OpenAI offers a broad range of models via the API for text generation, reasoning, coding, tool use, structured outputs, and document-centric workflows.
According to the official model overview, the current models support text and image input, text output, multilingual capabilities, and vision; they are available through the Responses API and client SDKs. For complex tasks, OpenAI generally recommends gpt-5.4; for lower latency and costs, OpenAI points to gpt-5.4-mini and gpt-5.4-nano
Open AI
LLM “Access our frontier models and APIs.”
Location: USA ⓘ OpenAI OpCo, LLC, 1455 3rd Street, San Francisco, CA 94158, USA
Batch / Flex / Priority / Scale Tier Options for cost and latency management for larger or predictable workloads.
Fine-Tuning / Evals / Tools / Agents Additional API features for customization, evaluation, agents, web search, File Search, Code Interpreter, Realtime, and structured outputs.
Data Residency / ZDR / EKM Enterprise-grade data controls with regional storage/processing, Zero Data Retention or Modified Abuse Monitoring, and external key management.
Target audience
OpenAI API LLMs are primarily aimed at development teams, SaaS providers, agencies, start-ups, internal automation teams, product organizations, and enterprise IT. For individual users without a technical build context, the API is less obvious than ChatGPT; however, for anyone looking to build their own applications, assistants, agents, automations, or document-centric workflows, it provides a very strong foundation. With its tiering from Nano to Mini to Frontier and Reasoning, OpenAI covers both high-volume and high-complexity scenarios.
Outstanding features
The greatest strength is the combination of model breadth and tool depth. OpenAI combines frontier models such as GPT-5.4 with more affordable variants like GPT-5.4 mini and nano, supplemented by reasoning models such as o3. The official docs also highlight Structured Outputs, Function Calling, Web Search, File Search, and, for selected models, Computer Use. For text-heavy business applications, the very large context windows of GPT-5.4 and GPT-4.1 are particularly interesting.
Key areas of application
OpenAI is particularly strong in coding, assistants, automation, document processing, internal knowledge systems, customer communication, research support, translation, and text generation. The model families are differentiated enough that you can choose an affordable model for simple classification or extraction and move to larger models for complex planning, multi-step reasoning, or high-quality output requirements. It is precisely this tiering that makes OpenAI so attractive for real-world API product development.
Usage & Notes
In practice, OpenAI is usually integrated via the Responses API or the client SDKs. For production use, model selection should be based not only on quality, but also on latency, context window, tool requirements, data protection requirements, and token price. It is also important to consider endpoint-specific data retention: by default, API data is not used for training, but depending on the endpoint, abuse monitoring logs may be retained for up to 30 days; Zero Data Retention or Modified Abuse Monitoring are available for eligible organizations. Data residency options are available for Europe, but only for supported or eligible configurations.
Calculate tokens and costs with the KIFOX Tokenizer
Brief Practical Context
For maximum quality in professional text, coding, reasoning, and agent applications: GPT-6 Astra
For new professional text/coding/agent applications:gpt-5.4, gpt-5.5 or gpt-5.6
For maximum quality on difficult tasks: gpt-5.5-pro.
For cost-effective, fast production workloads: gpt-5.4-mini.
For very low-cost classification, routing, and extraction: gpt-5.4-nano.
For image generation and image editing: gpt-image-2.
For live voice agents: gpt-realtime-2.
For live translation: gpt-realtime-translate.
For live transcription: gpt-realtime-whisper.
For coding agents: gpt-5.3-codex.
