GPT (Generative Pre-trained Transformer) is the AI model family from OpenAI, the company behind ChatGPT, Codex and the OpenAI API. The current reference is the GPT-6 generation with four tiers: Astra (top model), Sol (complex coding and agentic workflows), Terra (fast model for image analysis, computer use and writing) and Luna (efficient model for focused, high-volume tasks). The newest models are GPT-6.1 Terra (gpt-6.1-terra, 26 October 2026) and GPT-6.1 Sol (gpt-6.1-sol, 29 September 2026); the previous GPT-5.6 family (Sol, Terra, Luna) remains available as a fallback.
For enterprises in the DACH region, EU availability is the deciding factor: Microsoft Foundry lists GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Sol, GPT-6 Luna and GPT-5.6 as Data Zone Standard (EU) in all nine European regions, and GPT-6 Astra in ten (North Europe in addition). Amazon Bedrock offers the OpenAI models without an EU-resident option. As an AI consulting firm based in Rosenheim, Germany, we help enterprises integrate OpenAI models in a GDPR-compliant way; with CompanyGPT you can use GPT models securely. This overview reflects the state as of October 2026.
Current models at a glance
| Model | API ID | Release | Context | Price input / output per 1M tokens | EU Data Zone (Foundry) | Focus |
|---|---|---|---|---|---|---|
| GPT-6.1 Sol | gpt-6.1-sol | 29 Sep 2026 | 1.05M | $2.00 / $10.00 (cached $0.10) | yes | Complex coding, agentic workflows |
| GPT-6.1 Terra | gpt-6.1-terra | 26 Oct 2026 | 1.05M | $1.00 / $5.00 (cached $0.05) | yes | Image analysis, computer use, writing |
| GPT-6 Luna | gpt-6-luna | 22 Sep 2026 | 1.05M | $0.10 / $0.50 (cached $0.01) | yes | High-volume workloads, RAG, classification |
| GPT-6 Sol | gpt-6-sol | 22 Sep 2026 | 1.05M | $2.00 / $10.00 (cached $0.20) | yes | Coding and agents (superseded by 6.1 Sol) |
| GPT-6 Astra | gpt-6-astra | 3 Sep 2026 | 1.05M | $10.00 / $50.00 (cached $1.00) | yes | Top model, computer use |
| GPT-5.6 Sol | gpt-5.6-sol | 9 Jul 2026 | 1.05M | $4.00 / $20.00 (promotional) | yes | Proven fallback |
| GPT-5.6 Terra | gpt-5.6-terra | 9 Jul 2026 | 1.05M | $2.00 / $12.00 | yes | Superseded by 6.1 Terra |
| GPT-5.6 Luna | gpt-5.6-luna | 9 Jul 2026 | 1.05M | $0.20 / $1.20 | yes | Fast and low-cost |
Prices per OpenAI (first party, standard up to 272k input tokens). All models accept text and images and output text.
GPT-6.1 Sol
GPT-6.1 Sol (gpt-6.1-sol) is an upgrade of GPT-6 Sol. Per OpenAI it comes close to GPT-6 Astra on agentic coding, computer use and knowledge work – at one-fifth of its standard token prices. The figures OpenAI cites:
| Benchmark (per OpenAI) | GPT-6.1 Sol | GPT-6 Sol | GPT-6 Astra |
|---|---|---|---|
| DeepSWE v1.1 | 75.2% (reasoning high) | 68.8% (max) | roughly on par, 6.1 Sol at about one-fifth of the cost per task |
| OSWorld 2.0, offline set (max) | 71.4% | – | 73.5% (6.1 Sol at about one-seventh of the cost per task) |
| AutomationBench (medium) | 31.7% | 26.9% | – |
| Factual-question error rate (reasoning low) | 7.7% | 11.4% | within 1.9 points across all levels |
OpenAI also cites better document understanding in multi-step workflows, more reliable code writing and debugging, and fewer cases where the model fails to report broken search tools or ignores explicit constraints. Independent measurements were not available at launch.
