Карта моделей OpenAI: GPT-5.5, GPT Image 2, Codex, Realtime, Audio и Sora 2

На 2026-05-08 GPT-5.5 Pro, GPT-5.5 Instant, GPT Image 2, GPT-5.3-Codex, gpt-realtime-1.5, gpt-audio-1.5, Sora 2 и Sora 2 Pro нельзя читать как один плоский список моделей. Это набор имен из разных поверхностей OpenAI: ChatGPT, API, Codex, Images, Realtime voice, REST-style audio и Sora video.

Практический порядок такой: сначала выбрать рабочую поверхность, затем найти публичное имя или ID для этой поверхности, и только потом сравнивать качество, latency, стоимость или доступность. GPT-5.5 Instant относится к ChatGPT; gpt-5.5 и gpt-5.5-pro относятся к API reasoning; GPT Image 2 maps to gpt-image-2; Codex может иметь отдельный picker; Realtime и Audio решают разные audio задачи; Sora требует отдельной проверки app access и legacy Videos API.

Production wording should stay conservative. If a decision depends on price, quota, plan, region, provider routing, refund policy, speed promise or API deprecation, do not preserve it from memory. Reopen the official page or the account console on the day of implementation.

GPT-5.5: Instant, API и Pro не являются одним handle

The easiest Russian-language mistake is to treat GPT-5.5 Instant as if it were a direct API model name. In ChatGPT, Instant is a product-facing mode and default experience. It tells a user what route the app is using; it does not by itself define the string that a backend should pass as the model parameter.

For API work, use the model pages and current endpoint guidance. gpt-5.5 and gpt-5.5-pro are public API reasoning names; chat-latest is the route mentioned in the Instant context. These names are related, but they answer different implementation questions. A codebase should not invent gpt-5.5-instant unless OpenAI documents that exact public ID for the relevant endpoint.

GPT-5.5 Pro should also be treated as a workload choice, not a universal upgrade. It is attractive for harder reasoning, long analysis, agentic planning and high-value responses, but it can change latency expectations and request handling. A team should know which prompts deserve Pro, which prompts can use standard GPT-5.5, and which prompts should not pay for heavier compute.

Troubleshooting follows the same split. A missing ChatGPT mode points to account, plan, rollout or model picker. A failed API call points to endpoint, organization access, billing, model availability and request shape. The names look similar, but the first diagnostic step is not the same.

Surface route: choose the product layer before the name

A model-name list becomes useful only after the product layer is separated. ChatGPT, API, Codex, Images, Realtime, Audio and Sora each have a different owner and a different failure mode. If the surface is wrong, a correct model name still produces the wrong implementation.

For a chat product, the first question is whether the backend needs gpt-5.5 or gpt-5.5-pro. For a user asking why ChatGPT changed, the first question is Instant, Thinking or Pro in the app. For an image workflow, the first question is gpt-image-2 and the image API shape. For a voice product, the first question is whether live interruption and streaming are required. For video, the first question is app route or legacy API risk.

This separation also keeps narrower topics from swallowing the map. Image price, free use, 4K output and cheapest provider claims need their own current evidence. Sora API setup needs a status page and a migration lens. Codex setup needs a route-specific guide because sign-in, API key, local CLI and cloud tasks can expose different choices.

Публичные ID: что можно безопасно писать в коде

Developer implementation should follow public IDs, not product labels. A product label can be helpful in a UI or status explainer, but code should use the current model name documented for that endpoint. The safest habit is to ask which surface exposes which ID before changing a request body.

This table is not a benchmark. It is a copy-safety table. It prevents a developer from pasting a ChatGPT display name into an API parameter, choosing an image model for a text workload, or starting a Sora backend without checking deprecations.

Codex: coding route depends on Codex, not only on the API catalog

Codex should be evaluated as a coding workflow. Public API availability is important, but Codex adds its own route: model picker, account sign-in, API-key authentication, local CLI behavior, IDE integration and cloud tasks. A model can be valid in one route and still unavailable or inappropriate in another.

When Codex exposes GPT-5.5, it is a strong first candidate for hard repository work. gpt-5.3-codex still matters when the task benefits from a Codex-tuned model or when the current route points there. The real comparison is not release date; it is patch quality, scope control, ability to read the repo, test interpretation and repair discipline.

