OpenAI API GPT and Image Generation

The OpenAI API provides programmatic access to OpenAI’s GPT model family, image generation, speech, and embedding models. It is the most widely used LLM API and has become the de facto standard that other providers emulate - chat-completions-style message arrays and function-calling schemas are now conventions the rest of the industry follows too. For enterprise AI projects, the OpenAI API is often the first integration point for proof-of-concept work, though production deployments may migrate to Azure OpenAI or Amazon Bedrock for compliance and enterprise support reasons.

Official documentation: https://developers.openai.com/api/docs (platform.openai.com/docs now redirects here)

Core API Endpoints

Chat Completions (and the newer Responses API) - The primary endpoints for text generation. You send a list of messages (system, user, assistant roles) or a structured input, and receive a model response. This powers conversational AI, content generation, summarization, classification, and virtually every text-based AI task. The Responses API is now the strategic endpoint rather than a parallel option: the Assistants API was shut down on 26 August 2026, with the Responses API and the Conversations API named as its replacements, and on GPT-6 Astra tool calling works only through Responses. New work should start on Responses unless you have a specific reason not to. The same pattern holds on GPT-6 Sol and Luna: built-in tools need Responses, and Chat Completions supports function calling on those two models only with reasoning_effort set to none.

Agents API (public beta) - Released 10 September 2026. It runs agents on a managed Codex harness, with OpenAI handling session orchestration, context compaction and recovery. Durable sessions carry work across turns and stream progress; you can connect your own tools and MCP servers, and run the agent in an OpenAI-hosted sandbox or one from your own infrastructure or a supported provider. It is a beta, so budget for interface changes before you put it on a critical path.

Live API - v1/live/sessions is the endpoint for GPT-Live 1 (see Audio below): full-duplex voice sessions that hand reasoning and tool calls to a backend model or agent.

Assistants and Videos are gone. The Assistants API shut down on 26 August 2026 (replacements: Responses API and Conversations API). The Videos API and the sora-2 / sora-2-pro models and snapshots shut down on 24 September 2026, and OpenAI’s deprecations table names no replacement. Any code path still calling either endpoint is already broken.

Embeddings - Converts text into numerical vectors for semantic search, clustering, and classification. text-embedding-3-small and text-embedding-3-large remain OpenAI’s current embedding models - OpenAI has not shipped a successor since these launched, so this part of the lineup is genuinely stable. Embeddings are the foundation for RAG systems: embed your documents, store the vectors, and retrieve semantically similar content at query time.

Images - DALL-E 2 and DALL-E 3 were retired on 12 May 2026. On 8 September 2026 OpenAI shipped a new image generation stack: two API models, gpt-image-2.5-flare - positioned as the default choice for most applications, with up to 50% lower generation latency than gpt-image-2 at higher quality - and gpt-image-2.5-sunburst, a premium tier for campaign creative and product imagery that trades longer generation times for tighter control across successive edits. Both bill at the same rates as gpt-image-2, work through the Image API and the Responses API image generation tool, and add new xhigh and max quality settings. The consumer counterpart, “ChatGPT Images 2.5”, rolled out the same day to all ChatGPT, ChatGPT Work and Codex tiers on desktop, mobile and web. gpt-image-2 (launched 21 April 2026) remains available and is the named migration target for the older line, but that older line is now on a clock: gpt-image-1 retires 23 October 2026, and gpt-image-1.5 and gpt-image-1-mini on 1 December 2026. Billing for image models is token-based for both the prompt and the generated image, not a flat per-image rate the way DALL-E was - check the pricing page directly for current per-resolution figures.

Audio - The transcription lineup turned over on 28 July 2026. gpt-transcribe is now OpenAI’s recommended general-purpose speech-to-text model at $0.0045 per minute of audio - 25% below the $0.006/min that whisper-1 and gpt-4o-transcribe charge - with a word error rate of 19.27% against 40.37% for whisper-1 on OpenAI’s launch benchmark (Common Voice, 22 languages). It accepts context prompts, keyword hints and multiple language hints, and reports the languages it detected. gpt-live-transcribe is the streaming counterpart for transcription as audio arrives. Do not start new work on the previous generation: on 26 August 2026 OpenAI deprecated whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize together, all shutting down 26 February 2027 with gpt-transcribe/gpt-live-transcribe as the named replacements. The realtime and audio-native models turned over too - gpt-realtime-2.1 (audio $32.00 in / $0.40 cached / $64.00 out; text $4.00 / $0.40 / $24.00 per 1M tokens, with a cheaper gpt-realtime-mini) replaces gpt-realtime and gpt-4o-realtime, and gpt-audio-1.5 replaces both gpt-audio and gpt-4o-audio; all four predecessors shut down 20 January 2027. For conversational voice, GPT-Live 1 (gpt-live-1) became generally available on 10 September 2026 on the new v1/live/sessions endpoint. It is a full-duplex model: it listens and speaks at the same time and delegates reasoning and tool use to a backend, either an OpenAI model through Responses delegation or your own backend through client delegation. Sessions cost $0.05 per minute, billed per second with no rounding up, and backend model and tool usage is billed separately at normal rates. Rate limits are counted in concurrent sessions (25 at Tier 1 up to 500 at Tier 5; the Free tier is not supported), and the model does not support structured outputs. Text-to-speech is the one part of the audio stack that has not moved: gpt-4o-mini-tts is still the current recommendation, described in OpenAI’s own guide as its newest and most reliable TTS model, ahead of the older tts-1/tts-1-hd.

