HomeAI models
OpenAI: GPT-3.5 Turbo Instruct
OpenAI
•
gpt-3.5-turbo-instruct
•
Released 
September 2023

OpenAI
gpt-3.5-turbo-instruct

OpenAI's instruction-tuned completion model for the legacy Completions endpoint, built for prompt-in, text-out workflows.

Modality:
Text
model ID
openai/gpt-3.5-turbo-instruct
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed

N/A
tok/s

Intelligence Index

N/A
/ 100

Context Window

4095
tokens

Input price

9
Anytoken

Output price

12
Anytoken
The GPT-3.5 Model Built for Raw Text Completion, Not Chat GPT-3.5 Turbo Instruct is OpenAI's instruction-tuned variant of GPT-3.5 Turbo built for the legacy Completions endpoint rather than Chat Completions. It follows a prompt-in, completion-out contract: you send a freeform text string and the model continues or responds directly, with no message-role structure. That makes it the natural successor to GPT-3 era completion models for teams maintaining prompt templates, few-shot pattern completion, and text-transformation pipelines that predate the chat format. It notably plays chess near 1750–1800 Elo, a signal of strong sequence modeling in the pure completion setting. Start building with GPT-3.5 Turbo Instruct via the AnyAPI.ai API.

Performance

Why the Completion Format Unlocks Behavior the Chat Model Lacks

GPT-3.5 Turbo Instruct's defining trait is strong single-turn sequence continuation without chat fine-tuning artifacts. The clearest evidence is chess: independent testing put it around 1750 Elo with an illegal-move rate under 0.1% across roughly 8,000 moves, while chat-tuned siblings and even GPT-4 struggled to play legally. This matters because RLHF chat tuning can degrade raw pattern completion; the Instruct model preserves it. In production, that translates into more reliable few-shot pattern completion, deterministic-friendly generation with precise max_tokens control, and predictable behavior for templated prompts where the chat message wrapper adds no value.

Benchmarks

Independent Evidence: Chess as a Sequence-Modeling Signal

The most cited independent measurement of GPT-3.5 Turbo Instruct is chess play through PGN-style completion prompting. Multiple external analyses placed it near 1750–1800 Elo, with one 150-game Stockfish evaluation reporting roughly 99.7% legal moves across about 8,000 moves. Researchers attribute this to the pure completion format: fine-tuned chat models produced far higher illegal-move rates under comparable prompts. Treat this as an interpretive proxy for reliable sequence continuation and low structural hallucination in templated tasks, not a general intelligence score. No broad official benchmark suite is published specifically for this variant.

Output Speed

N/A
tok/s

Intelligence Index

N/A
/ 100

MMLU

Broad world knowledge and problem-solving
0
%

GPQA

PhD-level scientific reasoning across physics, biology, chemistry.
0
%

HLE

Adherence to multi-step structured instructions.
0
%

LiveCodeBench

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

GPT-3.5 Turbo Instruct is a text-only model on OpenAI's legacy Completions endpoint (v1/completions), not Chat Completions. It accepts a freeform prompt and returns text, with a combined 4,096-token context shared between prompt and output. The two production-critical facts: the small context window rules out long-document workloads, and there is no chat message-role or system-prompt concept, so system instructions must be folded into the prompt. It supports streaming and logprobs, giving fine-grained control over token probabilities that many newer chat models expose less directly.
Verified Specifications — 
gpt-3.5-turbo-instruct
Input modalities
Text
output modalities
Text
Context window
4095
 tokens
Maximum output tokens
4096
Reasoning
No
Knowledge cutoff
September 2023
Pricing (standard)
9
 AnyTokens in
 / 
12
 AnyTokens out

