OpenAI’s GPT-3 API text-davinci-002 and text-davinci-003 models powered early AI writing applications until their shutdown on January 4, 2024. OpenAI ended support for these legacy completion models to shift developers to newer chat-based endpoints like gpt-3.5-turbo.
What It Was
GPT-3 launched in June 2020 as OpenAI’s breakthrough in large language models, with API access expanding in late 2021. The text-davinci series formed GPT-3’s top capabilities: text-davinci-001 in 2021, text-davinci-002 with RLHF improvements in 2022, and text-davinci-003 for better instruction following in November 2022.
(prompt input, text output) before chat formats emerged. Pricing reached $0.02 per 1,000 tokens for text-davinci-003. Applications included blog generation, marketing copy, summarization, question-answering, early chatbots, and code tasks.
What Happened
OpenAI announced deprecations for older GPT-3 completions models including text-davinci-002 and text-davinci-003 in July 2023, setting January 4, 2024, as the shutdown date. Post-shutdown, calls routed to gpt-3.5-turbo-instruct as a bridge, though outputs differed enough to affect some uses.
This followed signals since ChatGPT’s November 2022 launch, pushing toward gpt-3.5-turbo chat models. The text-davinci models joined prior GPT-3 deprecations in a pattern of retiring legacy endpoints.
Why It Died
Newer models like GPT-3.5 (2022) and GPT-4 (2023) outperformed text-davinci on benchmarks while cutting costs: gpt-3.5-turbo at $0.002 per 1,000 tokens, a 10x drop. OpenAI maintains 12 to 24 months of overlap before full shutdowns, as seen with this GPT-3 cycle completing amid GPT-4 adoption.
Retaining infrastructure for low-usage models proved inefficient as capabilities advanced.
Current Alternatives
Teams replacing text-davinci models use these active APIs as of 2026-04-15:
- gpt-5.1-chat-latest via OpenAI API: Successor post-GPT-4o shutdown on February 16, 2026; handles text generation with top performance.
- Claude via Anthropic API: Matches GPT capabilities for writing; reduces OpenAI reliance.
- Gemini via Google API: Competitive on instruction tasks; viable migration option.
- Llama 4 (open-source): Self-hostable at low cost; approaches closed-model quality for text tasks.
Lessons
OpenAI’s model lifecycle shows rapid iteration: expect 12 to 24 month support windows before deprecations, as with text-davinci to GPT-4o, now to GPT-5 family. Developers must audit dependencies yearly and test replacements early, since bridges like gpt-3.5-turbo-instruct alter behaviors.
Diversification cuts risks; relying on one provider invites disruptions, as GPT-4o retirement affected API users despite ChatGPT variances. Open-source options like Llama provide hedges against proprietary shutdowns.