement. TechTarget’s June 2025 article discussed o3, o3-pro, o3-mini, and o4-mini as distinct options and reported historical trade-offs. Those descriptions predate the current deprecation notice; don’t treat them as a current availability or recommendation chart. (TechTarget, published June 13, 2025)
What to verify before using any reasoning model in production
Check the documentation for the exact model and API channel you intend to use:
- Availability and identifier: Is the model available for your account, and is its ID current rather than deprecated?
- Input and output support: Does it accept the text, images, or other inputs your application sends, and return the format you require?
- Tools and features: Are the specific API tools and features you need documented for that model? ChatGPT’s interface features do not by themselves establish API support.
- Limits: Check context and output limits, rate limits, and any account or tier conditions that affect your workload.
- Latency and failure handling: Measure response times on your own requests and design for delays or failures; don’t infer a service guarantee from a model’s reasoning label.
- Billing: Use current API prices and billing rules for the chosen model, including applicable cached-input and cache-write rates or other charges. Keep API usage separate from ChatGPT plan costs. (API pricing)
Bottom line: o3 was intended for complex reasoning across text and images, but OpenAI now marks it as deprecated in the API and unavailable in ChatGPT. For a complex reasoning workload, evaluate gpt-6-astra; if cost or high-volume processing dominates, evaluate gpt-6-luna. Confirm account access and feature fit, then compare candidates on your own quality, latency, capability, and cost requirements. (o3 model documentation; launch announcement; reasoning guidance; model catalog)
Sources
Introducing OpenAI o3 and o4-mini
Understanding and counting tokens
GPT-6 Astra Model | OpenAI API
OpenAI o3 and o4 explained: Everything you need to know | TechTarget