GPT-3.5 Turbo

The Model That Made Chat APIs Economical at Scale

GPT-3.5 Turbo sits at an important point in the history of production LLM applications. It made conversational interfaces inexpensive enough for broad experimentation while also becoming a default model for background automation that did not require GPT-4-level capability.

The current gpt-3.5-turbo alias uses the later 16K-class model generation rather than the earliest 4K GPT-3.5 Turbo releases. OpenAI lists a 16,385-token context window, up to 4,096 output tokens, and a September 1, 2021 knowledge cutoff.

The model accepts text and produces text. It does not support image, audio, or video input.

By current standards, GPT-3.5 Turbo is no longer the natural low-cost default. OpenAI has recommended newer small models for years and now marks the GPT-3.5 Turbo line for final retirement.

• Model ID: gpt-3.5-turbo.
• Optimized historically for Chat Completions and useful for non-chat text tasks.
• 16,385-token context window.
• Text-only input and output.
• Deprecated with final shutdown scheduled for October 23, 2026.

Look for Small Calls Hidden Throughout the Product

A mature application may use one stronger model for its headline AI feature while still relying on GPT-3.5 Turbo behind the scenes.

Common examples include routing a request to the correct workflow, extracting a small set of fields, translating short text, rewriting copy, generating tags, classifying support tickets, or summarizing content before another process runs.

These tasks are good migration candidates because current small models can often deliver stronger capability at lower cost. But each replacement should still be evaluated against the actual task contract.

A classifier with 99% acceptable accuracy can be more valuable than a more articulate model that changes labels unpredictably. An extraction step may care more about parse success than prose quality. Migration criteria should match the workload.

16K Was a Major Upgrade, but It Is Small Today

Earlier GPT-3.5 Turbo versions were associated with much smaller context budgets. In November 2023, OpenAI updated the default GPT-3.5 Turbo generation to a 16K-class context window.

That expansion made longer conversations, more examples, and larger source texts practical without immediate chunking. It also reduced the pressure to aggressively summarize chat history.

By 2026, 16K is a legacy-scale context window. Current models can support substantially larger prompts, which means older trimming, summarization, and chunking logic may be more restrictive than necessary after migration.

Do not remove that logic blindly, however. Selective context can still improve cost and relevance. The migration task is to determine which constraints were required by GPT-3.5 Turbo and which remain good application design.

• Context window: 16,385 tokens.
• Maximum output: 4,096 tokens.
• Larger than early GPT-3.5 Turbo generations.
• Much smaller than modern 128K and million-token models.

Do Not Assume Modern API Features Are Available

GPT-3.5 Turbo accepts and generates text. Image and audio modalities are not supported.

Fine-tuning remains listed as supported. However, OpenAI's current model documentation lists streaming, function calling, Structured Outputs, and predicted outputs as unsupported for the current deprecated model entry.

That matters when documenting or migrating old systems. Historical GPT-3.5 snapshots introduced capabilities such as function calling, but the current supported surface for the deprecated alias should be taken from current provider documentation rather than old integration assumptions.

A replacement model can therefore provide not only better quality but also a broader application contract: multimodal input, structured schemas, tool calling, larger context, or reasoning controls depending on the chosen model.