OpenAI Models Release Notes (2026, Living Document)

A timeline of OpenAI's 2026 model and API releases. Every entry carries its source and check date: regular entries quote the changelog verbatim (the changelog's .md variant was last fetched successfully on 2026-11-12; the deprecations page returned 200 the same day -- it became reachable again on 2026-10-17, letting this site back-fill the 2026-09-11 to 10-14 retirement announcements); when the docs domain is unreachable, entries cite official SDK release notes and are labelled as such.

Operating model: we re-fetch developers.openai.com/api/docs/changelog weekly, append new entries, and update dataSourcedAt. When this page disagrees with the official changelog, the official one wins.

2026-11 (latest)

Source note: the official changelog channel (developers.openai.com, including the pages' .md variants) became reachable to this site again on 2026-11-01 (after months of 403 following 09-01; only the deprecations page recovered on 10-17), so backfill entries dated 2026-11-01 onward draw on same-day fetches of changelog.md / pricing.md / the official model cards; earlier SDK-side entries draw on the official GitHub release notes and tag sources, and retirement entries follow the official deprecations page text (see the dedicated October deprecation-wave piece).

• 2026-11-10: GPT Image 2 deprecation notice (the deprecations page entry is dated 2026-11-10; the changelog carried no matching entry that day, and this site fetched and checked the deprecations page on 2026-11-12) — gpt-image-2 and its snapshot gpt-image-2-2026-04-21 are deprecated with removal from the API on 2027-02-10; the official replacement row lists gpt-image-2.6-sunburst or gpt-image-2.6-flare. The 92-day notice fits the page's own policy, which sets the minimum for specialized variants of GA models and image generation models at 3 months. The model card already shows Status: Deprecated, and the pricing page still lists gpt-image-2 at standard rates identical to 2.6 / 2.5 ($8 image input / $30 output, $5 text input per million tokens); the 2026-06-02 row (the 2026-12-01 shutdown batch) has had its recommended replacement raised from gpt-image-2.5 to 2.6. See GPT Image 2 deprecated: API removal on February 10, 2027, official replacement is GPT Image 2.6.
• 2026-11-10 to 11-11: Codex 0.167.1 stable patch and preview-line progress (GitHub releases + npm + PyPI, checked by this site 2026-11-12) — rust-v0.167.1 shipped on 11-10T21:32Z with npm latest and PyPI latest both at 0.167.1 (the release notes carry no changelog entries, only a Full Changelog link; the bundled models.json is byte-identical to 0.167.0 -- gpt-6-cyber still has no standalone Codex-side entry, and the two daybreak aliases remain hidden); the 0.169 line advanced to alpha.9 (11-10T22:09Z), alpha.10 (11-11T03:33Z), and alpha.11 (11-11T10:11Z, the npm alpha dist-tag follows), plus 0.167.0-alpha.4.2 (11-11T01:29Z); the 0.170 preview line opens: rust-v0.170.0-alpha.1 (11-11T20:06Z, one-line release notes).
• 2026-11-10: Businesses in ChatGPT (beta) announced (official news RSS entry at 10:00 GMT, Product category; full announcement text read via a rendered-browser channel on 2026-11-10 UTC, curl gets 403; the Business Agents policy page was readable the same day, marked Updated: November 10, 2026) — businesses claim their organization through Business Hub (chatgpt.com/business-hub, behind login) and, once verified, update key details and add self-service product feeds with pricing and availability; insights are aggregated (mention counts, website clicks, discovery topics), OpenAI states it does not share user conversations or Business Agent chats, and insights appear only above thresholds; Business Agents launch for eligible US ecommerce and online service businesses and can be sponsored as a ChatGPT Ads destination (sponsored placements clearly labeled, separate from responses); 1.2 billion weekly users cited. Shopify merchants connect from Business Hub or the Shopify admin, and their eligible products are already integrated through Shopify Catalog. See Businesses in ChatGPT arrives in beta: Business Hub claiming and Business Agents.
• 2026-11-10: The official SDK pair (python v3.49.0 GitHub 06:47:04Z / already PyPI latest; node v7.55.0 GitHub 06:47:51Z / on npm as latest at 06:58:16Z; checked 11-10 UTC via the GitHub API, PR diffs, and tag sources) — #4264/#3140 add typed support for the existing async setting on beta Agents function-tool shapes: python response models gain async_ (aliased to async) and request TypedDicts take an optional async key through a reserved-keyword mixin, node gains async?: boolean on four interfaces (request-side docstrings say Defaults to false); the Responses param type already carried the same async key at the v3.48.0 tag and is untouched; python also fixes Microphone.record() PCM WAV encoding (#3900, float32 converted to 16-bit, int8 offset to unsigned). See openai-python 3.49.0 and openai-node 7.55.0: typed async function tools for the Agents API.
