GPT Models: Model IDs, ChatGPT Defaults and Where They Run

Choose GPT models by the product you use as well as the task: OpenAI GPT-6.1 Sol is a starting point for complex Work and Codex tasks, GPT-6 Luna fits focused repeatable work, and GPT-6 Astra is the flagship for demanding problems. Regular ChatGPT follows a different map: Free and Go default to GPT-5.6 Luna, while GPT-6 is rolling out to eligible paid accounts. A model name alone does not tell you what your account can run.

Verified against OpenAI's model cards, deprecations page and Help Center on October 25, 2026. The recommendations below are analysis of those sources, not claims from hands-on model testing.

GPT Models Comparison: IDs, Context and Where They Run

OpenAI's current flagship API lineup is GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. GPT-6 Sol remains separately documented; GPT-6 Pro is a ChatGPT option powered by Astra; gpt-oss is the downloadable branch. The distinction matters when a founder asks a builder to reproduce an answer from ChatGPT inside an application. OpenAI model catalog.

The purpose column below reproduces OpenAI's short descriptions, with the Pro row using its description of the Pro option. Context values are the exact token counts on the individual model cards, not guaranteed ChatGPT conversation limits. An API ID is the identifier software sends to select a model. Work is ChatGPT's experience for longer tasks and finished deliverables; Codex is its software development agent.

Availability sources: Sol and Luna launch, GPT-6.1 Sol launch, Astra launch, ChatGPT Work and Codex, and gpt-oss deployment partners.

Azure: unconfirmed means the OpenAI pages checked do not establish that model's Azure availability. It does not mean Microsoft cannot offer it. Even a named deployment channel does not establish access in your particular region or subscription. An operator procuring through Azure should have the builder confirm the exact deployable model before promising it to colleagues.

OpenAI's descriptions are intended positioning, not a guarantee that every task follows the same ranking. For the family background, see GPT-6 explained.

ChatGPT Models Explained: Your Plan's Default

Your Chat plan determines the default and the thinking levels you can select. The published rollout target is not proof that every existing conversation has switched. ChatGPT model availability.

Sources: general Chat plan table, Business models, Enterprise models.

ChatGPT Free Models: Chat and Desktop Access Differ

A Free or Go account can encounter different Luna generations in different experiences. OpenAI's Sol/Luna launch gives those users GPT-6 Luna access in the desktop app. That does not replace the Help Center's GPT-5.6 Luna default for regular Chat. Check the experience before comparing answers. Desktop access.

For a founder using Chat to draft a customer update, the practical choice is the model and thinking level visible in that conversation. For a builder opening Codex from the same account, inspect Codex's selection separately. A subscription is access to products with their own controls; it is not one model permanently attached to your name.

GPT-5.6 Models: Existing Selections Can Stay Put

An existing explicit GPT-5.6 Sol selection stays on GPT-5.6. The rollout moves the former Latest selection to GPT-6 while preserving its effort. Existing-selection behavior.

That is why colleagues on the same plan may report different models. Check their conversation settings and workspace access before treating the difference as a faulty answer or an account problem. Business and Enterprise model permissions can also constrain what a user can select.

OpenAI Model Comparison: Separate Model, Product and Effort

Record the model, product and reasoning effort together whenever a result matters. Think of a model as the engine, the product as the vehicle, and effort as how much work the engine is asked to do. The vehicle determines which controls and tools you have; an engine specification alone cannot describe the whole trip.

In a weekly operating review, “ChatGPT wrote this” is too vague to reproduce. Record the selected model, whether the task ran in Chat, Work, Codex or an API application, and the thinking setting. Keep the source documents and instructions too. This gives the next operator something concrete to repeat when the output changes.

More effort does not automatically select another model. In Plus and Pro Chat, asking an Instant conversation to think harder does not automatically move it to a higher thinking level. Use the control. Business can add reasoning automatically through its Higher intelligence setting, so a workspace's behavior may differ. Personal-plan behavior, Business controls.

For an operator reviewing a launch plan, increase effort when the task requires reconciling constraints. If the problem is a missing source document, supply that document first. Extra reasoning is not a substitute for the facts the task needs.

Which GPT Model Is Codex Giving You?

Codex does not have one reliable universal default across every plan, client and workspace. OpenAI recommends GPT-6.1 Sol for complex coding and agentic workflows when available, and Luna for focused, repeatable tasks. Availability varies with sign-in method and client. Work and Codex models.

Codex keeps a model you selected manually. Workspace administrators can set starting models and reasoning levels for Work and Codex separately from Chat. If a CTO wants a consistent baseline for code review, set and document that baseline rather than assuming every engineer inherited it. Starting defaults.