Function Calling

Function calling enables the model to generate structured JSON that maps to functions you define. Instead of parsing free-text responses, you declare functions with typed parameters, and the model returns a function call when appropriate. This is the foundation for tool-using agents: the model decides which tool to use and provides the arguments.

Define functions in the API request, and the model either responds with text or requests a function call. Your application executes the function and returns the result to the model for further processing. This loop continues until the model has enough information to provide a final response.

Function calling is more reliable than prompt-based JSON extraction because the model is specifically trained to generate valid function call structures. It supports parallel function calls (requesting multiple tools simultaneously) and strict mode (guaranteeing JSON schema conformance). Since 3 September 2026 the platform also supports async tool calling and mid-turn steering over WebSockets, plus adjusting reasoning effort mid-conversation - though the Azure-hosted models do not support the last two.

Structured Outputs

The response_format parameter constrains model output to valid JSON matching a provided schema. This eliminates the parsing failures that plague free-text extraction. For enterprise applications that feed model outputs into downstream systems, structured outputs are essential for reliability.

Current Model Lineup

Here is the lineup as of 3 October 2026. The short version: GPT-6 has four shipping models, Astra (3 September), Sol and Luna (22 September) and GPT-6.1 Sol (29 September), and there is no GPT-6 Terra. The GPT-5.6 Sol/Terra/Luna family remains available alongside it with no retirement date. Context for how we got here: the GPT-5 family superseded GPT-4o and GPT-4 Turbo and folded the earlier “o-series” reasoning models (o1, o3) into unified models that reason natively rather than requiring a separate model choice. Point releases 5.1, 5.2 (including 5.2-codex and 5.2-chat), 5.3 (chat/codex) and 5.4 (mini/nano/pro) followed in between, and the 5.1 variants retired on 23 July 2026 and gpt-5.2-chat-latest on 10 August 2026, all before GPT-5.5 in April 2026 and the named GPT-5.6 tiers in July.

GPT-6 Astra - launched 3 September 2026 and the most capable model in the API, under model ID gpt-6-astra. It carries a 1,050,000-token context window (922,000 max input, 128,000 max output), a 30 April 2026 knowledge cutoff, and text-plus-image input with text-only output - no audio or video. Reasoning effort runs low / medium / high / xhigh / max; there is no none level, and custom temperature, top_p and logprobs are not accepted. It is served through the Responses API and Chat Completions, with tool calling available only on Responses, and supports web search, file search, code interpreter and computer use. Multiple outlets reported it as the first OpenAI model classified at the “Critical” cybersecurity threshold under OpenAI’s own Preparedness Framework; OpenAI’s launch post is not machine-retrievable, so treat that classification and the widely-repeated account of it chaining two unknown vulnerabilities into a working exploit as secondary-sourced rather than confirmed from OpenAI’s own words. See this wiki’s full writeup: Astra becomes the first OpenAI model to cross the “Critical” cyber threshold .

The rollout was staged and publicly rocky - Sam Altman apologised for the “messy rollout”. Day one, 3 September, was limited to organisations already in OpenAI’s Daybreak security program; the API opened around 5 September; ChatGPT Plus was reached around 7 September; and GA on Amazon Bedrock landed 8 September 2026. Correcting a common misreading of the access story: as of 9 September 2026 the base model is not gated. OpenAI’s model page for gpt-6-astra lists no eligibility or trusted-access requirement and the rate-limits guide sets no tier minimum for it, so there is no request step to budget for on the direct API. What is gated is the elevated-cybersecurity configuration, behind the Daybreak program’s two access tiers - Daybreak Blue for general security work and Daybreak Red for pen-testing and exploit validation, both launched 7 August 2026 - which means published Astra cyber results describe Daybreak-configured access rather than default production behaviour. Two caveats worth checking against your own account rather than assuming: trade coverage from 6-7 September described per-organisation entitlement as still uneven with some organisations waitlisted, and Altman said on 5 September that OpenAI was “working towards getting Astra in everyone’s hands”, so no source confirms the API rollout is finished. On Azure the gate is different again - quota, not eligibility (see Enterprise Considerations).