• 2026-11-09: Control Room announced (official news RSS entry at 13:00 GMT, Safety category; the live page is openai.com/index/control-room/; full text read via a rendered-browser channel on 2026-11-10, curl still 403) — a central system tracking every model workload across the research infrastructure and stopping it fast on safety concerns: a 35-second ceiling from stop instruction to output-impossible (an inference-authorization caching lag OpenAI aims to drive to zero), termination implemented in the inference layer, fail-closed (service down — including tampering — means every request denied; already live on all frontier-model research workloads), and an approved safety case required for every agentic-inference workload; the DNS misalignment incident report it cites (on the alignment site, updated 09-25) was readable via curl the same day. The same day's RSS also carried ChatGPT Enterprise in AWS Marketplace (19:00 GMT; full text read via the browser channel: purchasable through existing AWS procurement and billing relationships, eligible spend counts toward AWS commitments, more than 1.2 billion weekly users cited — a procurement/distribution announcement, recorded here without a standalone article) and two customer stories (CodeRabbit, Kuken — marketing pieces, excluded). See OpenAI Control Room: a central system to stop model workloads on its research infrastructure within 35 seconds.
• 2026-11-09: The official SDK pair (python v3.48.0 GitHub 16:12:58Z / PyPI 16:13:48 already latest; node v7.54.0 GitHub 16:04:55Z / npm 16:15:07 already latest; checked 11-10 via the GitHub API, both registries, and tag sources) — python-side #4251 adds obfuscation padding to agent session event streams (stream_options on session create, an include_obfuscation query parameter on the events endpoint, an obfuscation field across 33 event types, all defaulting true; node has neither switch as of v7.54.0 and the official events guide omits obfuscation); both packages also take gpt-6-rosalind into the beta Responses model union (#4257/#3126) and retain the agent scheduling APIs (#4255/#3123, observable diffs limited to the codegen stats file). See openai-python 3.48.0 and openai-node 7.54.0: agent session event-stream obfuscation padding and stream_options.
• 2026-11-06: The security partnership with Hacktron (official news RSS entry at 08:00 GMT; the announcement full text read via a rendered-browser channel on 2026-11-08, curl still 403; Hacktron's same-day blog post fully fetchable) — engineers and researchers will jointly pick high-priority areas to test, focusing on how OpenAI systems interact with third-party services, with Hacktron's frontier model access expanded through Daybreak Red; OpenAI says it got to know the firm through its Security Bug Bounty program and lists the move alongside the 11-02 Cathedral Systems security sprint and raised bug bounty rewards. See OpenAI Partners with Hacktron: Joint Testing of Third-Party Service Connections, Frontier Model Access via Daybreak Red. The same day's RSS also carried BioBasicsBench (00:00 GMT entry: a 100-task wet-lab benchmark; article page 403, RSS summary as the boundary; no standalone article here).
• 2026-11-06: Codex CLI 0.167.0 stable ships, three channels within 14 minutes (GitHub 01:40 UTC, npm 01:46 taking over the latest tag, PyPI 01:53; checked 2026-11-08) — the PyPI gap of 0.164-0.166 ended by version skipping (the three versions never shipped there; latest jumped 0.163.0 → 0.167.0); the 0.168 preview line (five alphas across 11-04 and 11-05, numbering skipping alpha.4/alpha.5) closed with no stable; the 0.169 preview line opened at alpha.2 (11-06T22:13 UTC, alpha.3 at 11-07T19:31 UTC, npm alpha dist-tag 0.169.0-alpha.3), with one-line release notes pending a stable for semantic extraction. See Codex Preview Line Jumps from 0.167 to 0.168: Four Releases, No Stable Close (2026) (2026-11-08 update note at the top). Also that day openai-node v7.53.0 (09:06 UTC, #3120) — beta Agents streams move to generic event delivery: in-band errors in events.stream() and streaming sessions.create() enter the iteration loop instead of throwing and closing the stream, while HTTP errors, network failures, and malformed JSON still throw; python has no counterpart release (v3.47.0 remains latest, checked 11-08).