1. 1

   Read the current selection

   In Work or Codex, inspect the model control beneath the composer. In an interactive Codex CLI session, use /model to inspect and change the model or effort.

2. 2

   Select a model explicitly for repeatable work

   For a new CLI session, OpenAI documents codex --model gpt-6.1-sol. Use a model your account and client actually expose; a name in a guide does not grant access.

3. 3

   Check the client and workspace

   Astra requires Codex CLI 0.153.0 or newer. For Enterprise, confirm model access with the administrator as well as updating the client.

Sources: Codex model controls, Astra client requirement.

One retirement illustrates the product boundary: GPT-5.5 retired from Codex with ChatGPT sign-in on October 14, 2026, at 19:00 UTC. OpenAI says API access and Codex sessions using an API key were unaffected by that change. Do not translate a Codex retirement into an API shutdown date. Codex retirement notice.

Context Windows: What the Numbers Let You Do

GPT-6.1 Sol's API model card lists 1,050,000 tokens of context, while Work and Codex default to 500,000. A token is a small unit of text or other model input. Context is the working space available to a request, not a promised number of complete documents or a guarantee that every detail will be used correctly. API specification, Work and Codex limit.

For a builder analyzing a repository, that means the API ceiling is not a safe assumption for a Codex task. Select the relevant files and preserve room for instructions, tool results and the answer. For an operator comparing supplier proposals, group the documents around the decision and ask for cited evidence; a larger pile is not inherently a better brief.

The GPT-6 model cards also distinguish the context window from 128,000 maximum output tokens. Input capacity and answer length are separate limits. The downloadable gpt-oss cards list 131,072 context tokens. Use the numbers for the exact deployment you are planning. Astra limits, gpt-oss limits.

The detailed migration tradeoff lives in GPT-6.1 Sol vs GPT-6 Sol.

gpt-oss: The Models You Can Download

Choose gpt-oss when controlling deployment is a requirement. Open weights means the learned model files are available to download and run. OpenAI publishes gpt-oss-120b and gpt-oss-20b under Apache 2.0, but does not serve them through its API or ChatGPT. A model card in the catalog does not make it an OpenAI-hosted endpoint. Open-weight model guidance.

OpenAI says gpt-oss-120b fits within 80 GB of memory, while gpt-oss-20b requires 16 GB. Those deployment figures help distinguish server-sized and smaller local experiments. They do not promise useful speed on every device with that much memory. OpenAI's deployment guidance.

For a CTO whose requirement is running inside controlled infrastructure, this branch deserves evaluation. For a founder who simply needs a hosted assistant, downloading weights introduces a separate job: operating the inference service, securing access and maintaining the surrounding application. Assign an owner to that job before choosing the model.

The gpt-oss models are text-only according to their cards. If the workflow depends on interpreting screenshots, do not assume the image input available in GPT-6 transfers to these weights. gpt-oss-20b modalities.

Older GPT Models: Retirement Dates That Matter

Audit exact model IDs, because a family can contain both retired and still-available snapshots. These are selected API shutdowns for general text and reasoning models, taken from OpenAI's deprecations page:

• October 23, 2026, already shut down: listed GPT-3.5 Turbo, GPT-4 and GPT-4 Turbo aliases/snapshots; GPT-4.1 nano; gpt-4o-2024-05-13; and listed o1, o1-pro, o3-mini and o4-mini models.
• December 11, 2026: GPT-5, GPT-5 mini, GPT-5 nano, GPT-5 pro, o3 and o3-pro, including the snapshots identified in the notice.
• January 14, 2027: gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14. OpenAI names gpt-6-luna as their replacement.
• April 1, 2027: gpt-5.3-codex, gpt-5.1, gpt-5.1-2025-11-13, gpt-5.4-nano and gpt-5.4-nano-2026-03-17.

Deprecated means a shutdown has been announced; it does not mean access has already ended. Legacy means the model no longer receives updates. OpenAI also exposes an optional shutdown_date through its model-list and retrieval endpoints. Status definitions and monitoring.

For an operator with a document-processing workflow, ask the builder for the configured identifier and any fallback identifier. Put each affected workflow on a migration schedule, then verify the replacement against representative documents. A migration is complete when the application still produces acceptable results, not when a configuration file contains a newer name.