For RAG/knowledge bases: gpt-5.4-mini or gpt-5.4-nano plus text-embedding-3-large/text-embedding-3-small.
| Target group | Assessment |
|---|---|
| Developers / software teams | Very suitable – for chatbots, agents, structured outputs, code, tool calling, RAG, automation, multimodality, audio, images, and production AI applications. |
| SaaS providers / product teams | Very suitable – if AI is to be embedded directly into proprietary products, platforms, or workflows. |
| SMEs with IT resources | Suitable – for support automation, internal search, document analysis, content processes, and data extraction. |
| Large enterprises | Very suitable – due to the broad model portfolio, data residency options, ZDR/Modified Abuse Monitoring, Enterprise Key Management, and governance capabilities. |
| Individuals without a technical background | Rather unsuitable – ChatGPT is a better fit for them; the API requires technical integration. |
All Models
GPT/Reasoning/Text Models via API
GPT-6 Astra (gpt-6-astra). This is now OpenAI’s most capable general-purpose API model and is explicitly recommended for complex reasoning, coding, computer use, research, and document creation. Context window: 1.05 million tokens, maximum output: 128,000 tokens.
gpt-5.6 / gpt-5.6-sol – OpenAI’s current flagship model for complex professional tasks, demanding coding, in-depth analysis, scientific research, computer use, design, and agentic workflows. Particularly suitable for long-running, multi-step tasks involving extensive tool use, high autonomy, and complex reasoning. Supports the additional “max” reasoning level and offers greater performance and efficiency compared to GPT-5.5.
gpt-5.6-terra – Powerful and more cost-efficient model in the GPT-5.6 family. Suitable for professional knowledge work, software development, agents, research, data analysis, computer use, and extensive business workflows where high quality is to be combined with lower costs than GPT-5.6 Sol. OpenAI positions Terra as a balanced choice between intelligence, speed, and cost.
gpt-5.6-luna – The fastest and most affordable model in the GPT-5.6 family for cost-sensitive applications and high request volumes. Suitable for classification, extraction, routing, support automation, summarization, sub-agents, standardized coding tasks, and other clearly defined workflows. In terms of positioning, Luna roughly corresponds to the Nano model class of earlier GPT-5 generations, but offers significantly stronger reasoning and agent capabilities.
gpt-5.5-pro – For highly demanding professional tasks, complex analyses, difficult programming tasks, multi-step reasoning, strategy, architecture, legally/technically sophisticated drafts, and maximum response quality. According to OpenAI, GPT-5.5 pro is the more precise, computationally intensive variant of GPT-5.5 and is intended for difficult problems.
gpt-5.5 – Best general-purpose frontier model for complex professional work, coding, analysis, technical concepts, agents, RAG workflows, structured outputs, and high-quality text generation. OpenAI describes it as its latest frontier model for complex professional work.
gpt-5.4-pro – For very difficult professional tasks when higher precision is more important than speed. Particularly suitable for complex problem-solving, long analyses, deep reasoning, and difficult code/architecture questions.
gpt-5.4 – For professional standard and enterprise workflows with high quality, but more affordable than GPT-5.5. Suitable for coding, analysis, documentation, agents, knowledge systems, consulting texts, and structured business applications.
gpt-5.4-mini – For fast, cost-efficient applications with good quality: chatbots, assistants, sub-agents, coding assistance, classification, data extraction, support automation, and high request volumes. OpenAI describes it as a strong mini model for coding, computer use, and subagents.
gpt-5.4-nano – For very low-cost, fast high-volume processing: classification, ranking, simple data extraction, routing, pre-filtering, tagging, short summaries, and sub-agents. OpenAI describes it as the lowest-cost GPT-5.4-class model for simple high-volume tasks.
gpt-5.3-chat-latest – A ChatGPT-like model that, according to OpenAI, points to the GPT-5.3-Instant snapshot. Suitable when ChatGPT-like response behavior is desired via the API; not the first choice for new technical systems if GPT-5.5 or GPT-5.4 are available.
gpt-5.2-pro – Previous Pro model for professional work. Suitable for highly complex tasks if GPT-5.5 pro is deliberately not used for compatibility, cost, or stability reasons.
gpt-5.2 – Previous frontier model for professional work with configurable reasoning. Suitable for existing applications optimized for GPT-5.2, as well as for complex analysis, coding, and agent workflows.
gpt-5.1 – Older GPT-5 model for coding and agentic tasks. Suitable for existing applications that are still optimized for GPT-5.1.
gpt-5-pro – Pro version of GPT-5. Suitable for more complex tasks than GPT-5, especially when greater computational effort and response precision are desired.