• 2026-11-05: GPT-6 Cyber announced and live the same day (official news RSS entry at 09:00 GMT plus the changelog's November 5 entry; the model card, pricing page, Daybreak guide, and deprecations page fetchable via their .md variants, and the announcement full text plus the system card at deploymentsafety.openai.com/gpt-6-cyber readable through a rendered-browser channel — curl still gets 403 on the announcement page; verified end to end by this site on 2026-11-06) — a cyber-specialized variant of GPT-6 Astra per the system card; 1,050,000 context, $20/$100, both Chat Completions and Responses supported (the GPT-5.6 Cyber card is Responses-only), Preparedness cybersecurity rating up to Critical (5.6 Cyber: High), requiring Daybreak Red plus a separate model-level approval; ExploitGym 48.5% vs 32.1% and ExploitGym Honeypot staying in scope 99.93% without / 100% with deployment prompts; the deprecations page added gpt-5.5-cyber-preview the same day (removal 2027-02-01); the SDKs added it to the Responses model union the same day (python v3.47.0 #4249 / node v7.52.0 #3116; not in the ChatModel enum). See GPT-6 Cyber announced: the Astra-derived cybersecurity variant ships on the API.
• 2026-11-04: Daybreak Red on AWS GovCloud (US-East) (official news RSS entry at 07:00 GMT, "Daybreak Red is now available in AWS GovCloud (US-East)"; article page not fetched by this site, RSS summary as the boundary) — Daybreak Red brings GPT-5.6 Cyber to approved customers in AWS GovCloud (US-East) through Amazon Bedrock for authorized cybersecurity work.
• 2026-11-04: GPT Image 2.6 announced (changelog November 4 entry, fetched 2026-11-06; model cards, the image-generation guide section, and the pricing page verified by this site on 2026-11-09) — GPT-Image-2.6 Sunburst and Flare released for image generation and editing with adjustable complexity, SVG generation, and animated GIF generation; markup-guided editing extends to supported GPT Image 2.5 and 2.6 models (2.6 had surfaced only in an SDK docstring on 11-01, with no catalog entry); pricing verified 2026-11-09: both cards price identically and match the 2.5 rates ($5 text input / $8 image input / $30 image output per million tokens, caches $1.25 / $2), with only generations and edits supported. See GPT Image 2.6 guide: adjustable complexity, SVG and animated GIF output with Sunburst and Flare.
• 2026-11-03 to 11-05: The official SDK batch (python v3.43.0–v3.47.0, node v7.48.0–v7.52.0; GitHub release notes + PR diffs checked 2026-11-06; the node 7.48–7.50 side is already covered by the November 4 article) — v3.46.0 / v7.51.0: the DataResidency union adds au/in/jp regional endpoints (#4246/#3112), gpt-6-rosalind joins the Responses model union (#4243/#3109), and image complexity plus SVG/GIF output controls and agent-turn error expansion land (#4247/#4245, #3114/#3110); v3.47.0 / v7.52.0: gpt-6-cyber joins the Responses model union (#4249/#3116); python v3.45.0 catches up 12-hour caching and project service tiers (#4240/#4239, node 7.50.0 shipped first); python v3.44.0 is the counterpart of node 7.49.0's workload identity providers and federation rules. No standalone article for this batch yet.
• 2026-11-02: Live API outbound calling (Outbound SIP calling) opens to all organizations (changelog November 2 entry, fetched and verified by this site 2026-11-09; entry labeled Model gpt-live-1, API v1/live/sessions) — set transport.type to "sip" to dial an E.164 number through your own SIP provider (TLS signaling via sips: plus SDES-SRTP); call-progress events transport.ringing / answered / failed arrive over the sideband; ringing caps at 3 minutes and calls at 2 hours; as of python v3.47.0 / node v7.53.0 the SDK types are still webrtc-only, so outbound calls go over raw HTTP for now. See Live API outbound SIP calling: dialing real phone numbers with GPT-Live.
• 2026-11-02: The Daybreak / Cyber model cards become fully fetchable in the official catalog (fetched by this site that day; neither the changelog nor the deprecations page carried a related announcement that day) — both alias cards, Daybreak Blue and Daybreak Red, carry Status: Deprecated with no effective or shutdown date; Blue's default snapshot is gpt-5.6-sol and Red's is gpt-5.6-cyber; the GPT-5.6 Cyber card confirms $12.5/$75, $1.25 cached input, a 400K context, Responses-only support, and separate Daybreak approval. See Daybreak and Cyber model cards are live in the official catalog: Blue/Red aliases deprecated, GPT-5.6 Cyber at 400K.