Best GPT Models for Founders, Operators and Builders

Start with the task's failure cost and deployment constraints. OpenAI's API catalog points uncertain users toward Astra; its Work/Codex guidance recommends GPT-6.1 Sol for complex coding and Luna for focused work. The following choices are practical starting points for evaluation, not a benchmark ranking. API starting point, Work/Codex guidance.

For a funded founder: use the strongest available model for a consequential strategy document with conflicting evidence, then review its reasoning against the sources. Try Astra for that task. Do not make Astra a prerequisite for routine copy changes when a lighter model already meets the brief.

For a senior operator: begin a repeated extraction or classification workflow with Luna and a clearly defined output. Include awkward examples, missing fields and contradictory records in the evaluation. Escalate when the failures require interpretation rather than better input instructions.

For a solo technical builder: start complex code changes with GPT-6.1 Sol where available. Keep a known-good original Sol configuration while checking tool calls and effort settings. Use Sol vs Luna for the deeper distinction between broad engineering work and tightly scoped tasks.

For a mid-market CTO: choose the allowed deployment first. OpenAI API access, Azure deployment and self-hosted gpt-oss are different operating decisions. Then compare eligible models against the same acceptance criteria. A stronger model in a deployment your organization cannot use is not a candidate.

For current rates, use the dedicated OpenAI API pricing guide.

What's Overhyped: The Newest Name Is Not a Deployment Plan

A newer model name does not settle access, integration or quality. “Uses GPT-6” leaves the actual family member unresolved. “Includes Astra” leaves the product and allowance unresolved. “Large context” leaves the product's working limit unresolved. Ask for the full configuration before making a business commitment.

The useful decision rules are straightforward:

• Act now if a configured model is retired or has an approaching shutdown, or if a replacement fixes a failure your workflow actually has.
• Evaluate before switching when changing Sol generations affects tool calling, reasoning settings or the documents your application can process.
• Wait when the workflow already meets its acceptance criteria and the only reason to change is a newer label. Keep its retirement date under review.
• Remain unchanged when a retirement applies to another product or authentication method, as with the Codex GPT-5.5 change and API-key sessions.

A model upgrade deserves a small, representative comparison using your inputs, expected outputs and review process. Vendor positioning is useful for choosing candidates; your acceptance criteria decide whether a candidate can take over the job.

Frequently Asked Questions

What model does OpenAI use?

OpenAI uses different models across products. Its flagship API catalog lists GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna; ChatGPT's selection depends on the plan and experience. Current catalog.

What are the different types of AI models?

For this decision, distinguish general text and reasoning models, specialized image or audio models, and downloadable open-weight models. The first distinction is what the task requires; the second is where the model can run. OpenAI separates these categories in its model catalog.

What is the GPT-2 model?

GPT-2 is an earlier OpenAI transformer language model announced in 2019, trained to predict the next word. It is historical context for this lineup, not a current ChatGPT plan label. OpenAI's GPT-2 announcement.

What are the limitations of GPT models?

Context, output length, supported input types and tool compatibility differ by model and product. A large context window also does not guarantee a correct answer. For an application, verify the model card and evaluate the complete workflow against known examples. GPT-6 guidance.

What shouldn't you tell ChatGPT?

Do not paste passwords or confidential material without authorization and an approved data-handling setup. OpenAI advises against entering sensitive information you would not want reviewed or used; check your account's Data controls and your organization's policy. Consumer data guidance.

Which GPT has no limit?

No option is unlimited in every dimension. Everyday chat allowances do not remove context limits, tool limits, reasoning allowances or safeguards. Business explicitly distinguishes virtually unlimited Instant chat from reasoning, Pro, Work and Codex usage. Business limits.

Which AI is uncensored?

Open weights describes access to the model files, not a promise of unrestricted behavior. OpenAI says gpt-oss underwent safety training. If deployment control is your requirement, evaluate that directly rather than relying on an “uncensored” label. gpt-oss release.

Is there a better AI than ChatGPT?

There is no useful universal answer without a task. ChatGPT is a product, while a GPT model is one component. Compare complete workflows on your documents, required tools, acceptable errors and review burden.

What's the dumbest AI ever?

That is not a useful selection criterion. A small model can succeed on a narrow extraction task while a more capable model fails with missing context. Define the job and inspect the failures rather than ranking systems by an insult.

Your Monday Move

Create a model inventory for the workflows your business relies on. For each workflow, record the product, exact model selection, effort, owner and any shutdown date. Ask the owner to identify an example that passes and an example that fails. Prioritize retired dependencies, then evaluate upgrades where the failure matters.

That inventory turns “which GPT are we using?” into an answer someone can verify and maintain.

For practical model and workflow updates, join the newsletter.