gpt-5 – Previous intelligent reasoning model for coding, agents, analysis, and general demanding text tasks. Today, it is more relevant for compatibility and existing implementations.
gpt-5-mini – Faster and more affordable GPT-5 variant for clearly defined tasks, precise prompts, high request volumes, chatbots, simple agents, and standard automation.
gpt-5-nano – Fastest and most affordable GPT-5 variant. Suitable for classification, summarization, simple extraction, routing, tagging, and very high volumes.
gpt-4.1 – Strong non-reasoning model. Suitable for fast, high-quality text generation, coding, instruction following, long contexts, and general API applications when deep reasoning is not required.
gpt-4.1-mini – Smaller and faster GPT-4.1 variant. Suitable for production chatbots, support, content creation, classification, and cost optimization.
gpt-4.1-nano – Very fast, inexpensive GPT-4.1 variant, now deprecated according to OpenAI. Only suitable for existing workflows, simple classification, and bulk processing.
gpt-4o – Multimodal GPT model for text and image understanding, fast chatbots, assistant systems, analysis of images/screenshots, and general applications. Still relevant for existing projects.
gpt-4o-mini – Affordable smaller GPT-4o variant for focused tasks, simple chatbots, classification, short texts, and high volumes. Still relevant for older implementations.
gpt-4-turbo – Older GPT-4 model, now more of a legacy option. Suitable only for existing applications that have deliberately not been migrated.
gpt-4 – Older high-intelligence model. Today primarily legacy/compatibility.
gpt-3.5-turbo – Legacy model for low-cost chat and text tasks. Today only useful for legacy systems.
o3-pro – More compute-intensive variant of o3 for better responses. Suitable for very difficult reasoning tasks when an older o-series model is required.
o3 – Reasoning model for complex tasks; according to OpenAI, now superseded by GPT-5. Suitable for existing applications optimized for o3.
o4-mini – Fast, cost-efficient reasoning model; according to OpenAI, now deprecated and replaced by GPT-5 mini. Suitable only for existing applications, quick reasoning tasks, coding, and visual tasks.
o3-mini – Older small reasoning model. Today only legacy/compatibility.
o1-pro – Older Pro variant of o1 for difficult reasoning tasks. Deprecated/legacy.
o1 – Earlier o-Series reasoning generation. Deprecated/legacy.
o1-mini – Earlier small o-Series variant. Deprecated/legacy.
o1-preview – Early preview version of the o-Series. Deprecated/legacy.
Coding Models
OpenAI offers its own Codex/coding models for software development and agentic coding tasks.
gpt-5.3-codex – Currently an important coding model for agentic software development, Codex-like workflows, refactoring, bug fixing, complex codebases, pull request work, and longer coding tasks. OpenAI describes it as a particularly powerful agentic coding model.
gpt-5.2-codex – Predecessor/legacy model for long agentic coding tasks, complex code changes, and software development.
gpt-5-codex – Older GPT-5-Codex model for agentic coding. More of a legacy model today.
gpt-5.1-codex – Older Codex model for coding agents and existing workflows. Deprecated/legacy.
gpt-5.1-codex-max – Variant for longer-running coding tasks. Deprecated/legacy.
gpt-5.1-codex-mini – Smaller, more affordable Codex variant. Deprecated/Legacy.
codex-mini-latest – Fast older reasoning model for Codex CLI. Deprecated/legacy.
Image Models
OpenAI lists GPT Image 2, GPT Image 1.5, chatgpt-image-latest, GPT Image 1, gpt-image-1-mini, as well as DALL·E 3 and DALL·E 2 in the API model overview.
gpt-image-2 – Current state-of-the-art image model for high-quality image generation and image editing. The official API name is GPT Image 2, not “GPT Image 2.0”. Suitable for realistic images, product images, illustrations, marketing graphics, image variations, inpainting/editing, and professional visual content.
gpt-image-1.5 – Previous image generation model. Suitable for existing workflows optimized for GPT Image 1.5.
chatgpt-image-latest – Previous image model from the ChatGPT context. Suitable when ChatGPT-like image output is desired; for new API projects, prefer GPT Image 2.
gpt-image-1 – Previous image generation model, now deprecated. Only relevant for legacy applications.
gpt-image-1-mini – Cost-efficient image model variant. Suitable for lower-cost image generation, simple variations, drafts, thumbnails, and scalable image workflows.
dall-e-3 – Older image generation model, now deprecated. Only relevant for existing projects.
dall-e-2 – First older DALL·E generation, now deprecated. Legacy only.
