OpenAI has 8 tracked models in this catalog: 8 with verified first-party prices and 8 with an available OpenRouter route. Use the model pages below for a focused answer and source history.
What are the current OpenAI API prices?
This table sorts tracked OpenAI models by a consistent example workload when a verified first-party price is available. The example uses 1M input tokens and 250K output tokens; it is a comparison aid, not a promise that the models are equivalent.
| Model | Official in / out | OpenRouter in / out | Context | Example official | Example routed | Sources |
|---|---|---|---|---|---|---|
| GPT-5.4 miniOfficial API price verified | In $0.75Out $4.50 | In $0.75Out $4.50 | 400K | $1.875 | $1.875 | |
| GPT-5.6 LunaOfficial API price verified | In $1.00Out $6.00 | In $0.50Out $3.00 | 1.05M | $2.50 | $1.25 | |
| GPT-4.1Official API price verified | In $2.00Out $8.00 | In $2.00Out $8.00 | 1.048M | $4.00 | $4.00 | |
| GPT-5.3 CodexOfficial API price verified | In $1.75Out $14.00 | In $1.75Out $14.00 | 400K | $5.25 | $5.25 | |
| GPT-5.2Official API price verified | In $1.75Out $14.00 | In $1.75Out $14.00 | 400K | $5.25 | $5.25 | |
| GPT-5.6 TerraOfficial API price verified | In $2.50Out $15.00 | In $1.25Out $7.50 | 1.05M | $6.25 | $3.125 | |
| GPT-5.4Official API price verified | In $2.50Out $15.00 | In $2.50Out $15.00 | 1.05M | $6.25 | $6.25 | |
| GPT-5.6 SolOfficial API price verified | In $5.00Out $30.00 | In $5.00Out $30.00 | 1.05M | $12.50 | $12.50 |
The first five model pages are also linked here for quick access: GPT-5.4 mini · GPT-5.6 Luna · GPT-4.1 · GPT-5.3 Codex · GPT-5.2. All prices should be confirmed on the linked source before budgeting.
How should you compare OpenAI models?
Start with the input and output mix of your actual workload, then account for cached input, context-window tiers, tools, retries, and expected monthly runs. A lower output price can still produce a higher monthly bill if the model needs longer responses or more retries.
Use the workload calculator to change token volume, cached-input percentage, and monthly runs. The calculator keeps the official direct route separate from OpenRouter instead of treating a routed price as a first-party quote.
What is the difference between official and OpenRouter OpenAI pricing?
Official pricing refers to a provider-owned API or documentation source. OpenRouter pricing refers to a separate routing and billing channel with its own model ID, availability, verification time, and account terms. The two prices may match, differ, or exist on only one side.
For OpenAI, a missing official hosted price is shown as “No first-party API” rather than being replaced with an OpenRouter number. That distinction is especially important for open-weight or provider-specific routes.
Where can you find each model's source?
Every row links to the provider source used for the official channel and, when available, the exact OpenRouter model record used for the routed channel. The model detail pages include the latest verification date and the number of recorded price-history events.
Prices are checked by the scheduled updater, but a blocked or suspicious source keeps the last verified value instead of silently publishing an unverified change. Read the pricing methodology for the full verification rules.
OpenAI API pricing questions
What is the cheapest OpenAI API model in this catalog?
GPT-5.4 mini has the lowest verified OpenAI example workload cost in this catalog at $1.875 for 1M input tokens and 250K output tokens. A different input-to-output ratio can change the ranking.
Does OpenAI pricing match OpenRouter pricing?
Not necessarily. The official and OpenRouter channels are stored and checked separately for each model. Compare the two columns and follow each source link before making a production decision.
How are OpenAI token prices displayed?
Published standard text-token rates are normalized to USD per one million tokens. Input, cached-input, and output rates remain separate when the source provides them.
Are these OpenAI models interchangeable?
No. Price is only one dimension. Context limits, quality, latency, tools, availability, safety behavior, and provider terms still need to be tested for the intended workload.