Technical details per the OpenAI model page: context window 1,050,000 tokens (922k input, 128k output), knowledge cutoff 30 April 2026, text and image input, text output. Five reasoning levels (low, medium as default, high, xhigh, max) – unlike GPT-6 Sol no none; OpenAI recommends low as the replacement when migrating. Endpoints: Responses, Chat Completions, Batch; no Realtime, no fine-tuning, no predicted outputs. The tool catalog matches the GPT-6 family; multi-agent features via the Responses API are in beta. Per the system card, OpenAI rates GPT-6.1 Sol like Astra as “Critical” for cybersecurity and “High” for bio/chem.
Pricing: $2.00/1M input, $0.10 cached input, $2.50 cache write, $10.00/1M output – the same as GPT-6 Sol, but with half the cached-input price. Above 272k tokens, $4.00 / $15.00 applies.
Assessment: GPT-6.1 Sol is our recommendation for complex coding and agentic workloads – same price as GPT-6 Sol, better benchmarks per OpenAI, EU Data Zone in Foundry from day one. When migrating from GPT-6 Sol, only the reasoning effort none needs replacing.
GPT-6.1 Terra
GPT-6.1 Terra (gpt-6.1-terra, 26 October 2026) closes the gap between GPT-6.1 Sol and GPT-6 Luna. OpenAI describes it as a fast, efficient model for visual reasoning, precise computer use and clear, natural writing. Typical tasks per OpenAI: matching product photos to catalog entries, comparing technical drawings, acting on application screens, and drafting and revising text.
Performance (per OpenAI and AWS): On Agents’ Last Exam, a test of how well agents use software at work, Terra scores 15 percentage points higher than GPT-5.6 Terra and on par with GPT-6 Astra per the AWS model card. OpenAI adds that Terra reaches this level at one-tenth of Astra’s price per token and finishes tasks on average more than three times faster than Astra and GPT-6.1 Sol. OpenAI did not publish further benchmark figures at launch, and independent measurements were not available. For in-depth research and complex coding, OpenAI itself points to GPT-6.1 Sol and Astra.
Technical details per the OpenAI model page: context window 1,050,000 tokens, 128k output, knowledge cutoff 1 June 2026, text and image input, text output. Six reasoning levels from none to max (default medium); OpenAI recommends starting with medium and increasing it only if the quality gain is worth the added time and tokens. Endpoints: Responses, Chat Completions and Batch; tool calling requires the Responses API. No fine-tuning, no Realtime, no Ultrafast mode. Per the system card addendum, OpenAI rates Terra like Astra and 6.1 Sol as “Critical” for cybersecurity and “High” for bio/chem and uses the same safeguards.
Pricing: $1.00/1M input, $0.05 cached input, $1.25 cache write (30 minutes), $5.00/1M output, half the price of GPT-6.1 Sol and GPT-5.6 Terra. Above 272k tokens, $2.00 / $7.50 applies.
Availability: in the OpenAI API, in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu, and from day one in Microsoft Foundry (Data Zone Standard EU across all nine European regions) and on Amazon Bedrock (without an EU-resident option). In the OpenAI API, EU data residency also works in fast mode.
Assessment: Terra is the new middle ground for tasks where Luna reaches its limits and Sol is oversized: image and document analysis, agents that operate applications through their interface, and writing. For applications on GPT-5.6 Terra, switching pays off. OpenAI recommends comparing both models with the same reasoning effort on representative images and workflows and moving tool calls from Chat Completions to the Responses API.
GPT-6 Luna
Per OpenAI, GPT-6 Luna (gpt-6-luna) is the “most efficient model for focused, high-volume tasks”. It offers the same context window as Astra (1.05M tokens, 128k output), full Responses API tool support (web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, tool search) and six reasoning levels from none to max. Knowledge cutoff: 18 May 2026. Function calling via Chat Completions works only with reasoning_effort: "none"; tool use with reasoning runs through the Responses API. None of the GPT-6 models supports fine-tuning.
Luna is the economical choice for chat assistants, classification, extraction, summarization, RAG and sub-agents. Its reasoning depth is below Sol and Astra; for demanding analysis and coding tasks, compare it against the larger tier. A special case is the Decisions API (see below), which focuses Luna on finite answer lists.