Realtime, Audio, and Speech Models
OpenAI lists realtime and audio models for live voice interaction, speech-to-speech, transcription, text-to-speech, and audio workflows.
gpt-realtime-2 – Currently the most important realtime language model for live voice agents, real-time dialogues, callbots, support assistants, voice-controlled agents, tool calling during conversations, and more complex live interactions. OpenAI describes it as a reasoning model for realtime voice interactions.
gpt-realtime-translate – Specialized model for live speech-to-speech translation. Suitable for multilingual real-time communication, customer support, education, meetings, international teams, and interpreting workflows. OpenAI describes it as a Streaming Speech-to-Speech Translation Model; the current announcement mentions real-time translation from over 70 input languages into 13 output languages.
gpt-realtime-whisper – Streaming speech-to-text model for live transcription. Suitable for meeting captions, live subtitles, call transcription, minutes, voice notes, and documentation workflows.
gpt-realtime-1.5 – Very good voice model for audio input/audio output. Suitable for live voice assistants, call center prototypes, voice UX, and interactive voice dialogues.
gpt-realtime – Realtime model for text and audio inputs as well as audio output. Suitable for older Realtime implementations and existing projects.
gpt-realtime-mini – Cost-efficient realtime variant. Suitable for lower-cost voice agents, simple voice dialogues, high request volumes, and prototypes.
gpt-audio-1.5 – Audio input/output model for Chat Completions-based audio workflows. Suitable for applications that do not necessarily require WebRTC/Realtime sessions.
gpt-audio – Audio model for audio inputs and audio outputs via Chat Completions. Suitable for legacy audio workflows, voice assistants, and multimodal audio apps.
gpt-audio-mini – Cost-efficient audio variant. Suitable for simpler audio tasks and scalable audio workloads.
gpt-4o-audio – Older/deprecated GPT-4o audio model. Only suitable for existing implementations.
gpt-4o-mini-audio – Older/deprecated smaller GPT-4o audio model. Only suitable for existing projects.
gpt-4o-realtime – Older realtime model for text and audio input as well as audio output. Suitable for existing realtime apps.
gpt-4o-mini-realtime – Smaller older Realtime variant. Deprecated/legacy.
gpt-4o-transcribe – GPT-4o-based speech-to-text model. Suitable for high-quality transcription, audio analysis, subtitles, meetings, interviews, and call analysis.
gpt-4o-mini-transcribe – Smaller, more affordable transcription variant. Suitable for high volumes, simple transcription, and cost-sensitive speech-to-text workflows.
gpt-4o-transcribe-diarize – Transcription model with speaker diarization. Suitable for interviews, meetings, conversation transcripts, and call center analysis when it is necessary to distinguish who is speaking when.
gpt-4o-mini-tts – Text-to-speech model based on GPT-4o-mini. Suitable for natural-sounding speech output, chatbot read-alouds, voice UX, learning content, and simple audio outputs.
tts-1 – Older text-to-speech model, optimized for speed. Suitable for fast TTS output.
tts-1-hd – Older text-to-speech model, optimized for quality. Suitable for higher-quality speech output in legacy workflows.
whisper – General speech recognition model. Suitable for classic transcription, audio-to-text, translation/transcription of older workflows and legacy applications.
Deep Research Models
o3-deep-research – Former Deep Research model for intensive research tasks, source work, and multi-step information analysis. According to the model overview, deprecated.
o4-mini-deep-research – Faster/cheaper Deep Research variant. According to the model overview, deprecated.
