What is davinci-002?
davinci-002 is a legacy OpenAI GPT-3 base model released in August 2023. It was introduced as a replacement for the original GPT-3 davinci and curie base models in OpenAI’s older completion-model lineup. The model generates a continuation of a text prompt: given the beginning of a sentence, paragraph, document, or program, it predicts what should come next.
This completion-oriented design is different from the instruction-following behavior expected from modern chat models. A prompt such as “Write a short product description for a bicycle” may produce useful text, but davinci-002 was not specifically trained to reliably interpret requests, maintain a conversational role, or obey complex multi-step instructions. It is better understood as a general-purpose text continuation engine than as a current assistant model.
Current status and positioning
OpenAI lists davinci-002 as deprecated. The supplied OpenAI model information gives a deprecation date of September 26, 2025 and schedules API shutdown for September 28, 2026. After that shutdown, the exact model is expected to become unavailable through the API.
OpenAI also states that new fine-tuning training runs on davinci-002 stopped being supported on October 28, 2024. Previously created fine-tuned models were not immediately affected by that cutoff, which is why the model may still matter to organizations maintaining older systems. However, continued availability should not be treated as a long-term product commitment.
OpenAI recommends migrating away from davinci-002 rather than starting a new application with it. The supplied research identifies gpt-5.6-terra as the recommended replacement. That recommendation is a migration reference, not evidence that davinci-002 has comparable capabilities or interface compatibility with the newer model.
Core capabilities and limits
| Specification | Verified information |
|---|---|
| Provider | OpenAI |
| Model family | GPT-3 |
| Primary endpoint | Legacy Completions API |
| Input | Text |
| Output | Text, including natural language and code |
| Maximum output | 16,384 tokens |
| Knowledge cutoff | September 1, 2021 |
| Input and output price | $2.00 per 1 million tokens each |
| API shutdown | September 28, 2026 |
The model page supplies a maximum output limit of 16,384 tokens, but the research does not separately verify a context-window value. The output limit should therefore not be presented as the total amount of prompt text plus generated text that the model can process.
Its September 1, 2021 knowledge cutoff means that the underlying model knowledge does not include later events or information. Web search and retrieval features are not supported by the model itself. An application could theoretically provide external text in a prompt, but that would be application-supplied context rather than an update to the model’s knowledge.
Modalities and API features
davinci-002 supports text input and text output only. It has no verified image, audio, video, or other non-text input or output capability. It also does not provide native tool or function calling, structured output, or JSON mode. Applications that need those features should use a model and API designed to support them rather than trying to build a modern tool-calling workflow around this legacy completion model.
The supplied specifications mark streaming as unsupported and batch API availability as supported. These feature flags should be interpreted in the context of the legacy API and the model’s deprecation status. Batch availability does not make davinci-002 a current high-volume recommendation; it simply indicates that batch processing is listed as available for the model.
Fine-tuning is also marked as unsupported for current use. This is consistent with OpenAI’s stated end to new davinci-002 fine-tuning training runs in October 2024. Existing fine-tuned models created before the cutoff are a separate case and may explain why some older deployments still use the model.
Reasoning, coding, speed, and cost trade-offs
davinci-002 does not have a separately identified reasoning mode. It can produce text that appears to involve reasoning, but it should not be treated as a modern reasoning model with deliberate multi-step inference features. It also lacks the instruction-following orientation that makes newer models easier to use for planning, analysis, and reliable task completion.
Code continuation is one of its more relevant uses. Given a partial function, configuration fragment, or code comment, the model can generate a continuation. This is different from providing a dependable coding agent: there is no native tool use, no stated repository awareness, and no verified structured-output feature for returning machine-validated code operations.
The supplied editorial evaluation rates reasoning at 3 out of 10, coding at 4 out of 10, speed at 6 out of 10, and cost at 4 out of 10. These are editorial scores, not OpenAI benchmarks or provider-published ratings. They reflect the model’s legacy position: it may be adequate for simple completion workloads and relatively responsive, but it is substantially less capable and less suitable for modern application patterns than current general-purpose models.
At $2 per million input tokens and $2 per million output tokens, its input and output rates are identical. Cost calculations should count the two directions separately. For example, one million input tokens and one million generated output tokens would represent $4 in listed token charges before any other application costs. The nominal price alone is not enough to justify adoption because migration risk and the scheduled shutdown are important operational costs.
Best use cases for davinci-002
davinci-002 is most defensible in narrowly defined legacy situations:
- Maintaining an existing completion integration: An application already built around the legacy Completions API may continue using the model temporarily while a migration is planned.
- Text continuation: The model can continue predictable prose, templates, or other partially written text when the application does not require instruction-following behavior.
- Code continuation: Older systems may use it to complete code fragments or generate text from programming-oriented prompts.
- Existing fine-tuned deployments: Previously created davinci-002 fine-tuned models may remain relevant until their replacement and migration path have been validated.
- Short-lived evaluation or archival compatibility: Teams may need to reproduce historical outputs or test behavior from a legacy application before retiring it.
These are compatibility and maintenance use cases, not recommendations for new development. Any deployment should include a migration plan before the September 28, 2026 shutdown.
When to choose this model—and when not to
Choose davinci-002 only when compatibility with an existing legacy completion workflow is more important than access to current model features. It may be a reasonable temporary choice if changing the model would alter established outputs, and the application can tolerate text-only completion, the September 1, 2021 knowledge cutoff, and the lack of tools or structured responses.
For a new production system, another option is generally more appropriate. Use a current instruction-following model when the application must reliably follow user requests, produce consistent formats, summarize supplied material, or support conversational interaction. Use a model with native tool support when the system must call functions, retrieve information, or take actions. Use a multimodal model for image, audio, or video inputs. Use a current coding-oriented model when software generation requires more than continuation of a prompt.
Even if davinci-002 appears fast enough or its per-token price seems acceptable, its deprecation, discontinued new fine-tuning, lack of modern API features, and planned shutdown outweigh those advantages for most new projects. A migration evaluation should compare output quality, prompt format, latency, token consumption, and application behavior rather than assuming that a newer replacement will be drop-in compatible.
Bottom line
davinci-002 is a historical GPT-3 base model that remains relevant mainly because older applications and pre-existing fine-tuned models may still depend on it. Its strengths are straightforward text and code completion, a clear legacy API role, and a documented maximum output of 16,384 tokens. Its limitations are equally important: it is not an instruction-following chat model, has no multimodal or tool features, has a 2021 knowledge cutoff, no new fine-tuning support, and a scheduled API shutdown on September 28, 2026.
For maintenance, compatibility testing, or controlled migration work, understanding davinci-002 remains useful. For new systems, its deprecated status means the practical decision is usually to select a supported current model and begin migration rather than build further dependence on this one.