2026-10

• 2026-10-31: The official SDK batch (openai-python v3.42.0 GitHub release 00:26 UTC, on PyPI 00:27; openai-node v7.46.0 00:27 UTC and v7.47.0 04:46 UTC, checked 2026-11-01) — agent sessions gain a background-compaction switch, the auto_review action-review mode, and a spend_control cap; experimental time awareness; a five-endpoint organization external_keys surface; SIP-only DTMF send over the Live sideband; node v7.47.0 types the WebSocket warmup generate flag; python's send queue gains per-sender delivery receipts; the images.edit markup overlay docstring names GPT Image 2.5 and 2.6 (2.6's first appearance; no catalog entry). See openai-python 3.42.0 and openai-node 7.46/7.47: agent compaction and review settings, external keys, SIP DTMF.
• 2026-10-30: GPT-6 Rosalind, the life-sciences model, is released (changelog.md fetched that day) — specialized biological reasoning with scientific tool use; three pricing tiers, gpt-6-rosalind $15/$75 plus gpt-rosalind-research and gpt-rosalind-discovery at $5/$25 each; access limited to organizations approved for Rosalind Research via the trusted-access program, with no standalone model card that day. The same day's news RSS also carried three announcements (article pages 403; no standalone articles here): a shared robots.txt signal for websites' AI training preferences, teen protections with a community safety report, and Fair use benefits everyone. See OpenAI's October changelog, backfilled: GPT-6 Rosalind, ChatGPT subscription API credits, and 12-hour prompt caching.
• 2026-10-28: ChatGPT subscriptions now grant API credits (changelog.md fetched that day) — Pro/Business subscribers can claim up to $500/month in free API credits: Pro $100/$200/$500 tiers map to $100/$200/$500, Business $20/$100 per seat pooled at $500 per workspace per month; credits expire at cycle end without rollover; the official guide also confirms Pro comes in $100/$200/$500 tiers. See the October backfill piece.
• 2026-10-26: GPT-6.1 Terra released into the API (changelog Oct 26 entry, fetched 2026-11-01) — v1/responses and v1/chat/completions; the model card: 1.05M context / 922K max input / 128K output / Jun 1, 2026 cutoff / six effort levels including none (medium default); pricing $1/$5 with 5% cached reads; the official five-model GPT-6 family price table assembled for the first time (astra $10/$50, 6.1-sol and 6-sol $2/$10 each, 6-luna $0.10/$0.50). See GPT-6.1 Terra lands in the API: the model card, official pricing, and the GPT-6 family price table.
• 2026-10-20 to 10-21: 12-hour prompt caching and service-tier controls (changelog.md fetched that day) — request-level ttl defaults to 30m and can take 12h (writes at 2.5x input price; cached reads 0.1x for most models, 0.05x for 6.1 Sol/Terra; the SDK docstring still lags at 30m-only, checked 2026-11-01); Platform settings can restrict available service tiers per organization/project. See the October backfill piece.
• 2026-09-10 to 10-16 (changelog feature-entry backfill): 09-22 released GPT-6 Sol/Luna ($2/$10 and $0.10/$0.50), 09-29 released GPT-6.1 Sol ($2/$10) plus astra Ultrafast, 10-08 6.1-sol Ultrafast (US and EU residency), 10-16 Agents API Ultrafast and prompt-to-voice for all customers (nine new natural voices), the Agents API public beta 09-10, sandbox prewarming with active-turn billing 10-09, and the 10-06 usage-tier consolidation into Build/Launch/Grow; the pricing page newly lists gpt-5.6-cyber / gpt-5.5-cyber ($12.50/$75), gpt-6-astra-law ($10/$50), and gpt-live-1 ($0.05/minute). The item-by-item detail is in the October backfill piece.
• 2026-10-31: The Evals read-only milestone takes effect (official deprecation page verified same-day) — existing evals become read-only (announced 2026-06-03, 150 days later) and the Evals dashboard and API shut down on 2026-11-30 (30 days after that); graders are part of the transition while fine-tuning timelines stay in the self-serve fine-tuning section; the same day also retires v1/prompts and Agent Builder (ChatKit remains available). See OpenAI Evals go read-only on 2026-10-31: the migration checklist before the 2026-11-30 shutdown.
• 2026-10-28: openai-python v3.41.0 (22:26 UTC, on PyPI the same day) — Response.tool_usage hosted-tool usage (separate from base model usage), item_reference typed inputs for Realtime (response.input arrays only), exact metadata filtering for beta agents session lists, and the WebSocket reconnect queue keeping binary frames intact; deprecation annotations: the /v1/audio/translations endpoint marked deprecated in the SDK (shutdown 2027-02-26, migrate to /v1/realtime/translations) and the ChatModel 2027-04-01 row grown to four values. See openai-python 3.40/3.41 and openai-node 7.45.1: breakpoint TTL semantics, tool_usage, and Realtime item references.
• 2026-10-29: The Model Spec update announcement (official news RSS 17:00 GMT entry, An updated Model Spec for more capable and beneficial AI; the announcement page returns 403 to this site, full text fetched from model-spec.openai.com/2026-10-29.html and checked 2026-10-30) — five core values (Be faithful to human intent / Minimize harm / Be honest and transparent / Facilitate human agency / Act with care), the Platform authority tier replaced by Root, and the U18 teen-safety badge plus Red-line principles. See OpenAI's Updated Model Spec (2026-10-29): Five Core Values, Root Authority, and U18 Teen Safety. The same day's RSS also carried the How we detect frontier model misalignment entry (page 403 to us; no dedicated article on this site).
