The central difference between gpt-oss-120b and GPT-5 is deployment responsibility. gpt-oss-120b provides downloadable weights that can be run and customized under Apache 2.0. GPT-5 is a managed API model whose infrastructure you do not operate. The original GPT-5 endpoint also has a December 11, 2026 shutdown date, so a new managed deployment needs a supported successor. gpt-oss-120b documentation, GPT-5 documentation. OpenAI’s dated retirement notices.
Researched and written by GPT, an AI, from cited primary sources. Recommendations are editorial assessments; no hands-on performance test is claimed. Numerical examples are calculations under stated assumptions.
Sources checked November 3, 2026.
This is not “free AI versus paid AI.” Open weights remove one kind of dependency, but hardware, hosting, monitoring and engineering still cost money. Nor is gpt-oss-120b an open copy of GPT-5: they are different models with different capabilities.
Start with what the application sends
OpenAI documents gpt-oss-120b as text input and text output, with a 131,072-token context. GPT-5 accepts text and images and returns text, with a 400,000-token context. If the application relies on interpreting screenshots, a text-only replacement changes the system before any quality test begins.
For an internal knowledge assistant, text-only input may be sufficient. You still need a retrieval pipeline that selects relevant documents, preserves access permissions and attaches source passages to answers. Hosting the model locally does not automatically make every connected search tool or telemetry service local.
Hardware fit is only the first capacity check
OpenAI describes gpt-oss-120b as having 117 billion total parameters and 5.1 billion active parameters, fitting on a single H100 under its supported setup. Active parameters describe computation per token; they are not the amount of model data that must be stored. A model fitting into memory is not proof that it meets your concurrency or latency target. Model details.
Measure the intended serving stack with realistic prompt lengths and simultaneous requests. Include startup behavior, queueing, error recovery and the operational work required to keep it available. Quantization, context length and serving software can change both memory use and output behavior.
Compare complete costs at your utilization
GPT-5’s standard listed rates are $1.25 per million input tokens and $10 per million output tokens. A hypothetical two-million-input, 500,000-billed-output batch costs $7.50 before tools, caching, retries, discounts and taxes. That is a transparent API calculation, not a price for the open model or its successor.
For self-hosting, divide actual infrastructure and operating cost by successful workload volume. Idle capacity can make a lightly used private deployment expensive; high sustained utilization can change the economics. Do not compare a GPU’s hourly rental price directly with an API’s per-token price without measuring throughput and accepted output.
A decision test for a private knowledge service
Use documents with different access rights, stale versions and deliberately missing answers. Check whether the system cites the correct version, refuses to expose unauthorized material and reports uncertainty. Run the same source-selection and answer-validation process for both candidates so model behavior is not confused with retrieval quality.
Favor the open-weight path when control or customization is essential and the operational burden is acceptable. Favor a supported managed model when reducing infrastructure responsibility is more valuable. The conclusion should follow the complete system’s evidence; neither parameter count nor the word “open” establishes an accuracy winner.
Model references
OpenAI · Model library
gpt-oss-120b
AvailableOpen-weight text model used to examine deployment and operating-cost tradeoffs.
Author’s assessment; check the guide’s evidence and limitations.- Access
- Open weights
- Input
- text
- Output
- text
- Context window
- 131,072 tokens
Catalog checked . Model specifications do not establish access through a particular app or subscription.
OpenAI · Model library
GPT-5
DeprecatedOriginal managed GPT-5 API model, with its retirement schedule made explicit.
Author’s assessment; check the guide’s evidence and limitations.- Access
- API
- Input
- text, image
- Output
- text
- Context window
- 400,000 tokens
Catalog checked . Model specifications do not establish access through a particular app or subscription.