Astra also introduces in-flight safety stops. OpenAI’s name for the mechanism is misalignment monitoring: in supported Responses API requests it checks agent work asynchronously and can raise a safety alert or stop a conversation for review part-way through a trajectory. These are not timeouts and not rate limits, and your retry logic should not treat them as either - the same request will be stopped again. Build for a terminal outcome rather than a transient one.

ChatGPT access to Astra is per-surface, not simply per-plan, and it is not the default anywhere. Per secondary reporting (OpenAI’s help center is not machine-retrievable, so this is not primary-verified): Pro, Business and Enterprise see it in the regular chat picker labelled “GPT-6 Pro” - Enterprise off by default until an admin enables it, Business limited on Standard and full on Premium - while Plus gets it only inside ChatGPT Work and Codex rather than the chat picker, and Free and Go are excluded. At the time of that reporting GPT-5.6 Sol was the ChatGPT default across every tier. That picture predates the 22 September launch of GPT-6 Sol and Luna, and this page has not verified the ChatGPT defaults since, so check them before relying on them. The Plus position was consistent across sources through 7 September but nothing dated 8-9 September confirms it either way, so verify before relying on it.

GPT-6 Sol and GPT-6 Luna - released 22 September 2026 under model IDs gpt-6-sol and gpt-6-luna. Both are reasoning models with text-plus-image input and text output, served through the Responses API, Chat Completions and Batch. Both have the same 1,050,000-token context window as Astra (922,000 max input, 128,000 max output). Knowledge cutoffs are 20 April 2026 for Sol and 18 May 2026 for Luna. Unlike Astra, both accept every reasoning-effort level: none, low, medium (the default), high, xhigh and max. Within GPT-6, Astra is the flagship, Sol is the mid tier and Luna the efficient tier. OpenAI describes Sol as built for complex coding and agentic workflows, and Luna as its “most efficient model for focused, high-volume tasks”. In the Responses API both support the full hosted tool set (web search, file search, image generation, code interpreter, hosted shell, apply_patch, skills, computer use, MCP and tool search). EU data residency is available only with Standard processing.

GPT-6.1 Sol - released 29 September 2026 at DevDay under model ID gpt-6-1-sol. It is a point upgrade to GPT-6 Sol rather than a new tier, and OpenAI positions it as approaching GPT-6 Astra on complex professional work at roughly one fifth of Astra’s token cost. Pricing is unchanged from GPT-6 Sol at $2/$10 per MTok, but cached input halves to $0.10 (OpenAI describes it as 95% below the standard input rate), which is the change that actually moves the bill on prompt-heavy workloads. OpenAI’s reported figures: matches GPT-6 Astra on DeepSWE v1.1 coding, approaches Astra’s state of the art on the GDP.pdf professional-document benchmark, scores 4.8 points above GPT-6 Sol on AutomationBench and 7 points above it on OSWorld 2.0 computer use, more than doubles GPT-6 Sol on Terminal-Bench Science, and cuts the factuality error rate from 11.4% to 7.7% at low reasoning effort. OpenAI notes its evaluations may differ from production ChatGPT output and that the tested scenarios do not represent typical usage. Even at 7.7%, roughly one claim in thirteen is wrong at that setting, so it is not a substitute for verification. At release it was in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, and in the API; it was not yet in standard Chat. OpenAI has not published its context window or rate limits. See GPT-6.1 Sol and DevDay 2026 .

DevDay 2026, the same day, also shipped GPT-6 Astra Ultrafast (up to 8x faster token generation in Codex at around 300 tokens per second, up to 6x in the API), a Decisions API that constrains Luna to a defined set of user-specified questions with text and image input, an Agents API with computer use, managed OpenAI agents on Amazon Bedrock through an AWS partnership, Codex Cloud and Codex Security Cloud, dots (always-on agents for Pro, Business and eligible Enterprise and Education plans), a Pro 500 tier, and a Private Inference confidential-computing preview for autumn 2026. The Bedrock item is the one that changes architecture decisions: OpenAI agents running natively in AWS removes the egress and data-residency objections that previously pushed AWS-based teams towards Azure OpenAI or Bedrock-native models.

The pricing is the headline. GPT-6 Sol lists at $2 / $10 per 1M tokens, the same input price as GPT-5.6 Terra ($2 / $12) with cheaper output. GPT-6 Luna lists at $0.10 / $0.50, half of GPT-5.6 Luna’s $0.20 / $1.20. So on list price the newer generation is the cheaper one at each tier (see Pricing). Launch coverage: GPT-6 Sol and Luna . Run your own evals before you switch: OpenAI has not retired or repriced the 5.6 tiers, and GPT-6 Sol at medium effort may spend a different number of reasoning tokens per task than 5.6 Sol, so the per-token saving may not match your per-task saving. Tier 1 rate limits for gpt-6-sol start at 500 RPM / 500,000 TPM and rise to 15,000 RPM / 40,000,000 TPM at Tier 5.