GPT-6 Sol
GPT-6 Sol (gpt-6-sol, knowledge cutoff 20 April 2026) was released on 22 September 2026 together with Luna and offers the same features and prices as GPT-6.1 Sol, but with cached input at $0.20 and the additional reasoning level none. OpenAI’s model page now points to GPT-6.1 Sol as the newer Sol model without deprecating GPT-6 Sol. New projects should start directly on 6.1 Sol.
GPT-6 Astra
GPT-6 Astra (gpt-6-astra) is the top model of the sixth generation, introduced on 3 September 2026; OpenAI describes it as its most intelligent and best-aligned model to date. Access started through the Trusted Access Program (Daybreak, for cybersecurity defenders) and, per the OpenAI developer documentation, is now regularly available through the API and in ChatGPT (Pro, Business and Enterprise directly, Plus via ChatGPT Work and Codex). Microsoft Foundry lists the model without an access restriction; on Amazon Bedrock it has been available since 8 September 2026.
Performance (per OpenAI): ARC-AGI-3 98.6%, FrontierMath Tier 4 v2 97.6%, GPQA Diamond 96%, DeepSWE v1.1 74.1%, BenchCAD 95.9%, OSWorld 2.0 72.6% at roughly 40 minutes per task (GPT-5.6 Sol: 65.7% at 75 minutes). The focus is computer use: Astra operates browsers, spreadsheets, websites and desktop applications through the human interface instead of APIs. On DeepSWE, OpenAI cites roughly 57% lower estimated API cost per task compared with the best GPT-5.6 Sol configuration. Independent measurements were not available at launch.
Technical details: context window 1,050,000 tokens (922k input, 128k output), knowledge cutoff 30 April 2026. Five reasoning levels (low to max, no none). Endpoints: Responses, Chat Completions, Batch; no Realtime, no Assistants API, no fine-tuning. Custom temperature and top_p values and logprobs are not supported; tool calling runs exclusively through the Responses API. New in the Responses API are asynchronous tool calls, mid-turn control via WebSocket and switching the reasoning effort mid-conversation while preserving the prompt cache.
Safety: Per OpenAI, Astra is the first model rated “Critical” for cybersecurity in the Preparedness Framework. The generally available variant refuses advanced cybersecurity tasks, and monitoring can slow or pause legitimate work; advanced cyber capabilities are available only through Daybreak Blue for vetted defenders. Teams using Astra in security or development should plan for these interventions in their operating concept. OpenAI did not release the GPT-6.1 Astra that had been expected for late September: per media reports (Wall Street Journal, CNBC, TechCrunch) it did not meet internal requirements on task scope and authorization; OpenAI has not named a new date. Governance questions for agentic systems – authorization, logging, duty to ask for confirmation – belong in every operating concept regardless of the model.
Pricing and tiers: Batch and Flex cost 50%, Fast Mode double ($20.00 / $100.00). Since 29 September 2026 Astra has an Ultrafast tier (service_tier: "ultrafast"): per OpenAI up to six times faster token output at six times the price ($60.00 / $6.00 / $300.00 per 1M tokens), in the OpenAI API only with global processing or US data residency, and per the AWS model card also on Bedrock.
GPT-5.6 (Sol, Terra, Luna)
The GPT-5.6 family has been generally available since 9 July 2026 (ChatGPT, Codex, API). The naming scheme separates the generation number from lasting performance tiers; mini/nano variants no longer exist. Sol (gpt-5.6-sol, alias gpt-5.6) is the flagship (Terminal-Bench 2.1: 88.8%, 91.9% with the Ultra configuration, per OpenAI), Terra (gpt-5.6-terra) the balanced tier, Luna (gpt-5.6-luna) the fast, low-cost tier – all three with 1,050,000 tokens of context (922k input, 128k output), knowledge cutoff 16 February 2026, six reasoning levels (none to max) and reasoning.mode: pro as the replacement path for the former Pro models. Programmatic tool calling, extended multi-agent capabilities and prompt-cache breakpoints were added. GPT-5.6 also launched ChatGPT Work, a Sol-powered work agent with Codex integration.