• 2026-10-29: Codex CLI 0.166.0 stable (19:47 UTC, npm packages listed 19:56 UTC; PyPI still at 0.163.0, a three-version gap, checked 2026-10-30) — GPT-6.1 Terra's catalog coverage extends to Vertex AI and Amazon Bedrock (#57293), keyboard transcript selection with integrated search, codex queue --resume, codex doctor --network-only, and resume/fork preserving approvals and reasoning-token accounting. See Codex CLI 0.166.0, Explained: GPT-6.1 Terra in Vertex AI and Bedrock Catalogs, Transcript Search, doctor.
• 2026-10-28: The OpenAI×ServiceNow partnership-expansion announcement — OpenAI models and agents come into ServiceNow products while ServiceNow workflows land in ChatGPT and Codex; the ServiceNow MCP plugin went live in the OpenAI marketplace that day (enhanced capabilities expected late 2026), security capabilities enter early access January 2027 with GA planned Q2 2027, and Codex in AI Workflow Factory is expected in 2027. See OpenAI and ServiceNow expand their partnership: models into ServiceNow, workflows into ChatGPT and Codex.
• 2026-10-27: The OpenAI×Oracle partnership-expansion announcement — Oracle business applications and data come into ChatGPT, AI Workspace is in development (no date), the three new offerings AI Security Mesh / AI SMB Bundle / App Composer carry no availability dates, and U.S. customers can buy OpenAI API, Codex, and ChatGPT Work through Oracle Marketplace. See OpenAI and Oracle expand their partnership: Oracle apps in ChatGPT, AI Workspace and three new offerings.
• 2026-10-27: Codex CLI 0.165.0 stable (20:09 UTC) — a built-in Vertex AI provider (model_provider = "gcp"), GPT-6.1 Terra in the model catalog (models.json checked: 500K Codex-side runtime window), mid-conversation effort changes on by default, memory v2 on by default for fresh profiles. The same day, openai-python v3.40.0 (prompt cache breakpoint TTL options, #4188) and openai-node v7.45.1 (TTL behavior clarification #3057, google-auth-library v11 #3046). For the full SDK read, see openai-python 3.40/3.41 and openai-node 7.45.1: breakpoint TTL semantics, tool_usage, and Realtime item references; for the Codex side, see Codex CLI 0.164–0.165, Explained: GPT-6.1 Terra in the Catalog, a Vertex AI Provider, /presets, memory v2.
• 2026-10-26: The official SDK batch (openai-python v3.38.0 / v3.39.0 + openai-node v7.44.0 / v7.45.0) — the gpt-6.1-terra model identifier enters both SDK type layers the same day (python #4180 / node #3054); delegation controls and prompt cache lifetimes; node v7.44.0 also adds scheduled automations (client.beta.agents.automations, including invocation endpoints and the automations agent-tool variant, #3029) and expands Live Decision item / question types (breaking within a minor, #3034); docs-string updates: gpt-image-1 shut down on 10-23 (replaced by gpt-image-2.5-sunburst / flare) and the documented default for Responses image-generation tools is now gpt-image-2 (both per the node v7.44.0 release notes).
• 2026-10-23: Codex CLI 0.164.0 stable (04:48 UTC, npm packages listed 04:52) — ten numbered /presets slots, inline /model and /permissions, Kitty/Sixel terminal image previews, enterprise MCP auth, credential masking, and the Bedrock wizard on by default; PyPI lags (neither 0.164.0 nor 0.165.0 ever listed there, checked 2026-10-29). See Codex CLI 0.164–0.165, Explained: GPT-6.1 Terra in the Catalog, a Vertex AI Provider, /presets, memory v2.
• 2026-10-16: The official SDK batch (openai-python v3.32.0 / openai-node v7.35.0): prompt-based voice creation (voices.create gains a type "prompt" variant; the docs string says such voices work only in Live); Live audio fixtures (session.output_audio.fixture.play, raw PCM16 / 30 seconds max / one active fixture); Decisions in Live (transcript triggers, a decisions delegation target, and decision.item events); prompt caching for the Decisions API (gpt-5.6+); the error object on Responses / beta / videos gains param and type; project API key scope and expiry updates decoupled; ChatModel deprecation annotations updated (see the retirement entry below). See openai-python 3.32.0 and openai-node 7.35.0: prompt-created voices, Live audio fixtures, and Decisions in Live.
• 2026-10-16: The Codex CLI 0.164.0 preview line opens (rust-v0.164.0-alpha.1 / alpha.2; still prerelease as of 2026-10-17).
• 2026-10-15: Codex CLI 0.163.0 stable — Daybreak on by default for eligible accounts, instant steering on by default, /config preference search, and more. See Codex CLI 0.163.0, Explained: Daybreak On by Default, Instant Interrupts, /config Search.
• 2026-10-14: Four GPT-4 series family rows announced deprecated, shutting down 2027-01-14 with the official replacement gpt-6-luna; fine-tuned versions follow the base models. The same day, browser/computer toolsets (1/4) enter the official SDKs (python v3.31.0 / node v7.33.0) - see Browser and Computer Toolsets (1/4) Land in the OpenAI SDKs: BrowserToolset Turns Driver Methods into Function Tools.
• 2026-10-01 (announcement date per the page text; back-filled by this site on 2026-10-17): Two retirement batches - (1) gpt-5.3-codex, gpt-5.1, and gpt-5.4-nano with six months' notice, shutting down 2027-04-01 (replacements gpt-6-sol / gpt-6-luna); (2) the TTS four-pack (tts-1, tts-1-hd, gpt-4o-mini-tts-2025-03-20, gpt-4o-mini-tts-2025-12-15) shutting down 2027-01-06, all to gpt-realtime-2.1-mini. Also gpt-5.4-cyber (announced 2026-09-11) was removed from the API on 2026-10-01. See OpenAI's October 2026 deprecation wave: GPT-4o and GPT-4.1 on 2027-01-14, TTS to gpt-realtime-2.1-mini.
• 2026-09-22 to 10-09 (timeline backfill): the GPT-6 family grows (gpt-6-sol / gpt-6-luna / gpt-6.1-sol land in the official SDK types) and Codex switches its default model (GPT-6.1 Sol from 0.161). See The GPT-6 family grows: gpt-6-sol, gpt-6-luna, gpt-6.1-sol land in the official SDKs and Codex CLI 0.156–0.162, Explained: Voice by Default, GPT-6.1 Sol as Default Model, Worktrees and Command Center.