GPT-5.6 - Sol / Terra / Luna - previously the self-serve, generally-available flagship family and still fully supported. Terra has no GPT-6 counterpart, so it remains the mid-tier option. The family was previewed in late June 2026 and GA 9 July 2026 (GA on Amazon Bedrock 13 July 2026). Point releases gave way to three named capability tiers, all sharing a 1,050,000-token context window, 128K max output, and full reasoning-mode support:

• Sol - flagship tier for the hardest reasoning, coding, and agentic work; knowledge cutoff 16 February 2026. The gpt-5.6 alias routes here.
• Terra - balanced cost and capability; the reasonable default for most production traffic, and OpenAI’s named migration target for gpt-3.5-turbo, gpt-3.5-turbo-instruct and o4-mini.
• Luna - cost-optimized, for high-volume/lower-complexity tasks: classification, simple extraction, routine summarization. Named replacement for gpt-4.1-nano, and the model behind the ChatGPT Free tier.

Prices fell twice over the summer: Luna and Terra were cut on 30 July 2026 (reported at roughly 80% and 20% respectively, secondary sources only), and Sol on 21 August 2026 by 20% on input and 33% on output. Sol’s cut comes with a string attached - see the Pricing section. Release detail: OpenAI ships GPT-5.5, then the GPT-5.6 Sol/Terra/Luna family .

GPT-5.5 - released 23 April 2026 (API 24 April) with a 1M-token context window and an agentic focus; GPT-5.6 Sol effectively supersedes it as the flagship choice, but it is absent from OpenAI’s deprecations table entirely and has no announced retirement date, so it remains a reasonable option if you’re already tuned to it and is still listed on Azure Foundry. GPT-5.5 Pro adds extended-compute reasoning at a premium price. No GPT-5.6 tier has a published retirement date either.

Legacy and retiring models - the last time this page was accurate, GPT-4o and gpt-3.5-turbo were the current models. Both are now superseded by the GPT-5.6 family, and OpenAI has since published dates for most of the rest of the back catalogue. GPT-4o was withdrawn from ChatGPT on 13 February 2026, with Business, Enterprise and Edu retaining it inside Custom GPTs until 3 April 2026, and chatgpt-4o-latest ended in the API in February 2026; the base gpt-4o snapshots (gpt-4o-2024-11-20, -08-06) are still listed with no announced retirement, but gpt-4o-2024-05-13 goes on 23 October 2026. That same date, 23 October 2026, is a single large sweep: gpt-3.5-turbo and gpt-3.5-turbo-0125, gpt-4 and gpt-4-0613, gpt-4-1106-preview, gpt-4-turbo, gpt-4o-2024-05-13, o1, o1-pro, o3-mini, o4-mini, gpt-4.1-nano and gpt-image-1 all shut down together (plus the fine-tuned variants of several of them). Around it: the Videos API with sora-2 and sora-2-pro shut down on 24 September 2026 with no replacement. On 28 September 2026 gpt-3.5-turbo-instruct, gpt-3.5-turbo-1106, babbage-002 and davinci-002 go, all pointing at gpt-5.6-terra. gpt-5.4-cyber follows on 1 October 2026, replaced by gpt-5.6-cyber (deprecation announced 11 September). On the platform side, the Evals platform turns read-only on 31 October 2026. The v1/prompts reusable-prompt objects, the Evals dashboard and API, and Agent Builder all shut down on 30 November 2026. Next, gpt-image-1.5, gpt-image-1-mini and chatgpt-image-latest on 1 December 2026, then a second sweep on 11 December 2026 taking o3 and o3-pro along with the entire GPT-5 base family - gpt-5-2025-08-07, gpt-5-mini, gpt-5-nano and gpt-5-pro - which is the one most likely to catch a codebase that migrated off GPT-4 only a year ago. Then the audio and realtime line (gpt-realtime, gpt-4o-realtime, gpt-audio, gpt-4o-audio) on 20 January 2027, and the transcription line (whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize) on 26 February 2027. o1-mini was already retired on 27 October 2025 (replacement o4-mini) and o1-preview on 28 July 2025 (replacement o3), so any code path still naming those two is already broken. o1 itself is not retired yet - it and o1-pro are both in the 23 October 2026 sweep above, each pointing at gpt-5.6-sol, o1-pro with reasoning.mode: pro. The Assistants API shut down on 26 August 2026 and is no longer available. Do not start new work against anything in this paragraph. Migrate to the named replacement above, or to whichever GPT-6 or GPT-5.6 tier matches your cost/quality target. OpenAI’s deprecations table still names GPT-5.6 tiers as the formal replacements.