Since 21 August 2026, GPT-5.6 Sol has been priced at a promotional $4.00 / $20.00 (previously $5.00 / $30.00); OpenAI guarantees this price until at least 21 November 2026, so plan longer-term budgets with the list price. Fast mode also supports prompts above 272k tokens for all three tiers (per OpenAI up to 2.5x faster); Sol additionally has an Ultrafast tier in limited preview (per OpenAI up to 14x faster).
In ChatGPT, per media reports, GPT-5.6 Luna has been the default model for Free and Go since 6 August 2026, and GPT-5.6 Sol the model for Plus, Pro, Business and Enterprise. Per the API changelog, the chat-latest snapshot points to the current ChatGPT model and was last updated on 7 October 2026; for production OpenAI recommends the GPT-6 family. At launch, GPT-5.6 was a government-cleared limited preview from 26 June to 8 July 2026: the US Department of Commerce (Center for AI Standards and Innovation) could review the model in advance because of its “High” cybersecurity rating. More in our blog post on GPT-5.6 and Claude Fable 5.
More models and APIs
Decisions API and DevDay features
Per OpenAI, the Decisions API (POST /v1/decisions) has been generally available since 22 October 2026 (public beta since 6 October). It currently supports only GPT-6 Luna and returns typed answers instead of free text. There are three question types: predicate (probability from 0 to 1 that a condition is true), choice (a pick from your own options with a probability distribution) and score (a probability-weighted value on ordered levels). Inputs are text, images, JSON (input_json) and conversation histories with tool calls; a request takes up to 400 questions (beta: 200). OpenAI calls the API roughly ten times faster than the Responses API (at DevDay: around 150 milliseconds versus 1.6 seconds).
- Pricing: uncached input $0.10 per 1M tokens, cache read and write rates as for GPT-6 Luna; output tokens are not charged. Regional-processing premiums and long-context multipliers still apply.
- Data protection: Zero Data Retention and HIPAA use are supported for eligible customers, data residency in all supported API regions, regional processing in the United States and Europe (EEA + Switzerland). Private Link is available for Decisions requests.
- Prompt caching: shared input is cached from a 1,024-token prefix; images can be attached per question.
- GPT-Live: since 15 October 2026 there is a native Decisions integration in GPT-Live, so applications can react while the user is still speaking.
This places OpenAI in the category of decision models that we also document with Microsoft-Decision-1, Cloudflare Clef, Jev by TypeSafe and Laya. The approach differs: the Decisions API calls a general-purpose language model, while the others are models trained specifically for decisions. Which route fits on accuracy, latency, cost and hosting can only be judged with your own labelled data and your data-protection requirements.
Further announcements from DevDay on 29 September 2026:
- Agents API (public beta): hosted execution, memory, tools and multi-agent support; per the OpenAI changelog now with computer use in an OpenAI-hosted browser. Per DevDay material, Managed Agents also runs inside Amazon Bedrock.
- Private Intelligence: Zero Data Retention with Private Safety Processing is available; a preview of Private Inference is announced for autumn 2026.
- Dots (persistent agents in ChatGPT, beta for Pro, Business and Enterprise) are per OpenAI not available in the European Economic Area, Switzerland and the United Kingdom.
- Codex: Codex Cloud runs tasks while the laptop is closed; the CLI gains voice control and an agent view.
- Performance: OpenAI cites 45% shorter time-to-first-token and over 30% faster tool calls in the API.
- Usage tiers: On 6 October 2026, OpenAI simplified API usage tiers from five to three (Build, Launch, Grow); rate limits depend on the respective tier.
Realtime, voice and transcription
- GPT-Realtime-2.1 (
gpt-realtime-2.1) and GPT-Realtime-2.1-mini are the current realtime voice generation and replace GPT-Realtime-2. - GPT-Live 1 (generally available since 10 September 2026) for full-duplex voice conversations: $0.05 per minute plus backend model and tools. Per the OpenAI changelog, voices can be created from a text description (prompt-to-voice), and nine more natural voices are available, including the German voices Nuri, Juni and Flitz.
