GPT-3.5

GPT-3.5 is a family of large language models developed by OpenAI. OpenAI used the name for the series from which it fine-tuned the model behind the original ChatGPT research preview, launched on November 30, 2022.^[1] The later GPT-3.5 Turbo line made related chat-oriented models available through the OpenAI API, beginning with gpt-3.5-turbo and the pinned gpt-3.5-turbo-0301 snapshot on March 1, 2023.^[3]

GPT-3.5 is a family label, not one unchanging model. OpenAI updated the Turbo line through dated snapshots, exposed an unpinned alias that could point to a recommended snapshot, and separately offered gpt-3.5-turbo-instruct through the legacy Completions interface.^[3][5] The ChatGPT product also changed its underlying models over time. It therefore cannot be treated as a permanent synonym for GPT-3.5.

OpenAI's public GPT-3.5 launch records and model documentation do not give the family's parameter count, layer count, full pretraining corpus, training-token total, training compute, or model weights.^[1][10] In particular, GPT-3's disclosed 175-billion-parameter configuration cannot be assigned to GPT-3.5 without a primary source. As of July 28, 2026, OpenAI described GPT-3.5 Turbo as a legacy, text-only API model with a 16,385-token context window, a 4,096-token maximum output, and a September 1, 2021 knowledge cutoff.^[10]

Name and scope

OpenAI introduced the "GPT-3.5 series" publicly in its ChatGPT launch article. That article says the initial ChatGPT model was fine-tuned from a model in the series and that the underlying GPT-3.5 model had finished training in early 2022.^[1] The statement establishes a direct relationship between the family and the first ChatGPT release, but it does not identify the underlying model with a public API snapshot.

The name later appeared in several related but distinct forms:

Scroll sideways for more →

This terminology also limits what can be concluded from tests. A result for the December 15, 2022 ChatGPT service, gpt-3.5-turbo-0301, or gpt-3.5-turbo-0613 describes that tested system and date. It is not automatically a measurement of every GPT-3.5 model.

Development and training record

Relationship to InstructGPT

InstructGPT supplied important methodological background for the original ChatGPT training process. In the InstructGPT study, OpenAI first performed supervised fine-tuning on human-written demonstrations, trained a reward model from human rankings of model outputs, and then optimized the policy with proximal policy optimization. The study tested this procedure on specific GPT-3 model sizes; its parameter counts and data totals are properties of those experiments, not published GPT-3.5 specifications.^[2]

For the November 2022 ChatGPT release, OpenAI described a related but not identical procedure. Human trainers wrote conversations while playing both the user and assistant roles. OpenAI mixed that dialogue dataset with reformatted InstructGPT data, collected rankings of alternative model responses, trained reward models from those comparisons, and ran several iterations of reinforcement learning using proximal policy optimization. OpenAI characterized the overall approach as reinforcement learning from human feedback, with differences from the earlier InstructGPT data-collection setup.^[1]

The launch article also states that ChatGPT and GPT-3.5 were trained on Azure AI supercomputing infrastructure.^[1] It does not disclose which GPT-3.5 base model was used, the size of that model, a complete dataset inventory, or the amount of compute consumed. Claims about distillation, a specific transformer configuration, or a particular mix of web and book data are not established by this source.

Product alignment and base-model facts

OpenAI's description separates two stages that are often conflated. A model in the GPT-3.5 series finished training in early 2022, and the initial ChatGPT system was then fine-tuned from a model in that series.^[1] The dialogue demonstrations, preference comparisons, reward model, and reinforcement-learning procedure describe the post-training of ChatGPT. They are not a complete account of GPT-3.5 pretraining.

The original ChatGPT release also should not be equated with the March 2023 API snapshot. OpenAI later said the gpt-3.5-turbo family released through the API was the same model used in the ChatGPT product at that time, but the API article did not identify that statement retroactively with every ChatGPT build since November 2022.^[3]

API releases

Chat Completions

OpenAI launched the GPT-3.5 Turbo API on March 1, 2023. Instead of a single free-form prompt, the Chat Completions interface accepted an ordered sequence of messages and role metadata. OpenAI said the messages were rendered internally into tokens using Chat Markup Language, or ChatML. The initial API price was $0.002 per 1,000 tokens, with no separate input and output rate stated in that announcement.^[3]

The launch established two versioning choices. Applications could call the generic gpt-3.5-turbo alias, which OpenAI said would receive the recommended stable model, or pin gpt-3.5-turbo-0301. A pinned snapshot provided behavioral stability, while an alias could change as OpenAI moved it to a newer snapshot.^[3] OpenAI later emphasized that model upgrades could improve many evaluations while still making some tasks worse, which is why snapshot pinning mattered for production testing.^[4]

The June 2023 update illustrated how that distinction worked. OpenAI announced that the generic gpt-3.5-turbo name would move to the new version on June 27, while applications that still needed the earlier behavior could continue to request gpt-3.5-turbo-0301 explicitly.^[4] The same announcement separated input and output billing for the updated standard model, at $0.0015 and $0.002 per 1,000 tokens respectively, and introduced a 16k model at $0.003 and $0.004 per 1,000 tokens.^[4] These were historical rates for the June models, not prices for every snapshot carrying the Turbo name.