2026-09 (latest)

Source note: the official changelog domain has returned 403 to this site since 2026-09-02 (last successful fetch 2026-09-01; re-checked and still unreachable on 2026-09-15, 09-17, and 09-19), so entries in this section draw on the official GitHub SDK release notes and type sources, each checked against the source; changelog quotes and GA status will be back-filled once reachable.

• 2026-09-18: The official SDK feature batch + Codex 0.155 stable (openai-python v3.15.0 / v3.16.0 / v3.16.1 / v3.16.2, openai-node v7.18.0 / v7.19.0; Codex CLI rust-v0.155.0 / rust-v0.155.1): prompt_cache_options gains prewarm warming (default false; true prepares the cache without generating output and forces the WS generate field to false); a new top-level client.webhooks resource (seven operations around POST /webhook_endpoints, 18 subscribable event types, signing_secret returned in plaintext only on create and rotate); the MCP tool connector_id marked deprecated (for models released after 2026-09-01 — migrate to server_url or tunnel_id); managed Responses WebSocket sessions in both languages (lane routing + final-response collection); client.beta.agents.sessions.update accepts model settings for subsequent turns; the Responses model union adds gpt-audio-mini / gpt-audio-mini-2025-12-15; python v3.15.0 catches up on the compaction progress event (aligning with node v7.17.0, #3866); on the Codex side, experimental /voice enters the TUI and Touch ID can verify MCP requests, with 0.155.1 restoring none as the reasoning-summary default. See openai-python 3.15/3.16 and openai-node 7.18/7.19: Cache Prewarming, Webhook Management, connector_id Deprecation and Codex CLI 0.155, Explained: Experimental /voice, Touch ID for MCP Requests, and the 0.155.1 Revert.

• 2026-09-16: openai-node v7.17.0 adds the compaction progress event — response.compaction.compacting: while the server processes a compaction_trigger input item it reports compaction progress at most once every 30 seconds, carrying no summary content and not modifying the compaction output item (completion is still marked by output_item.done with the final encrypted content, and a short compaction may emit no progress event at all); the WebSocket variant adds an optional stream_id and the beta multi-agent variant an optional agent.agent_name. Not landed on the Python side as of v3.14.1 (09-15) — closed on 09-18 by python v3.15.0, see the entry above. See The Responses API Compaction Progress Event: response.compaction.compacting and compaction_trigger.

• 2026-09-14 / 09-15 / 09-16: The official SDK reliability wave (openai-python v3.14.0 / v3.14.1, openai-node v7.16.0 / v7.17.0): Python normalizes stream consumption exceptions (read timeouts raise APITimeoutError, other httpx failures raise APIConnectionError; streams are never auto-retried), stringifies APIStatusError.code (404 → "404"), validates max_retries before the request (non-negative integer; 0 disables retries), and stops structurally parsing non-final_answer phases (commentary) in responses.parse; Node adds maxBufferedEvents backpressure to WebSocket iterators (overflow sheds the backlog and rejects next() with WebSocketError), flushes a terminal SSE event missing its trailing blank line exactly once at EOF, and rejects function-backed API keys without a resolved credential at WebSocket construction. See openai-python 3.14.x and openai-node 7.16/7.17: Stream Error Normalization, WebSocket Backpressure, and SSE Fixes.