- GPT-Realtime-Translate: live translation in real time, 70+ input languages, 13 output languages, $0.034/min.
- Transcription:
gpt-transcribe($0.0045/min, batch),gpt-live-transcribe($0.017/min, streaming) andgpt-realtime-whisper(live speech-to-text in the Realtime API, $0.017/min). Details on the open-source model: OpenAI Whisper.
Embeddings, images and cybersecurity
- Embeddings:
text-embedding-3-large($0.13/1M tokens) andtext-embedding-3-small($0.02/1M tokens); no successor is announced. - Image generation:
gpt-image-2(since 21 April 2026;gpt-image-1was shut down on 23 October 2026), complemented since 8 September 2026 by GPT Image 2.5 Sunburst (editing precision) and GPT Image 2.5 Flare (fast generation); both support the quality levelsxhighandmaxand run at gpt-image-2 token prices. - Cybersecurity models: The tiers Daybreak Blue and Daybreak Red (
gpt-daybreak-red-latest) have been available to cleared defenders since 7 August 2026; on Bedrock, GPT-5.6 Cyber and GPT-5.6 Sol (Daybreak) are available only In-Region in us-east-2. Access requires clearance;gpt-5.4-cyberwas replaced bygpt-5.6-cyberon 1 October 2026. - Video: Sora 2 and the Videos API were shut down on 24 September 2026 without a successor, see OpenAI Sora.
Pricing
All prices in USD per 1 million tokens, OpenAI first party, standard up to 272k input tokens:
| Model | Input | Cached input | Cache write | Output | Input / output above 272k tokens |
|---|---|---|---|---|---|
| GPT-6 Astra | $10.00 | $1.00 | $12.50 | $50.00 | $20.00 / $75.00 |
| GPT-6.1 Sol | $2.00 | $0.10 | $2.50 | $10.00 | $4.00 / $15.00 |
| GPT-6.1 Terra | $1.00 | $0.05 | $1.25 | $5.00 | $2.00 / $7.50 |
| GPT-6 Sol | $2.00 | $0.20 | $2.50 | $10.00 | $4.00 / $15.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.125 | $0.50 | $0.20 / $0.75 |
| GPT-5.6 Sol (promotional) | $4.00 | $0.40 | $5.00 | $20.00 | $8.00 / $30.00 |
| GPT-5.6 Terra | $2.00 | $0.20 | $2.50 | $12.00 | $4.00 / $18.00 |
| GPT-5.6 Luna | $0.20 | $0.02 | $0.25 | $1.20 | $0.40 / $1.80 |
Prices per provider. Batch and Flex cost 50%, Fast Mode double, regional processing (data residency) +10%. The cache write column applies to 30-minute retention; since 20 October 2026 the API also offers 12-hour caching, whose cache writes cost 2.5 times the input price. Ultrafast mode at six times the price is available only for GPT-6 Astra and GPT-6.1 Sol ($12.00 / $0.60 / $60.00). Prompts above 272,000 tokens are billed at twice the input price and 1.5 times the output price. OpenAI argues that comparisons should be made on price per completed task rather than per token; for budget planning we recommend measuring exactly that on your own tasks. In Microsoft Foundry, prices differ in part (Astra Data Zone US: $11.00 / $55.00).
EU availability and GDPR
Microsoft Foundry (Azure OpenAI)
Per the Microsoft Learn region table, GPT-6.1 Sol, GPT-6.1 Terra (since 26 October 2026), GPT-6 Sol, GPT-6 Luna and all three GPT-5.6 tiers are listed as Data Zone Standard (EU) in all nine European Foundry regions: France Central, Germany West Central (Frankfurt), Italy North, Norway East, Poland Central, Spain Central, Sweden Central, Switzerland North and West Europe. Prompts and responses are processed within the EU data zone; data at rest stays in the Azure geography. GPT-5.5 and GPT-5.4 are also available there.