Open-Weight Models
gpt-oss-120b – Open-weight model under the Apache 2.0 license. Suitable for self-hosting, proprietary infrastructure, research, customization, and scenarios where model weights are relevant. OpenAI describes it as the most powerful open-weight model that fits on an H100 GPU.
gpt-oss-20b – Smaller open-weight model. Suitable for lower latency, local/self-hosted applications, experiments, and more resource-efficient deployments.
Additional API models that are often forgotten
computer-use-preview – Specialized model for computer use/browser/GUI automation. According to OpenAI, deprecated; only relevant for existing workflows.
gpt-4o-search-preview – Older GPT model for web search in Chat Completions. Deprecated; now generally replaced by web search tools/Responses workflows.
gpt-4o-mini-search-preview – Small older Search Preview variant. Deprecated/Legacy.
omni-moderation – Moderation model for detecting potentially harmful content in text and images. Suitable for safety checks, user-generated content review, and compliance filters.
text-moderation – Older text moderation model. Deprecated/Legacy.
text-moderation-stable – Older stable text moderation variant. Deprecated/legacy.
text-embedding-3-large – Powerful embedding model. Suitable for semantic search, RAG, vector indexes, similarity search, knowledge bases, and clustering.
text-embedding-3-small – More affordable embedding model. Suitable for scalable RAG systems, search functions, classification, and semantic similarity at lower cost.
text-embedding-ada-002 – Older embedding model. Today mainly relevant for legacy vector indexes.
sora-2 – Video generation model with synchronized audio; according to the model overview, deprecated.
sora-2-pro – Advanced Sora-2-Pro variant; according to the model overview, deprecated.
Hosting & Data
1) On-prem / local hosting
Meaning: The company operates the solution on its own hardware or within its own infrastructure. In the strictest sense, not only the application runs locally, but ideally the model as well.
2) Private cloud / data center
Meaning: The solution runs in a dedicated or more clearly separated cloud environment, often with a hosting provider or hyperscaler, but in a German data center or in a particularly controlled environment.
3) EU SaaS / managed
Meaning: The provider operates the solution itself as a service. The company uses the tool as a ready-made cloud service, ideally with EU data residency.
4) Hybrid
Meaning: One part of the processing remains internal / local / in a private cloud, while another part runs in an external cloud or EU SaaS.
5) AVV / DPA
Meaning: This is the data processing agreement or Data Processing Addendum. It governs that the provider processes personal data on behalf of the customer and is bound by the customer's instructions.
6) No training
Meaning: The provider does not use your prompts, uploads, attachments, chat histories, or outputs for training or improving the general model — ideally excluded by contract.
7) Open-source / transparency path
Meaning: There is a path toward greater technical transparency and sovereignty, for example through:
- open models
- documented components
- self-hostable parts
- traceable architecture
- export / switching options
| On-prem / local hosting | ❓ |
| Private cloud / data center | ⚠️ |
| EU SaaS / Managed | ✅ |
| Hybrid | ⚠️ |
| DPA / AVV | ❓ |
| No training on customer data | ✅ |
| Open source / transparency path | ⚠️ |
On-Prem / local hosting: indirect / not available
The website does not document any on-premise, self-hosting, or locally self-operated model option for the OpenAI API. Instead, it documents a hosted API service.
Private Cloud / data center: partial
Data residency controls per project are documented, including a European region with regional storage and regional processing for supported services. However, a dedicated private cloud or single-tenant environment is not explicitly described on the website.
EU SaaS / Managed: covered
The website documents an OpenAI-operated SaaS/API option with EU data residency for 'Europe (EEA + Switzerland)' via 'eu.api.openai.com'. For supported services, regional storage and regional processing are specified, though only under conditions such as approval and additional controls.
Hybrid: partial
The website mentions local execution in individual contexts, such as 'Skills' in local execution alongside hosted container-based execution. This describes a technical mixed path for partial components, but not a fully documented general hybrid operating model for the entire core solution.