Snapshot names made some experiments more reproducible, but pinning did not make a model permanent. OpenAI extended the availability of several early snapshots after developer feedback and later shut them down.^[4][11] An application therefore needed both behavioral regression tests and an availability plan: the generic alias could change before the application changed its code, while a dated snapshot could eventually stop accepting requests.

Release chronology

Scroll sideways for more →

Function calling did not itself execute external code. The -0613 models could produce a JSON object containing arguments for a function described by the developer. The application remained responsible for deciding whether to run a tool, validating the arguments, executing the operation, and returning any result to the model. OpenAI warned that untrusted tool output could contain instructions that influence a model and recommended user confirmation before actions with real-world consequences.^[4]

The November 2023 JSON mode likewise concerned output syntax. OpenAI said gpt-3.5-turbo-1106 could be constrained to produce valid JSON and could request several function calls in one message.^[7] Neither feature guaranteed that generated values were factually correct or safe to execute.

Completions and GPT-3.5 Turbo Instruct

The chat-oriented Turbo models were not the only API form. In July 2023, OpenAI announced gpt-3.5-turbo-instruct for the older Completions interface. It described the model as InstructGPT-style, trained similarly to text-davinci-003, and intended as a drop-in replacement for older instruction-following completion models.^[5]

That replacement relationship does not prove that every predecessor in OpenAI's migration table had the same architecture or belonged to the GPT-3.5 family. It only documents the migration path OpenAI provided. The current model documentation listed gpt-3.5-turbo-instruct beside the Turbo snapshots as of July 28, 2026, while the deprecation record treated it as a separately scheduled model.^[10][11]

Documented specifications

OpenAI's model page described the following properties for the default GPT-3.5 Turbo API model as of July 28, 2026:

Scroll sideways for more →

These values apply to the documented default Turbo model, not automatically to the original ChatGPT model, -0301, the separate 16k -0613 snapshot, or Turbo Instruct.^[10] Historical release notes used rounded labels such as "4k" and "16k"; the current card gives the exact 16,385-token context and 4,096-token output limits for -0125.

The knowledge cutoff is also distinct from a release date. A January 2024 snapshot can have a September 2021 knowledge cutoff because the snapshot date identifies the served model build, not the latest event represented in its learned parameters.^[8][10] Information supplied in a prompt can be newer than the knowledge cutoff, but a longer context window does not guarantee that the model will use every part of that prompt reliably.

Pricing history

GPT-3.5 Turbo's API price fell across its major releases. The units and billing structure changed, so historical rates should be tied to a date and snapshot.

Scroll sideways for more →

The January 2024 rates remained on the GPT-3.5 Turbo model card as of the research cutoff.^[10] They should not be used to calculate the cost of earlier snapshots, fine-tuned inference, or unrelated models.

Fine-tuning

OpenAI made supervised fine-tuning for GPT-3.5 Turbo available on August 22, 2023. The launch workflow involved preparing examples, uploading a training file, creating a fine-tuning job, and then calling the resulting model. OpenAI presented instruction adherence, response formatting, and consistent tone as possible narrow uses, but those benefits depended on training data and task-specific evaluation.^[6]

The original fine-tuning prices were:

Scroll sideways for more →

OpenAI's worked example used a 100,000-token training file for three epochs. That represents 300,000 billed training tokens, so the stated expected training cost was $2.40, not $0.80.^[6] Later announcements changed fine-tuned inference prices for newer snapshots, which means the launch rates must not be presented as timeless.

The August launch concerned supervised fine-tuning of the 4k GPT-3.5 Turbo model. OpenAI said support for function calling and the 16k model was still planned for later that year.^[6] It also said submitted training data would pass through the Moderation API and a GPT-4-powered moderation system intended to detect training data that conflicted with its safety standards. This was a description of OpenAI's screening process, not evidence that a fine-tuned model would be safe or accurate for a particular use.

OpenAI's examples of possible gains, including better instruction adherence, more reliable response formats, and a more consistent tone, came from the vendor's private beta and early tests.^[6] They were not guarantees. A deployment still required a held-out evaluation set that represented its own task, including cases where the correct action was to abstain or reject malformed input.