- Data Zone Provisioned Managed (EU): GPT-5.6 Sol, Terra and Luna as well as GPT-5.5 and GPT-5.4 are listed, but not the GPT-6 models, including GPT-6.1 Terra. There is therefore no option for GPT-6 with reserved EU capacity; Global Provisioned Managed lists GPT-6 Astra, Sol and 6.1 Sol in the EU regions but processes worldwide.
- Data Zone Batch (EU): Not listed for GPT-5.5, GPT-5.6 and GPT-6 Astra; the category only extends to GPT-5.4/GPT-5.4-mini and the o-series.
- Global Standard: deployable, but inference may run worldwide. For EU data processing, explicitly choose Data Zone Standard (EU).
- GPT-6 Astra: Data Zone Standard (EU) in ten regions, North Europe in addition to the nine above. Reserved capacity is only available as Global Provisioned Managed (worldwide processing) or as Data Zone Provisioned Managed in the US.
- Quota: Tier 5 and tier 6 subscriptions have standard quota; lower quota tiers must submit a quota request. The classic GPT-5 family is also available via the Azure OpenAI Service in West Europe, Germany West Central and Sweden Central.
Amazon Bedrock
On Bedrock, GPT-5.5, GPT-5.4 and Codex have been available since 1 June 2026, GPT-5.6 since 13 July, GPT-6 Astra since 8 September, GPT-6 Sol and Luna since 22 September, GPT-6.1 Sol since 29 September and GPT-6.1 Terra since 26 October 2026 (model IDs openai.gpt-5.6-sol, openai.gpt-6-astra, openai.gpt-6-sol, openai.gpt-6-luna, openai.gpt-6.1-sol, openai.gpt-6.1-terra etc.). There is no EU-resident option: via the bedrock-mantle endpoint the models run exclusively In-Region in individual US regions (e.g. GPT-6 Sol/Luna, 6.1 Sol and 6.1 Terra only in us-east-1), without cross-region inference. Via bedrock-runtime, GPT-5.6, GPT-6 and GPT-6.1 have a Geo profile (US and partly Canada only) and a Global profile that can be called from EU regions such as eu-central-1 (Frankfurt) but explicitly processes worldwide. In-Region prices include a 10% surcharge; Global prices match OpenAI’s rates. Per the model cards the context window is 1 million tokens (Astra 1,050,000); the 272k limit is only the price threshold. Prompt caching offers a 90% discount on cached inputs (for GPT-6.1 Sol the model card lists no explicit prompt caching, for GPT-6.1 Terra explicit caching via the Responses API). AWS describes “zero-operator access”; traffic flagged by classifiers is stored for up to 30 days. For workloads with an EU residency requirement, Bedrock is therefore ruled out for OpenAI models.
OpenAI direct: retention, ZDR and data residency
The OpenAI platform API stores abuse-monitoring logs for up to 30 days by default. Zero Data Retention (ZDR) is possible after prior approval by OpenAI at organization or project level; store is then forcibly treated as false. ZDR-eligible endpoints include /v1/chat/completions, /v1/responses, /v1/images/*, /v1/embeddings, /v1/audio/*, /v1/realtime and /v1/moderations, but not Assistants, Conversations and Vector Stores. For Europe (EEA and Switzerland), OpenAI offers data residency (+10% for models from 5 March 2026), per OpenAI for GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Sol and Luna with Standard, Flex and Batch processing; the data residency table also lists GPT-6 Astra with EU processing for Responses, Chat Completions and Batch. Fast Mode with EU residency is available for 6.1 Sol, 6.1 Terra and Luna, but not for GPT-6 Sol and Astra; Ultrafast mode for GPT-6.1 Sol supports US and EU data residency, Ultrafast for Astra only US residency and global processing. Since 21 August 2026, regional processing can also be chosen per request (prefix domain with the API key of a project with the “Global” geography); since 29 August 2026, mutual TLS (mTLS) and X.509 workload identity federation are generally available. Since 5 October 2026, eligible organizations can enable the HIPAA compliance support flow in the API settings.