A Russian-speaking engineering team should test Codex with real tasks: a small bug fix, a constrained refactor, a failing test, an API integration and a review pass. General chat quality is not enough. Good Codex behavior means the model changes the right files, explains the risk and avoids unrelated rewrites.

GPT Image 2: model name, workflow and claims are separate

GPT Image 2 has a clear developer ID, gpt-image-2, but that does not make every image question simple. Image generation, image editing, multi-image consistency and combined text-image reasoning can use different request shapes. A direct Images API call is not the same as a larger workflow that reasons, selects assets and then generates output.

The volatile part is pricing and access language. Free use, limits, exact cost, no-rate-limit claims, provider availability and failure billing are not safe to copy into a general model map. The map can say that image work starts with gpt-image-2; a price or provider page must recheck the current evidence.

Provider pages can be useful for discovery, but they do not become official OpenAI facts. Treat them as access layers. Official model identity, API behavior, deprecation status and safety caveats should come from OpenAI or the relevant first-party documentation.

Realtime 1.5 и Audio 1.5: voice session is not an audio job

gpt-realtime-1.5 is the route for low-latency interactive audio. It belongs in products where the user speaks, the agent can respond quickly, and the session needs interruption, streaming and state. This is a different engineering problem from sending one request with audio and receiving one response.

gpt-audio-1.5 is the better first inspection point for request/response audio jobs in Chat Completions. It can belong in pipelines, batch processing, audio understanding, audio answer generation and backend steps where persistent realtime interaction is not required.

Before migrating old voice code, check the deprecations page. Older preview names, audio endpoint assumptions and realtime session details can become stale quickly. The model choice should be tied to interface behavior, not merely to the word audio.

Sora 2 и Sora 2 Pro: приложение отдельно, API-риск отдельно

Sora is the highest-risk place to blur product access and API build readiness. A user may mean the Sora app or web product. A developer may mean sora-2 or sora-2-pro model pages. Those are not the same decision, especially when the API pages are marked legacy and the deprecations page lists the Videos API shutdown date as 2026-09-24.

This boundary should appear before any implementation advice. App access can be useful for creation, review and workflow planning. It does not prove that a new backend video product should be built on the same path. Backend work needs current support status, migration path, account access, region, quota, cost and output-rights checks.

A conservative Sora plan records two rows: consumer or creator workflow through the app, and API or automation workflow with deprecation risk. If the second row cannot name a supported current build path, the project should pause or choose a different route instead of treating app availability as backend evidence.

Какую модель OpenAI выбрать первой?

Start from the job. A chat assistant or backend agent begins with gpt-5.5 and moves to gpt-5.5-pro only when the workload deserves it. A ChatGPT user begins with Instant, Thinking or Pro in the app. An image workflow begins with gpt-image-2. A coding agent begins with Codex. A live voice product begins with Realtime. A REST-style audio job begins with Audio 1.5. A video creator begins with Sora app access, while an API builder begins with the legacy warning.

The safest final rule is simple: do not let a model-name announcement become an implementation contract. A production contract needs surface, endpoint, model ID, account access, deprecation state and a current check for volatile facts.

Часто задаваемые вопросы

GPT-5.5 Instant можно использовать как API model ID?

Treat it as a ChatGPT route label. For API calls, use the documented route such as chat-latest where applicable or public model names like gpt-5.5 and gpt-5.5-pro.

Когда нужен GPT-5.5 Pro?

Use it for harder reasoning, long analysis and high-value agentic work. Do not make it the default for every request before checking latency, cost and background handling.

GPT Image 2 и ChatGPT Images 2.0 — одно и то же?

No. GPT Image 2 is the model/API wording, with gpt-image-2 as the developer ID. ChatGPT Images 2.0 is an app-facing product route.

Почему Codex надо проверять отдельно?

Codex availability can depend on model picker, sign-in, API key route, local CLI, IDE integration and cloud tasks. API model visibility alone does not settle the Codex workflow.

Выбирать gpt-realtime-1.5 или gpt-audio-1.5?

Choose gpt-realtime-1.5 for live, interruptible voice sessions. Choose gpt-audio-1.5 for request/response audio work.

Sora 2 безопасен для API-продукта?

Only after a current deprecation and support-path check. Sora app access and legacy API model pages do not prove a stable new backend route.