What is babbage-002?
babbage-002 is a lightweight OpenAI GPT-3 base language model for completing text. Given a prompt, it predicts a continuation that may be natural language or code. OpenAI introduced it on August 22, 2023, as a replacement for the original GPT-3 ada and babbage models.
The word base is important. Unlike an instruction-tuned chat model, babbage-002 was not trained to reliably interpret a request such as “summarize this” or “write a formal email.” Developers generally need to design prompts carefully, use delimiters and stop sequences, or apply fine-tuning to make a particular completion workflow predictable. It therefore belongs to an older generation of prompt-completion systems rather than the modern chat and agent model category.
OpenAI provides babbage-002 through its legacy Completions API. Its primary role is maintaining applications that already depend on completion-style behavior, not serving as a general replacement for newer OpenAI models.
Capabilities and technical profile
babbage-002 accepts text and produces text. The text output can include code, so it may be useful for straightforward continuation or generation tasks where the application controls the prompt format. It does not accept images, audio, or video, and it does not directly generate non-text media.
| Specification | Research-supported detail |
|---|---|
| Provider | OpenAI |
| Model family | GPT-3 |
| Model type | Lightweight base language model |
| Input | Text only |
| Output | Text only |
| Knowledge cutoff | September 1, 2021 |
| Maximum documented output | 16,384 tokens |
| Context window | Not separately documented in the supplied model information |
| Primary API | Legacy Completions API |
The September 1, 2021 knowledge cutoff means the base model should not be treated as a source of current facts. The model does not include first-party web search or another documented retrieval feature that would automatically ground its answers in newer information. Applications needing current information would have to provide that information through their own workflow, although the supplied documentation does not describe a modern tool-grounding interface for babbage-002.
API features and limitations
babbage-002 supports a comparatively narrow set of API capabilities. The supplied model documentation does not list streaming, function calling, structured outputs, or predicted outputs as supported. It is also not documented as supporting modern action-taking or tool-use features.
These limitations affect how an application must be designed. A current chat or agent system can often ask a model to call a function, return a constrained object, or stream partial output to a user. With babbage-002, developers should instead treat the result as ordinary generated text and implement any parsing, validation, control flow, or external actions outside the model.
OpenAI’s current information lists batch access for the model. Fine-tuning requires additional caution: new fine-tuning training runs for babbage-002 have not been supported since October 28, 2024, while existing compatible fine-tuned models were handled separately under the deprecation policy. Consequently, “fine-tuning” should not be interpreted as an available route for starting a new training project on this model.
Pricing, speed, and capability trade-offs
The listed base-model price is $0.40 per 1 million input tokens and $0.40 per 1 million output tokens. This low token price is one of babbage-002’s clearest practical advantages for an existing workload that already fits its completion behavior.
The supplied editorial assessment rates the model’s speed at 7 out of 10 and its cost at 8 out of 10. Those are evaluation scores, not scores published by OpenAI. They indicate the expected trade-off: babbage-002 is relatively inexpensive and lightweight, but its lower price and speed profile come with substantially fewer capabilities than newer instruction-following models. The research also rates its reasoning at 2 out of 10 and coding at 3 out of 10; these are editorial assessments rather than standardized benchmark results or provider claims.
Cost should therefore be evaluated against the whole application rather than token price alone. A cheap model that needs elaborate prompting, custom output parsing, or a separate control layer may be less suitable than a more capable model for a new system. Conversely, migrating a stable legacy completion pipeline may create more engineering work than continuing to use babbage-002 during its remaining availability period.
Status and sunset date
OpenAI marks babbage-002 as deprecated and has announced that access to the base model will shut down on September 28, 2026. The supplied research identifies gpt-5.6-terra as the recommended replacement, but it does not provide a detailed capability or pricing comparison between the two models.
The shutdown date makes lifecycle risk the central issue for anyone evaluating babbage-002 today. It may still be relevant for a short-term maintenance task or for reproducing an established behavior, but it is a poor foundation for a new production service expected to operate beyond the announced retirement date. Teams using it should identify a migration target, test prompt and output differences, and avoid treating current availability as a long-term guarantee.
When to choose babbage-002
Choose babbage-002 only when its legacy behavior is itself part of the requirement. Suitable cases include:
- Maintaining an application that already calls the legacy Completions API.
- Generating controlled text or code continuations from carefully designed prompts.
- Reproducing historical GPT-3 base-model behavior for compatibility or evaluation.
- Operating an existing compatible fine-tuned babbage-002 derivative while a migration is planned.
- Running a simple, high-volume text-completion workload where the documented token price matters and modern API features are unnecessary.
Before selecting it, confirm that the application does not require instruction following, current knowledge, multimodal input, structured responses, function calling, streaming, or a long support horizon. The model’s maximum documented output is 16,384 tokens, but the supplied research does not provide a separate context-window value, so systems with demanding prompt-plus-output requirements should not assume a larger usable context than is documented.
When another option is more appropriate
A newer instruction-following model is a better fit for applications that need reliable responses to natural-language tasks, conversational interaction, complex reasoning, or code generation guided by explicit requirements. A tool-capable model is more suitable for agents that must call functions or interact with external systems. A multimodal model is necessary when images, audio, or video are part of the input or output.
Even for text-only work, a newer model may be preferable when the application needs structured output, streaming, web-grounded information, or continued availability after September 2026. babbage-002’s low price does not offset the shutdown risk for a new system, and its September 2021 knowledge cutoff makes it unsuitable as a standalone source for current information.
Bottom line
babbage-002 is best understood as a low-cost compatibility model, not a current general-purpose OpenAI model. It can continue text or code from prompts through the legacy Completions API and may remain useful for existing systems that depend on its behavior. However, it is text-only, not instruction-tuned, limited in API features, based on dated knowledge, and scheduled for shutdown on September 28, 2026. New projects should generally use a currently supported model with the capabilities and lifecycle needed by the application.