In Microsoft Foundry, a 30-day abuse-monitoring retention applies that cannot be switched off by the customer; modified abuse monitoring or ZDR is possible only after approval by Microsoft (limited access) and usually with an Enterprise Agreement or MCA-E. We clarify these points in projects before go-live and document them in the record of processing activities.
Migration and breaking changes
For gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini, which are shut down in January 2027, OpenAI names GPT-6 Luna as the replacement; fine-tunes of these models follow the base model’s schedule. In practice:
- Reasoning levels: GPT-6 Luna has six levels from
nonetomax. Function calling via Chat Completions is only possible withreasoning_effort: "none"; tool use with reasoning runs through the Responses API. - Fine-tuning: there is no fine-tuning for the GPT-6 models, so existing fine-tunes on GPT-4o or GPT-4.1 cannot be carried over one to one.
- Speech and transcription: text-to-speech (
tts-1,tts-1-hd,gpt-4o-mini-tts) moves togpt-realtime-2.1-mini; transcription togpt-transcribeorgpt-live-transcribe. Output formats and diarization logic may change, see OpenAI Whisper. - Foundry: Microsoft runs its own retirement schedules, which can differ from those of the OpenAI API. Check the dates per platform.
- Monitoring: OpenAI exposes an optional
shutdown_datefield viaGET /v1/models, so shutdowns can be monitored automatically. - From GPT-5.6 Terra to GPT-6.1 Terra: OpenAI recommends comparing both models with the same reasoning effort on representative images and workflows and moving tool calls from Chat Completions to the Responses API. Terra keeps
none; for GPT-6.1 Sol and Astra, replace it withlow.
Model lifecycle
Older models remain available for now and are still EU-suitable in Foundry:
- GPT-5.5 (23 April 2026, codename “Spud”): API from 24 April, $5 / $30 per 1M tokens, 1M context; Thinking and Pro variants. GPT-5.5 Instant (5 May 2026) is available as
chat-latestand in Foundry asgpt-chat-latest; OpenAI cites 52.5% fewer hallucinations than GPT-5.3 Instant on high-stakes prompts. - GPT-5.4 (5 March 2026): native computer control (75% on OSWorld-Verified, above the human baseline of 72.4%), 1,050,000 tokens of context and tool search, which cuts token costs in tool-heavy workflows by around 47%. Variants: GPT-5.4 ($2.50 input), Pro ($30.00), mini and nano (nano is deprecated).
- GPT-5.3 Codex (February 2026): specialized model for agentic coding, also available on Bedrock; deprecated for 1 April 2027.
Status per the OpenAI deprecations page (shutdown = removal from the API):
| Model | Status | Shutdown | Replacement |
|---|---|---|---|
| Assistants API | shut down | 26 Aug 2026 | Responses API + Conversations API |
sora-2, sora-2-pro, Videos API | shut down | 24 Sep 2026 | no successor |
gpt-5.4-cyber | shut down | 1 Oct 2026 | gpt-5.6-cyber |
o1, o1-pro, o3-mini, gpt-4, gpt-4-turbo, gpt-4o-2024-05-13 | shut down | 23 Oct 2026 | GPT-5.6 Sol (o1-pro with reasoning.mode: pro) |
o4-mini, gpt-3.5-turbo | shut down | 23 Oct 2026 | GPT-5.6 Terra |
gpt-4.1-nano | shut down | 23 Oct 2026 | GPT-5.6 Luna |
gpt-image-1 | shut down | 23 Oct 2026 | gpt-image-2.5-sunburst or gpt-image-2.5-flare |
Evals platform (dashboard and API), Agent Builder, reusable prompts (v1/prompts) | deprecated | 30 Nov 2026 | depends on product (Promptfoo, Agents SDK, prompts in application code) |
gpt-image-1-mini, gpt-image-1.5, chatgpt-image-latest | deprecated | 1 Dec 2026 | gpt-image-2.5-sunburst or gpt-image-2.5-flare |
gpt-5, o3 | deprecated | 11 Dec 2026 | GPT-5.6 Sol |
gpt-5-mini | deprecated | 11 Dec 2026 | GPT-5.6 Terra |