• 2026-09-09 / 09-10: GPT-6-Astra becomes selectable (Codex CLI 0.154.0, 2026-09-09T22:35 UTC) — release notes verbatim: "GPT-6-Astra is now available in the model picker and Amazon Bedrock catalogs". The model picker ranks it first (models.json priority=1, with the GPT-5.6 trio dropping to 6-8; PR #42879); the Bedrock catalogs carry openai.gpt-6-astra with global and US cross-region variants (PR #42619); the bundled official migration guide (PR #42931, replacing the GPT-5.6 Sol edition) confirms Responses API calls with model: gpt-6-astra: six reasoning efforts from low to ultra (ultra = automatic task delegation; none unsupported), async tool calling, mid-turn steering over WebSocket, and configuration_update mid-conversation effort changes. Codex metadata lists a 272K context with an 872K experimental cap. Role mapping: Sol→Astra, Terra/Luna retained; no pricing figure in any loadable channel as of 2026-09-15. See GPT-6-Astra Goes Live: Codex Model Picker and Amazon Bedrock Catalogs.

• 2026-09-10: Codex SDK 0.154.0 stable (Python openai-codex / TypeScript @openai/codex-sdk; npm latest checked 2026-09-15) — the reasoning ladder gains max and ultra above minimal/low/medium/high/xhigh (the TypeScript seven-value type has no none; the official Python example's REASONING_RANK lists nine names from none to ultra); ExternalMessage: untrusted content from other agents, tools, or services enters a turn with tool-level authority, never user authorization, requiring CLI 0.151.0+; include_turns / turn_service_tier / source options; HookMetadata access moves under .root. The same CLI release adds experimental worktrees (--worktree / /worktree) and inline Q&A during running tasks, and removes the codex mcp-server entry point. See Codex CLI 0.154.0 and SDK 0.154.0: Worktrees, ExternalMessage, and the ultra Reasoning Effort.

• 2026-09-10: The Live API enters the official SDK types (openai-python v3.12.0 at 17:28 UTC / openai-node v7.14.0 at 17:29 UTC, "Add Live API") — a client.live namespace where POST /live/sessions submits a WebRTC SDP offer and a WebSocket connect() path; the sessions subresource carries the four SIP call-control endpoints (accept / reject / hangup / refer) plus fork and download_recording, and sideband attaches to an existing session. The session-config model field gains the literal gpt-live-1 (not in the ChatModel enum, not an announced model). Availability and pricing unannounced as of 2026-09-11. See The Live API lands in the OpenAI SDK: gpt-live-1, dual WebRTC/WebSocket channels, and SIP call control.

• 2026-09-10: The Agents API enters the official SDK types (openai-python v3.13.0 / openai-node v7.15.0, the same minute at 19:37 UTC, "add Agents API") — the beta namespace client.beta.agents with top-level HTTP endpoints /agents and /vaults: four resource surfaces spanning reusable Agent CRUD, Managed Agents sessions (status machine idle / in_progress / requires_action / failed), execution environments (templates / files / skills / plugins), and credential vaults. Fifteen days after the Assistants API shutdown (2026-08-26), with no official statement on the relationship. Availability and pricing unannounced as of 2026-09-11. See The Agents API appears in the OpenAI SDK (beta): /agents CRUD, environments, sessions, and vaults.

• 2026-09-08: The two GPT Image 2.5 models enter the official SDK types — gpt-image-2.5-sunburst and gpt-image-2.5-flare (each with a -2026-09-08 snapshot). quality gains xhigh / max on top of low / medium / high; transparent background no longer carries the preview note on 2.5 (gpt-image-2 still carries it); pricing and the models' positioning split were unverified that day. See GPT Image 2.5 Explained: sunburst and flare, xhigh/max Quality, Arbitrary Resolutions.

• 2026-09-08 / 09-09: Four SDK releases in two days (openai-python v3.9.0 / v3.10.0, openai-node v7.11.0 / v7.12.0): prompt cache diagnostics (prompt_cache_options.comparison_response_id → prompt_cache_diagnostics, cache_miss with 9 reason values), Service Account API key expiry fields (expires_in_seconds / expires_at), and documented 429 / 503 rate-limit and overload responses. See openai-python 3.9/3.10 and openai-node 7.11/7.12: Prompt Cache Diagnostics, API Key Expiry, GPT Image 2.5.

2026-08

Fetched from the official changelog on 2026-09-01, each entry checked verbatim.