Pricing

OpenAI bills per token, input and output separately, with a discount for cached input (repeated prompt prefixes) and a “long-context” surcharge once a request’s input exceeds 272K tokens. Verified against OpenAI’s own pricing and model documentation on 25 September 2026, with the GPT-6.1 Sol row added from its 29 September release post on 3 October 2026:

All figures are $ per 1M tokens.

* GPT-5.6 Sol’s rate is promotional, not standing. OpenAI’s pricing page and the gpt-5.6-sol model page both carry a footnote that the promotional pricing is “available at least through November 21, 2026”. The $4.00/$20.00 figures are the result of the 21 August 2026 cut from $5.00/$30.00. If you are building a cost model that runs past November 2026, budget for the possibility of reverting to the pre-promotion rate - a 25% increase on input and 50% on output - rather than treating today’s numbers as permanent.

Astra’s long-context threshold is documented: prompts above 272,000 input tokens are billed at 2x input and cache and 1.5x output, for the whole request rather than only the tokens past the threshold, which is exactly the $20 / $2 / $75 row. Fast mode is 2x standard across the board, and Batch and Flex are each 50% of standard. † The two multipliers are documented separately and OpenAI does not publish a combined Fast-plus-long-context row, so the $40.00 / $4.00 / $150.00 figures above (and cache writes at $25.00 short-context, $50.00 long) are the arithmetic those multipliers imply, not a quoted price - treat them as an estimate and confirm against an actual invoice before you budget on them.

GPT-6 Sol and Luna use the same rules. Above 272K input tokens the whole request bills at 2x input and cache and 1.5x output, which gives the long-context column above. Batch and Flex are 50% of Standard, Fast mode is 2x the applicable rate, and regional processing adds 10% where available.

Three billed items the table does not show. Cache writes are charged separately from cache reads: 1.25x the uncached input rate, which is $12.50 per 1M tokens for Astra, $2.50 for GPT-6 Sol and $0.125 for GPT-6 Luna at short context. So a workload that churns its prompt prefix pays for the privilege rather than saving. Flex processing sits alongside Batch at 50% of standard. And per-request regional processing selection, added 21 August 2026, carries a 10% uplift for data-residency-eligible models released on or after 5 March 2026. Image and audio models (the gpt-image line, gpt-transcribe, TTS, Realtime) are priced separately in different units and aren’t shown above.

The Batch API remains a flat 50% discount off standard synchronous token rates for eligible models, with results delivered within 24 hours - still the right default for document processing, classification, and extraction workloads that don’t need a live response. For cost optimization generally: GPT-6 Sol, the mid tier, lists at GPT-5.6 Terra’s input price with cheaper output, and GPT-6 Luna comes in below GPT-5.6 Luna. The natural starting point for new work is therefore GPT-6 Sol, with GPT-6 Luna for high-volume, low-complexity work and Astra for tasks where the extra reasoning changes the outcome. Validate that choice against your own evals before migrating tuned 5.6 workloads. Keep repeated prompt content (system prompts, few-shot examples, long shared context) stable across calls so it qualifies for cached-input pricing.

One specialist model to know about: GPT-Rosalind (gpt-rosalind-research) became generally available on 8 September 2026 through a trusted-access program for approved life-sciences research. It costs $5 / $0.50 cached / $25 per 1M tokens, and billing begins 5 October 2026.

Prompt Caching

Caching changed materially for GPT-5.6 and later models, including all three GPT-6 tiers:

• Cache writes are billed. A write costs 1.25x the uncached input rate and each read 0.1x. OpenAI’s own arithmetic: one write plus one full reuse costs 1.35x the ordinary input cost, against 2x for sending the prefix twice uncached. One write plus nine reads costs 2.15x against 10x.
• Explicit breakpoints. Set prompt_cache_options.mode to explicit and mark boundaries with prompt_cache_breakpoint: { "mode": "explicit" } on a content block inside an input message. Content after the last breakpoint is billed at the uncached rate with no write charge, so you avoid paying to cache text that changes on every call. Each request can create up to four cache writes. In implicit mode OpenAI places a breakpoint at the end of the latest eligible message, and you can add explicit breakpoints on top of it. Top-level instructions cannot hold a breakpoint, so put reusable developer instructions in an input_text block inside a developer message.
• Change effort without breaking the cache. On GPT-6 models, append a configuration_update input item to change reasoning effort between responses and leave the request-level reasoning.effort unchanged. That preserves the cached prefix, whereas editing the top-level setting can rewrite the hidden system instructions.
• Tooling. A Prompt Caching dashboard (20 August 2026) shows hit rate, reads per write and the split between cache-read, cache-write and uncached tokens. Prompt Cache Diagnostics became GA in the Responses API on 8 September 2026. It compares cache reuse against a previous response and reports why a request missed.

On models before GPT-5.6, only implicit caching applies and there is no write charge. A stable prompt_cache_key remains the main lever there.