Limitations & Trade-offs

Where gpt-3.5-turbo-instruct falls short

1
Small 4K context window. The combined prompt-plus-completion budget is only 4,096 tokens, so long documents, large few-shot sets, or extended histories will not fit. This directly rules out long-context RAG, multi-document summarization, and long transcripts. When context length is the requirement, GPT-4o mini or another large-context model is the correct choice.
2
No native tool calling or structured outputs. Because it runs on the legacy Completions endpoint, GPT-3.5 Turbo Instruct has no function calling, no JSON schema mode, and no chat system-role concept. Any structure must be enforced through prompt design and parsing. For agentic workflows, reliable tool invocation, or guaranteed JSON, a modern Chat Completions model is required.
3
Dated knowledge and no multimodality. Training data extends to September 2021 and the model is text-only, so it lacks knowledge of recent events and cannot process images, audio, or files. Applications needing current facts, vision, or document parsing should use a newer multimodal model instead.
4
Effectively legacy status. OpenAI now steers new GPT-3.5 Turbo usage toward GPT-4o mini as cheaper, more capable, and multimodal. GPT-3.5 Turbo Instruct remains available but is best treated as a compatibility and completion-format choice rather than a forward-looking default. Teams starting fresh should weigh migration risk before building new dependencies on it.

Best-Fit Workloads

Where this model earns its place

01

Legacy Completions integrations

‍
The model's primary reason to exist is serving pipelines built on the legacy Completions endpoint, the same interface GPT-3 era models used. Teams maintaining backwards-compatible systems can preserve existing prompt templates and the prompt-in, text-out contract without rewriting for the chat message format, avoiding migration rework while staying on a supported model.

02

Few-shot pattern completion

‍
Freeform prompt continuation makes GPT-3.5 Turbo Instruct well-suited to templated generation and few-shot pattern completion, where you bake examples directly into the prompt and let the model continue the pattern. Its low illegal-move rate in chess evaluations is an interpretive signal of reliable sequence continuation, useful for structured text transformation where deviations are costly.

03

Text transformation and templated generation

‍
For prompt-driven rewriting, formatting, and single-turn text transformation, the completion contract is often semantically cleaner than wrapping every request in a chat message array. Precise max_tokens control and logprobs support give developers tight control over length and token-level probabilities, valuable for deterministic-leaning generation and downstream scoring.

04

Probability-sensitive tasks with logprobs

‍
Because the model exposes logprobs on the Completions endpoint, it fits classification, ranking, and confidence-estimation tasks that consume token log probabilities. This makes it a practical choice for lightweight scoring pipelines where you need the probability of a completion rather than only the text, provided the 4K context and dated knowledge are acceptable.

Pricing in anytokens via AnyAPI
Input
9
₳
Output
12
₳
Cache write
—
₳
Cache read
—
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access gpt-3.5-turbo-instruct via AnyAPI.ai

Access gpt-3.5-turbo-instruct through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate gpt-3.5-turbo-instruct without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access gpt-3.5-turbo-instruct and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test gpt-3.5-turbo-instruct against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use gpt-3.5-turbo-instruct from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use gpt-3.5-turbo-instruct for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

GPT-3.5 Turbo Instruct is an instruction-tuned OpenAI model for the legacy Completions endpoint. It takes a freeform text prompt and continues it directly, making it suited to few-shot pattern completion, templated generation, text transformation, and legacy pipelines built before the chat message format existed.

GPT-3.5 Turbo Instruct has a 4,096-token context window, shared between the input prompt and the generated output. This small window makes it unsuitable for long documents or large few-shot sets; workloads needing more context should use a larger-context model such as GPT-4o mini.

No. GPT-3.5 Turbo Instruct is compatible only with the legacy Completions endpoint (v1/completions), not Chat Completions. It uses a freeform prompt string rather than a message-role structure and has no system-prompt concept, so instructions must be included directly in the prompt.

Independent testing placed GPT-3.5 Turbo Instruct near 1750–1800 Elo with an illegal-move rate under 0.1% in PGN-style prompting. Researchers attribute this to the pure completion format: chat fine-tuning appears to degrade raw sequence continuation, which the Instruct variant preserves.

OpenAI recommends GPT-4o mini for most new GPT-3.5 Turbo use cases, citing lower cost, more capability, multimodality, and comparable speed. Choose GPT-3.5 Turbo Instruct specifically when you need the legacy Completions format, logprobs, or completion-native pipelines; otherwise GPT-4o mini is the better default.