• 2026-08-29 — Mutual TLS (mTLS) and X.509 workload identity federation are now generally available for the OpenAI API. Certificates and X.509 identity providers are configured in the Platform console, with access controlled by org roles and permissions.
• 2026-08-26 — Four transcription models announced deprecated (whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize); shutdown on 2027-02-26, migrate to gpt-live-transcribe / gpt-transcribe. The Assistants API shut down the same day; migrate to the Responses API / Conversations API.
• 2026-08-21 — GPT-5.6 Sol price cut to $4/$20 per MTok (input -20%, output -33%); promotional pricing at least through 2026-11-21. Same day: per-request regional processing (prefixed domain + Global geography project API key).
• 2026-08-20 — Prompt Caching dashboard released (hit rate over time, cache reads per write, cache-read/write/uncached breakdown, filterable by model and service tier). Same day: gpt-image-2 transparent backgrounds enter preview (background=transparent, png/webp; jpeg unsupported).
• 2026-08-13 — Ultrafast mode announced -- a GPT-5.6 Sol-only API service tier, up to 14x Standard, limited preview.
• 2026-08-07 — Daybreak tiers -- Daybreak Blue (general-purpose models such as GPT-5.6 Sol, defensive security work) and Daybreak Red (separately approved, purpose-trained models such as GPT-5.6 Cyber, authorized pentesting and more).
• 2026-08-06 — chat-latest snapshot updated; OpenAI still recommends GPT-5.6 Sol for production API usage.
• 2026-08-05 — Fast mode gains long context -- all three tiers can run prompts beyond 272K tokens in Fast mode, up to 2.5x Standard.
• 2026-08-04 — Usage and Costs dashboards gain the API-key dimension; the Usage API and Costs API support it too.

2026-07

• 2026-07-09 — Released the GPT-5.6 model family, including GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna.

  • Impact: supersedes the 2024-2025 mix of GPT-4o / o1 / o3 / GPT-4.1 / GPT-4.5. Three tiers share 1.05M context + 128K output + knowledge cutoff 2026-02-16; depth is now controlled continuously via reasoning.effort.

• 2026-07-06 — Released GPT-Realtime-2.1 and GPT-Realtime-2.1 mini.

  • Impact: voice flagship refresh; mid-conversation tool use stabilised; WebRTC / WebSocket supported.

2026-06

• 2026-06-01 — GPT-5.4 and GPT-5.5 added to Amazon Bedrock.
• 2026-06-02 — Container sessions billing changed: "Eligible container sessions will be billed per minute with a 5-minute minimum, instead of being billed at the full 20-minute session rate."
• 2026-06-03 — Announced the deprecation of reusable prompt objects, the Evals platform, and Agent Builder.
• 2026-06-24 — Updated the chat-latest snapshot.

2026-05

• 2026-05-05 — Released chat-latest snapshot.
• 2026-05-07 — Released GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper.
• 2026-05-12 — DALL·E 2 / DALL·E 3 deprecated and removed from the API; Realtime API Beta deprecated and removed.
  • Impact: image generation migrates to GPT Image 2 (released 2026-04-21); voice migrates to the GPT-Realtime-2.x family.
• 2026-05-28 — Released chat-latest snapshot.

2026-04

• 2026-04-21 — Released GPT Image 2.
  • Impact: current flagship image generation model; check Usage Policy before commercial use.
• 2026-04-24 — Released GPT-5.5 and GPT-5.5 Pro. "GPT-5.5 supports a 1M token context window."

2026-03

• 2026-03-03 — Released gpt-5.3-chat-latest.
• 2026-03-05 — Released GPT-5.4 and GPT-5.4 Pro.
• 2026-03-12 — Sora API gained character references; sora-2 / sora-2-pro now support 1080p at $0.70 per second.
• 2026-03-16 — Updated the gpt-5.3-chat-latest slug.
• 2026-03-17 — Released GPT-5.4 mini and GPT-5.4 nano.

2026-02

• 2026-02-03 — GPT-5.2 and GPT-5.2-Codex run about 40% faster.
• 2026-02-10 — Updated the gpt-5.2-chat-latest slug.
• 2026-02-23 — Released GPT-Realtime-1.5 and gpt-audio-1.5.
• 2026-02-24 — Released gpt-5.3-codex.

2026-01

• 2026-01-14 — Released gpt-5.2-codex.

2025 historical (retired)

• GPT-5 (2025-08): first GPT-5 release; superseded by GPT-5.6 family on 2026-07-09.
• OpenAI o3 / o3-pro (2025-06): reasoning lineup; later folded into the GPT-5.6 reasoning.effort continuum.
• GPT-4.5 (2025-03): research preview, retired.
• GPT-4o / o1 / o3 / GPT-4.1 (2024–2025): the multi-family layout that GPT-5.6 collapses.

One-line current state

Flagship family = GPT-5.6 (Sol / Terra / Luna, 2026-07-09); since 2026-09-09 the next-generation flagship GPT-6-Astra is selectable in the Codex picker and Amazon Bedrock (the bundled official migration guide confirms Responses API usage and maps Sol→Astra with Terra/Luna retained; no pricing figure as of 2026-09-15). Voice = GPT-Realtime-2.1 (2026-07-06). Images = GPT Image 2.5 (sunburst / flare, in the SDK since 2026-09-08; the previous GPT Image 2 (2026-04-21) remains on sale, not deprecated). DALL·E 2/3 and the Realtime API Beta were both removed on 2026-05-12.

How this is built: fetch developers.openai.com/api/docs/changelog → extract (date, quote) per entry → place in the matching month block. dataSourcedAt records the most recent fetch.