Rate Limits and Scaling

OpenAI enforces rate limits per organization/project across several metrics - RPM (requests/minute), RPD (requests/day), TPM (tokens/minute), TPD (tokens/day), and IPM (images/minute) for image models - whichever limit a request hits first applies. Accounts move up automatically through a ladder of usage tiers (a geography-limited Free tier, then Tier 1 through Tier 5) as cumulative spend increases, with each tier raising the ceiling; there is no separate manual increase request outside of that spend-based graduation. For production workloads, implement retry logic with exponential backoff and plan capacity against your current tier’s ceiling rather than assuming it scales instantly with demand.

Two additions to that ladder matter for anyone sizing production capacity. Long-context models carry a separate long-context rate limit on top of the standard ladder, so a workload that routinely pushes past the 272K threshold can be throttled independently of its ordinary TPM headroom - GPT-5.5 was the first model to make this visible and it now applies across the long-context lineup. And OpenAI offers a Reserved Tier capacity option for GPT-5.6-and-later models, which is the right answer when you need predictable throughput rather than best-effort shared capacity. For high-throughput batch processing, the Batch API (see Pricing above) uses a separate, larger rate-limit pool that doesn’t compete with your synchronous traffic.

Error semantics changed on 2 September 2026: a 429 slow_down (back off and retry) is now distinct from a 503 server_is_overloaded, so retry logic should branch on the two rather than treating every failure the same way. Astra’s in-flight safety stops are a third category again and should not be retried at all.

Enterprise Considerations

By default, API inputs and outputs are retained for up to 30 days for abuse monitoring and then deleted, and API data is not used to train OpenAI’s models unless you explicitly opt in. Enterprise and Business customers can request Zero Data Retention (ZDR) for eligible endpoints, which removes even that 30-day window. SOC 2 Type II compliance remains available. Always confirm current retention and training defaults against OpenAI’s own enterprise-privacy documentation before committing contractually, since these terms move.

Credential controls tightened over August and September 2026. Mutual TLS and X.509 workload identity federation went GA on 29 August. Since 10 September you can set expiration dates on project API keys, and administrators can enforce a maximum key lifetime at organization or project level. On 15 September OpenAI added key-creation governance: administrators can allow only service-account keys, allow only user-owned project keys, or disable new key creation entirely. Organization restrictions override project settings, and existing keys are not affected. If your security baseline requires short-lived credentials, these controls mean you can now enforce that inside the platform rather than only in your own tooling.

For organizations that need Azure- or AWS-native infrastructure, data-residency guarantees beyond OpenAI’s own locations, or procurement through an existing cloud agreement, Azure OpenAI and Amazon Bedrock offer the same underlying models through each cloud’s enterprise compliance and billing framework. Astra reached the hyperscalers unusually fast - GA on Amazon Bedrock on 8 September 2026, five days after launch, and generally available in Microsoft Foundry as model version 2026-09-03. (The exact date Foundry flipped from limited access to GA is disputed: Microsoft’s own blog announcing general availability carries a 3 September byline, but contemporaneous 3-4 September analyses still described Foundry access as a limited-access program requiring approval, and one account dates the catalog flip to 7 September. What is verifiable today is that the Foundry catalog entry reads lifecycle status “Generally available (GA)”.)

The material constraint on Azure was geography and quota. Geography has since moved. At launch Astra was offered in Global and US Data Zone deployments only, with no EU Data Zone. On 22 September 2026 Microsoft brought GPT-6 Sol and Luna to general availability in Foundry. It now lists Standard deployment for Astra, Sol and Luna across all 28 Global regions and in both the US and EU Data Zones, Provisioned Throughput for Astra and Sol, and Priority Processing for Sol (Global and US Data Zone). Quota for Astra was tier-gated at launch: Tier 5 and Tier 6 subscriptions had quota by default and everyone else had to file a request. Check whether that still applies to your subscription. Azure prices Astra Global Standard at $10 / $1 / $50 per 1M tokens short-context and $20 / $2 / $75 long. US Data Zone costs 10% more ($11 / $1.10 / $55) and EU Data Zone 20% more ($12 / $1.20 / $60 short-context, $24 / $2.40 / $90 long). GPT-6 Sol and Luna on Azure Global Standard match OpenAI’s direct list prices. See Azure OpenAI for the full table. Two capability gaps relative to the direct API: Azure does not support changing reasoning effort mid-conversation or mid-turn steering. Microsoft’s documentation also notes that Astra may apply enhanced safety controls at inference time, including classifier threshold changes and system-injected safety instructions - worth knowing if you are benchmarking Azure output against the direct API and finding it differs.