Fine-tuning availability narrowed in 2026. OpenAI's deprecation page says that, beginning May 7, organizations that had never run fine-tuning could not create training jobs. Beginning July 2, new jobs were also unavailable to organizations that had not run inference on a fine-tuned model during the previous 60 days. Those restrictions were already in effect by July 28, 2026.^[11]

At that cutoff, inference on existing fine-tuned models remained available, but OpenAI had announced an October 23, 2026 shutdown for fine-tuned GPT-3.5 Turbo models. The date was scheduled, not yet completed.^[11]

Capabilities and interfaces

OpenAI described GPT-3.5 Turbo as optimized for chat while also suitable for non-chat text and code tasks.^[10] The messages interface let an application distinguish instructions, user content, and prior assistant responses. Historical snapshots added structured-output features, function-call requests, and a larger context window.^[3][4][7]

Those interfaces made the family useful for tasks such as drafting, extraction, classification, summarization, question answering, and code generation, but the model card does not assign a universal accuracy level to any of them. Results depend on the snapshot, prompt, evaluation data, decoding settings, external context, and application controls. OpenAI also noted that a new snapshot could improve most internal metrics while regressing on some tasks.^[4]

The GPT-4 technical report illustrates another reason to preserve version boundaries. Its GPT-3.5 comparisons used a standard ChatGPT snapshot for exam tests, while other evaluations used different GPT-3.5 systems. The report found GPT-4 ahead on most tested exams and on OpenAI's internal factuality evaluations, but it also warned that both model generations could produce reasoning errors and false statements.^[16] These comparisons do not supply an architecture or parameter count for GPT-3.5.

Evaluations and limitations

Limitations identified at launch

OpenAI listed several limitations for the November 2022 ChatGPT system. It could produce plausible but incorrect or nonsensical answers, change its response after small prompt rephrasings, write excessively long answers, guess at ambiguous intent instead of asking a clarifying question, and sometimes respond to harmful instructions or show biased behavior.^[1] These are documented observations about the launch system, not a numerical error rate for every later Turbo snapshot.

Post-training also introduced tradeoffs. OpenAI wrote that making the model more cautious could cause it to decline questions it could answer, while training against human demonstrations could reward an answer that matched what the demonstrator knew rather than what the model knew.^[1] A fluent response therefore was not evidence that its factual claims had been checked.

Snapshot-specific research

Independent and peer-reviewed studies reinforce the need to identify the tested version:

Scroll sideways for more →

These studies do not yield one permanent "GPT-3.5 score." They used different interfaces, snapshots, prompts, datasets, and outcome definitions. They do show recurring application risks: plausible falsehoods, sensitivity to how and where information is presented, uneven performance across tasks and languages, and behavior that changes under adversarial instructions.

The study designs also limit comparisons between their numbers. Bang et al. interacted with the public ChatGPT service without API access and reported the service version so readers could reproduce the setup.^[12] DecodingTrust used an API snapshot and evaluated eight distinct trust dimensions rather than a single aggregate notion of safety.^[13] Lost in the Middle varied the position of relevant evidence while holding the task structure under experimental control.^[14] Zhou et al. evaluated five benchmark families with 15 prompt variants and classified responses as correct, incorrect, or avoidant.^[15] Those are different questions, so a score from one study should not be substituted for another.

For the multi-document question-answering experiment in Lost in the Middle, the reported closed-book accuracy for GPT-3.5 Turbo was 56.1 percent. In the worst 20-document and 30-document placements, performance fell below that closed-book result even though the prompt contained the relevant document.^[14] This does not mean that adding documents generally hurts performance. It shows that, in that controlled setup, access to evidence was not sufficient when the model failed to use evidence positioned in the middle.

Practical interpretation

A 16,385-token context window is a capacity limit, not a guarantee of faithful use. The long-context study found that relevant evidence could be present within the window yet still be used less effectively when placed in the middle.^[14] Applications that depend on supplied documents therefore need evaluations that vary document order and context length, rather than testing only whether the prompt fits.

Likewise, JSON mode and function calling constrain the form of an output but do not verify the substance. Generated arguments can still be wrong, incomplete, or manipulated by untrusted context.^[4][7] A production system should validate structured data, restrict tools, require confirmation for consequential actions, and check factual claims against authoritative sources.