gpt-5-nano | deprecated | 11 Dec 2026 | GPT-5.6 Luna |
gpt-5-pro, o3-pro | deprecated | 11 Dec 2026 | GPT-5.6 Sol with reasoning.mode: pro |
tts-1, tts-1-hd, gpt-4o-mini-tts (both snapshots) | deprecated | 6 Jan 2027 | gpt-realtime-2.1-mini |
gpt-4o (2024-08-06, 2024-11-20), gpt-4o-mini, gpt-4.1, gpt-4.1-mini | deprecated | 14 Jan 2027 | GPT-6 Luna |
gpt-realtime, gpt-audio, gpt-4o-audio, gpt-4o-realtime, mini variants, gpt-4o-mini-transcribe-2025-03-20 | deprecated | 20 Jan 2027 | gpt-realtime-2.1, gpt-realtime-2.1-mini, gpt-audio-1.5 |
whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize | deprecated | 26 Feb 2027 | gpt-transcribe, gpt-live-transcribe |
gpt-5.3-codex, gpt-5.1 | deprecated | 1 Apr 2027 | GPT-6 Sol |
gpt-5.4-nano | deprecated | 1 Apr 2027 | GPT-6 Luna |
If you use Whisper via the OpenAI API for meeting minutes, call-center transcripts or diarization, plan the migration, as output formats and diarization logic can change with the model switch.
Integration with CompanyGPT
With CompanyGPT you can use GPT models GDPR-compliantly in your company – without your data being used for training.
Our recommendation
For GDPR-sensitive workloads we recommend the GPT-6 generation via Microsoft Foundry with Data Zone Standard (EU), in clearly separated roles:
- GPT-6 Luna – default for most enterprise workloads, also as the replacement for GPT-4o and GPT-4.1. At $0.10 / $0.50 per 1M tokens with 1.05M context, full tool support and an EU Data Zone in all nine regions, Luna is usually the economically right choice for chat assistants, classification, extraction, summarization, RAG and sub-agents. Limits: lower reasoning depth than Sol and Astra, double input price above 272k tokens and no Data Zone Provisioned Managed option.
- GPT-6.1 Terra – for image and document analysis, computer use and writing. At $1.00 / $5.00 per 1M tokens half the price of GPT-6.1 Sol, with an EU Data Zone in all nine Foundry regions and, per OpenAI, on par with Astra on Agents’ Last Exam. Because little comparison data is available at launch, measure Terra on your own tasks against Luna and 6.1 Sol.
- GPT-6.1 Sol – for complex coding and agentic workflows. Same price as GPT-6 Sol, half the cached-input price and per OpenAI close to Astra. Escalate to it selectively where quality requires it; when switching from GPT-6 Sol, replace
nonewithlow. - GPT-5.6 Sol, Terra and Luna remain a proven fallback, for example when prompts and evals are not yet calibrated for GPT-6 or when Data Zone Provisioned Managed is required. GPT-5.6 Terra has been superseded by GPT-6.1 Terra at half the price since 26 October 2026.
- GPT-6 Astra: evaluate first, do not migrate broadly: the EU Data Zone in Foundry and EU processing in the OpenAI API are available, but without reserved EU capacity, without fast mode under EU residency and without an EU-resident option on Bedrock; add cyber monitoring and five times the price of GPT-6.1 Sol. Compare on price per completed task.
- Existing projects: GPT-5.5, GPT-5.4 and GPT-5.5 Instant (
gpt-chat-latest) continue to run in EU regions; a move to GPT-5.6 or GPT-6 usually pays off in the short term. If you still run ongpt-4o,gpt-4.1orgpt-4o-mini, you must move to GPT-6 Luna (or another model) by 14 January 2027, and if you usegpt-5,o3orgpt-5-mini, by 11 December 2026;o1,o3-miniando4-miniare already shut down. Plan fine-tunes separately. - Amazon Bedrock is of interest for OpenAI models only as a multi-cloud path for US workloads, not where EU residency is required.
We continuously check EU availability against the Microsoft Learn region table and the AWS model cards and advise project by project on model choice, deployment type and data protection.