Is GPT-OSS-120B free?
The model weights are available under Apache 2.0, but that does not make hosted inference free. OpenAI says GPT-OSS weights are not served through ChatGPT or the OpenAI API. You can download the weights and run them on infrastructure you control, or use a hosting provider; in either case, compute and hosting may cost money. (help.openai.com)
A third-party listing, BenchGecko, reported free input and output prices and “OpenAI Free” availability as of July 11, 2026. That listing does not establish API access or unlimited use, and it differs from OpenAI’s current documentation, which says the weights are not served through ChatGPT or its API. Treat the listing as a dated third-party claim—not confirmation that you can use GPT-OSS-120B for free in ChatGPT or through OpenAI’s API. (help.openai.com)
| Access route | What “free” means | Budget implication |
|---|---|---|
| Download and run the weights | OpenAI describes GPT-OSS-120B as open-weight under Apache 2.0 and links to the model download. | The weights are available to download; you still need suitable compute, which may have costs. (developers.openai.com) |
| ChatGPT or OpenAI API | OpenAI says GPT-OSS weights are not served through ChatGPT or the OpenAI API. | Don’t budget on using GPT-OSS-120B through those routes. (help.openai.com) |
| Third-party hosted API | A provider runs the model and charges according to its own terms. | Check that provider’s rate, access limits, and deployment details. Listed rates below are examples, not OpenAI prices. |
Open weights are not free API calls
“Open-weight” describes access to model weights; it is not a promise of free hosted inference. OpenAI’s model page describes GPT-OSS-120B as available under Apache 2.0 and offers a download. OpenAI’s Help Center also says the weights are not served through ChatGPT or the OpenAI API. (developers.openai.com)
So the practical choices are to run the model yourself—paying for the infrastructure you use—or choose a hosting provider and pay any applicable inference charges. A provider’s free offer, if available, would be specific to that provider and subject to its terms; the captured listings do not establish a universal free quota.
Third-party hosted API price examples
The following are provider-listed prices, not OpenAI API prices. The available listings do not provide effective dates for these rates, so confirm the selected model and route’s price before committing a production budget.
| Provider and route | Input | Output | Other listed rates |
|---|---|---|---|
| Novita serverless | $0.05 per 1 million tokens | $0.25 per 1 million tokens | No on-demand deployment price was provided in the captured listing. |
| ShareAI | €0.13524 per 1 million tokens | €0.540958 per 1 million tokens | Cache read and cache write: €0.13524 per 1 million tokens each. ShareAI says rates may depend on the selected tag, pricing plan, and token range; it says cached tokens are counted once. |
Example budget using Novita’s listed serverless rates: Suppose a workload uses 10 million input tokens and 2 million output tokens in a billing period:
- Input: 10 × $0.05 = $0.50
- Output: 2 × $0.25 = $0.50
- Estimated total: $1.00
This is arithmetic from the listed rates, not a tested bill or a guarantee of the amount charged. It excludes any fees or credits not specified in that listing. Actual cost depends on billed token counts and the provider’s applicable terms.
Production checks before choosing a route
Before sending production traffic, confirm these details for the exact provider, model tag, and deployment you plan to use:
- Current price and billing terms: Check input and output rates, cache treatment, any minimums or additional charges, and whether the account has credits or a free allowance.
- Access and limits: Confirm your account can use the model, along with applicable quotas, rate limits, and context and output limits.
- Endpoint and model identifier: Verify the current endpoint and model ID in the provider’s live documentation or console. A code example alone does not prove that a route is enabled for your account.
- Deployment cost: If considering dedicated GPUs rather than serverless billing, get the deployment price and billing unit directly. Novita’s captured page describes dedicated-GPU on-demand deployments but does not give their price.
- Workload fit: Compare the confirmed limits with your prompt sizes and expected output lengths. Check provider terms relevant to your application before launch.
The captured provider listings show why route-specific checking matters. Novita lists a 128K-token context window and 32K maximum output. ShareAI lists 131,072 tokens for both context and maximum output, but cautions that deployment limits can be lower. These are different provider-page figures, not one universal GPT-OSS-120B limit. ShareAI’s page also explicitly says its displayed endpoint for this tag has not been confirmed; verify it in the API reference or console before using it.
Bottom line: GPT-OSS-120B’s open weights can be downloaded, but OpenAI says it does not host them through ChatGPT or its API. For hosted use, budget against the specific provider’s current terms—not the word “free” in a third-party listing or the model’s open-weight status.
Sources
OpenAI open-weight models (gpt-oss) — OpenAI Help Center
gpt-oss-120b Model — OpenAI API
gpt-oss-120b (free) · Benchmarks and Pricing — BenchGecko