OpenAI’s commercial center of gravity has shifted alongside the model lineup: CFO Sarah Friar told investors in August 2026 that enterprise revenue (~$40B annualized run rate) now exceeds ChatGPT consumer revenue for the first time, about six months ahead of OpenAI’s own prior guidance - see OpenAI’s enterprise revenue overtakes consumer . On the consumer side, free and low-cost ChatGPT tiers picked up unpersonalized advertising in the US and, from 24 August 2026, across 31 European markets; paid tiers (Plus, Pro, Business, Enterprise, Education) remain ad-free - see ChatGPT ads reach 31 European markets . Neither change affects API terms directly, but as enterprise becomes OpenAI’s primary business, it’s a reasonable moment to push for stronger SLAs, longer deprecation notice periods, and version-pinning guarantees than you might previously have been offered.

Sources

1. OpenAI, API Pricing: https://developers.openai.com/api/docs/pricing
2. OpenAI, API Models reference: https://developers.openai.com/api/docs/models
3. OpenAI, gpt-6-astra model page (context window, cutoff, pricing, 272K long-context threshold, Fast/Batch/Flex multipliers, no gating note): https://developers.openai.com/api/docs/models/gpt-6-astra
4. OpenAI, gpt-5.6-sol model page (16 February 2026 cutoff, gpt-5.6 alias, promotional pricing through 21 November 2026): https://developers.openai.com/api/docs/models/gpt-5.6-sol
5. OpenAI, Deprecations (every retirement date cited above, with named replacements, including Videos API/Sora 2 on 24 September 2026, the 28 September 2026 GPT-3.5/babbage/davinci sweep and gpt-5.4-cyber on 1 October 2026), checked 25 September 2026: https://developers.openai.com/api/docs/deprecations
6. OpenAI, API changelog (22 Sep GPT-6 Sol/Luna; 15 Sep key-creation governance; 10 Sep Agents API beta, GPT-Live 1 GA, key expiration; 8 Sep GPT Image 2.5 Flare/Sunburst, Prompt Cache Diagnostics, GPT-Rosalind; 3 Sep GPT-6 Astra, async tool calling, mid-turn steering; 2 Sep error codes; 29 Aug mTLS; 26 Aug Assistants API shutdown and transcription deprecations; 21 Aug Sol price cut and regional processing; 20 Aug Prompt Caching dashboard; 7 Aug Daybreak Blue and Daybreak Red), checked 25 September 2026: https://developers.openai.com/api/docs/changelog
7. OpenAI, Rate limits guide (long-context rate limit; Reserved Tier for GPT-5.6 and later; no tier minimum for Astra): https://developers.openai.com/api/docs/guides/rate-limits
8. OpenAI, gpt-transcribe model page ($0.0045/min): https://developers.openai.com/api/docs/models/gpt-transcribe
9. OpenAI, Speech-to-text guide: https://developers.openai.com/api/docs/guides/speech-to-text
10. OpenAI, Text-to-speech guide (gpt-4o-mini-tts still current): https://developers.openai.com/api/docs/guides/text-to-speech
11. OpenAI Developer Community, “Introducing GPT-5.6 series: Sol, Terra and Luna - coming July 9, 10am PT”: https://community.openai.com/t/introducing-gpt-5-6-series-sol-terra-and-luna-coming-july-9-10am-pt/1384931
12. OpenAI Developer Community, “Introducing gpt-image-2 - available today in the API and Codex”: https://community.openai.com/t/introducing-gpt-image-2-available-today-in-the-api-and-codex/1379479
13. AWS, What’s New, 8 September 2026, “OpenAI GPT-6 Astra on Amazon Bedrock”: https://aws.amazon.com/about-aws/whats-new/2026/09/openai-gpt-6-astra-on-amazon-bedrock/
14. Microsoft Learn, “Foundry Models sold directly by Azure” (Astra version 2026-09-03, context sizes, Tier 5/6 quota, enhanced safety controls, unsupported mid-conversation features): https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
15. Microsoft Foundry model catalog, gpt-6-astra (lifecycle status “Generally available (GA)”): https://ai.azure.com/catalog/models/gpt-6-astra
16. Microsoft Azure blog, “GPT-6 Astra: frontier intelligence for work, now generally available in Microsoft Foundry” (Global and US Data Zone pricing; Standard and Provisioned Throughput): https://azure.microsoft.com/en-us/blog/gpt-6-astra-frontier-intelligence-for-work-now-generally-available-in-microsoft-foundry/
17. Computerworld, 7 September 2026, “Sam Altman calls GPT-6 Astra rollout ‘messy’ as enterprise users wait for access” (rollout timeline; uneven entitlement): https://www.computerworld.com/article/4219264/sam-altman-calls-gpt-6-astra-rollout-messy-as-enterprise-users-wait-for-access-2.html