DPA / DPA: unclear
I could not find a DPA or a corresponding agreement on the specified domain. Not specified on the website.
No training: covered
In the data controls table, 'Data used for training' is listed as 'No' for the listed API endpoints in each case. In addition, 'Zero Data Retention' and 'Modified Abuse Monitoring' are described, which can further restrict abuse logging, subject to approval and conditions.
Open source / transparency path: partial
There is a documented transparency path through open-source Codex components on GitHub, including Codex CLI, Codex SDK, Codex App Server, Skills, and the 'Universal cloud environment'. At the same time, according to the website, central parts such as 'IDE extension' and 'Codex web' are not open source; a self-hostable end-to-end path for the OpenAI API is not documented.
Data processing
The website describes the OpenAI API as a hosted service. For many endpoints, it is documented that API data is not used for training. By default, depending on the endpoint, there are retention periods for abuse monitoring and in some cases application state; for '/v1/responses', a 30-day application-state storage period is described by default, unless restricted by appropriate controls. For European data residency, the website names the region 'Europe (EEA + Switzerland)' with regional storage and, for many supported services, also regional processing. At the same time, according to the website, data residency does not include 'system data', certain functions have limitations, and data sent to third parties such as remote MCP servers is subject to their own policies.
Conclusion
For the EU/EEA area, based on the documented best-case option, OpenAI is not universally usable, but conditionally usable: positives include EU data residency, regional processing for many services, and documented no-training/retention controls. Negatives include the dependency on approval, the lack of concrete information on a DPA, missing subprocessors information on the domain, the lack of an on-prem option for the core API, as well as several functional limits and exceptions. For a European directory, 'conditional' is therefore the most defensible classification.
Sources
Strengths & weaknesses at a glance
| Strengths | Weaknesses |
|---|---|
| - Very broad model coverage, from affordable to frontier. | - The portfolio is complex; model selection, pricing tiers, context limits, and tool costs require explanation. |
| - Strong suitability for coding, agents, tool calling, structured outputs, and long contexts. | - The most powerful models are significantly more expensive than Mini/Nano variants. |
| - According to OpenAI, API/business data is not used for training on inputs/outputs by default. | - Data protection and data residency options are not identical across all cases, but are sometimes tied to the organization type, endpoint, or enablement. |
| - Data residency, Zero Data Retention, and DPA are | - Older model families that remain available increase operational complexity in selection and lifecycle management. |
Reviews
0 reviews in total
There are no confirmed reviews for this tool yet.
Submit review
Your review will only become visible after email confirmation. This protects the portal against abuse.
Report review
Please select the reason why this review should be checked.
GDPR-compliant usage possible?
For users in the European region, GDPR-compliant use is generally possible under the best available option documented on the website, but only under certain conditions. Positively, OpenAI describes EU data residency for 'Europe (EEA + Switzerland)' on its developer website, with regional storage and regional processing for supported services, and also documents project-specific data controls such as 'Zero Data Retention' and 'Modified Abuse Monitoring'. At the same time, according to the website, restrictions apply: data residency is eligibility-based, subject to additional requirements for non-US regions, does not apply to system data, and does not apply equally to all features.
Positive
The website documents EU data residency for 'Europe (EEA + Switzerland)' via 'eu.api.openai.com', regional storage and regional processing for many supported API services, as well as data controls according to which API content is not used for training according to the table. In addition, 'Zero Data Retention' and 'Modified Abuse Monitoring' are available as configurable controls following approval.
Negative
The website does not name a specific EU server location at country level or a particular data center. Data residency is not freely available by default, but only following review or approval. It expressly does not apply to 'system data', and certain features are restricted or cannot be used in a GDPR-compliant manner, such as notes regarding Realtime Tracing or certain storage-related or third-party-adjacent features. I was unable to find a Data Processing Agreement (DPA) or a list of subprocessors on the specified domain.
Server location
For data residency, the website names the region 'Europe (EEA + Switzerland)' with the endpoint 'eu.api.openai.com'. The website does not specify a particular country, server location, or data center within the EU/EEA.