Role in ChatGPT

The initial ChatGPT release is central to GPT-3.5's history but should be described precisely. OpenAI launched ChatGPT as a free research preview on November 30, 2022 and said it was fine-tuned from a GPT-3.5-series model.^[1] When the GPT-3.5 Turbo API arrived in March 2023, OpenAI described that API family as the model then used in the ChatGPT product.^[3]

Neither statement means the consumer service stayed on one frozen snapshot. ChatGPT was a continuously updated product, while API users could choose an unpinned alias or a dated model. Claims measured through the web product need a service date, and API findings need a model ID wherever the study provides one.

On July 18, 2024, OpenAI introduced GPT-4o mini and made it available to ChatGPT Free, Plus, and Team users in place of GPT-3.5. OpenAI priced GPT-4o mini at $0.15 per million input tokens and $0.60 per million output tokens and described it as cheaper, more capable, and multimodal relative to GPT-3.5 Turbo.^[9] Those comparative claims came from OpenAI's release evaluation, but the replacement date and product tiers are directly documented.

GPT-3.5 Turbo remained available through the OpenAI API after its removal from those ChatGPT tiers. The product replacement in July 2024 and the later API deprecations are separate events.^[9][10][11]

Deprecation and availability

OpenAI defines deprecation as notice that a model or endpoint will be retired, followed by a later shutdown date. A deprecated model may remain callable between those events.^[11] That distinction is necessary for GPT-3.5 because several shutdowns were announced but had not occurred at the July 28, 2026 research cutoff.

Scroll sideways for more →

The model card still labeled GPT-3.5 Turbo "legacy" and "still available for use in the API" at the cutoff.^[10] "Legacy," "deprecated," and "shut down" are therefore not interchangeable. The first two described its documented status on July 28, 2026; the last applied only to snapshots whose shutdown dates had already passed.

Significance

GPT-3.5 connected three stages in OpenAI's language-model development: the instruction-following methods studied with GPT-3, a conversational product fine-tuned from the GPT-3.5 series, and a low-cost messages-based API with named snapshots.^[1][2][3] Its API history also made versioning visible to developers. An application could follow a moving alias or pin a snapshot, but either choice required evaluation when behavior or availability changed.^[3][4]

The family also shows why model documentation must separate known facts from plausible inheritance. GPT-3 has a published architecture and parameter count, while the GPT-3.5 records cited here focus on post-training, interfaces, context limits, prices, and release history. Treating the predecessor's specifications as GPT-3.5 facts would erase that evidentiary gap.

By July 2026, newer models had replaced GPT-3.5 in ChatGPT and OpenAI had scheduled the remaining API line for shutdown.^[9][11] Its historical importance rests on documented product and API transitions, not on unsupported claims about proprietary internals or unverified adoption totals.

See also

• GPT-3
• GPT-4
• Hallucination
• AI model lifecycle calendar
• AI model comparison finder
• AI API cost and context planner

References

1. ^OpenAI. "Introducing ChatGPT." November 30, 2022.
2. ^Long Ouyang et al. "Training language models to follow instructions with human feedback." *Advances in Neural Information Processing Systems* 35, 2022.
3. ^OpenAI. "Introducing ChatGPT and Whisper APIs." March 1, 2023; updated April 24, 2024.
4. ^OpenAI. "Function calling and other API updates." June 13, 2023; updated July 20, 2023.
5. ^OpenAI. "GPT-4 API general availability and deprecation of older models in the Completions API." July 6, 2023; updated April 24, 2024.
6. ^OpenAI. "GPT-3.5 Turbo fine-tuning and API updates." August 22, 2023.
7. ^OpenAI. "New models and developer products announced at DevDay." November 6, 2023.
8. ^OpenAI. "New embedding models and API updates." January 25, 2024.
9. ^OpenAI. "GPT-4o mini: advancing cost-efficient intelligence." July 18, 2024.
10. ^OpenAI. "GPT-3.5 Turbo Model." OpenAI API documentation, retrieved July 31, 2026; article status statements are bounded to July 28, 2026.
11. ^OpenAI. "Deprecations." OpenAI API documentation, retrieved July 31, 2026; only events and announced schedules known by July 28, 2026 are used.
12. ^Yejin Bang et al. "A Multitask, Multilingual, Multimodal Evaluation of ChatGPT on Reasoning, Hallucination, and Interactivity." *Proceedings of IJCNLP-AACL 2023*, pages 675-718.
13. ^Boxin Wang et al. "DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models." *Advances in Neural Information Processing Systems* 36, 2023.
14. ^Nelson F. Liu et al. "Lost in the Middle: How Language Models Use Long Contexts." *Transactions of the Association for Computational Linguistics* 12, 2024, pages 157-173.
15. ^Lexin Zhou et al. "Larger and more instructable language models become less reliable." *Nature* 634, 2024, pages 61-68.
16. ^OpenAI. "GPT-4 Technical Report." March 2023.