18. Requesty, 6 September 2026, “GPT-6 Astra: production readiness, gated access, the EU gap and safety stops”: https://www.requesty.ai/blog/gpt-6-astra-production-readiness-gated-access-eu-gap-safety-stops
19. Winbuzzer, 4 September 2026, “GPT-6 Astra arrives with major gains, staged access and new questions about its benchmarks” (secondary source for the Preparedness Framework “Critical” classification): https://winbuzzer.com/2026/09/04/gpt-6-astra-arrives-with-major-gains-staged-access-and-new-questions-about-its-benchmarks-xcxwbn/
20. Christopher Alarcon, 5 September 2026, “GPT-6 Astra across ChatGPT plans” (secondary source for the per-surface ChatGPT access matrix): https://christopheralarcon.com/blog/gpt-6-astra-chatgpt-plans
21. Unite.AI, 8 September 2026, “OpenAI releases ChatGPT Images 2.5 with Sketch and two new API models”: https://www.unite.ai/openai-releases-chatgpt-images-2-5-with-sketch-and-two-new-api-models/
22. Artificial Analysis, GPT-6 Astra pricing/benchmarks: https://artificialanalysis.ai/models/gpt-6-astra
23. Artificial Analysis, GPT-5.6 Sol pricing/benchmarks: https://artificialanalysis.ai/models/gpt-5-6-sol
24. CloudZero, “OpenAI API pricing in 2026: every model after the July price cuts”: https://www.cloudzero.com/blog/openai-pricing/
25. VentureBeat, “OpenAI is ending API access to fan-favorite GPT-4o model in February 2026”: https://venturebeat.com/ai/openai-is-ending-api-access-to-fan-favorite-gpt-4o-model-in-february-2026
26. Codersera, “GPT-3.5 Turbo Shutdown (Oct 2026): Dates & What to Use Instead”: https://codersera.com/blog/gpt-3-5-turbo-shutdown-2026-what-to-use-instead/
27. Captain Compliance, “OpenAI Pledges No User Data Retention - What It Means for Enterprise Privacy Compliance”: https://captaincompliance.com/news/openai-pledges-no-user-data-retention-what-it-means-for-enterprise-privacy-compliance/
28. OpenAI Cookbook: https://cookbook.openai.com/
29. This wiki, “Astra becomes the first OpenAI model to cross the ‘Critical’ cyber threshold”: /news/openai-astra-critical-cyber-threshold/
30. This wiki, “OpenAI ships GPT-5.5, then the GPT-5.6 Sol/Terra/Luna family”: /news/openai-gpt-5-5-and-5-6/
31. This wiki, “OpenAI’s enterprise revenue overtakes consumer, six months early”: /news/openai-enterprise-revenue-overtakes-consumer/
32. This wiki, “ChatGPT ads reach 31 European markets”: /news/chatgpt-ads-europe/
33. OpenAI, gpt-6-sol model page (cutoff 20 April 2026, effort levels, endpoints, tools, pricing, rate limits): https://developers.openai.com/api/docs/models/gpt-6-sol
34. OpenAI, gpt-6-luna model page (cutoff 18 May 2026, pricing): https://developers.openai.com/api/docs/models/gpt-6-luna
35. OpenAI, gpt-live-1 model page ($0.05/min billed per second, v1/live/sessions, concurrent-session limits): https://developers.openai.com/api/docs/models/gpt-live-1
36. OpenAI, Prompt caching guide (cache-write charge, explicit and implicit breakpoints, configuration_update): https://developers.openai.com/api/docs/guides/prompt-caching
37. OpenAI, Agents API overview: https://developers.openai.com/api/docs/guides/agents-api/overview
38. OpenAI, gpt-image-2.5-flare model page (quality settings, snapshot gpt-image-2.5-flare-2026-09-08): https://developers.openai.com/api/docs/models/gpt-image-2.5-flare
39. Microsoft Azure blog, 22 September 2026, “GPT-6 Astra, Sol, and Luna: For production agents in Microsoft Foundry” (regions, Data Zones, PTU, Priority Processing, pricing): https://azure.microsoft.com/en-us/blog/gpt-6-astra-sol-and-luna-for-production-agents-in-microsoft-foundry/
40. This wiki, “GPT-6 Sol and Luna”: /news/gpt-6-sol-and-luna/
41. OpenAI, “Introducing GPT-6.1 Sol” (29 September 2026, fetched 3 October 2026 - $2/$0.10/$10 pricing, the one-fifth-of-Astra cost claim, the benchmark figures, the factuality error rates and the ChatGPT and API availability): https://openai.com/index/introducing-gpt-6-1-sol/
42. OpenAI, “DevDay 2026 Recap” (29 September 2026, fetched 3 October 2026 - Astra Ultrafast, Decisions API, Agents API with computer use, managed agents on Amazon Bedrock, Codex Cloud and Security Cloud, dots, Pro 500, Private Inference): https://openai.com/index/devday-2026-recap/
43. OpenAI, “Addendum: GPT-6.1 Sol” deployment safety (29 September 2026): https://deploymentsafety.openai.com/gpt-6-1-sol