← the radar · releases

New AI model releases

Every model ArtificialWatch has seen go live, newest first, and the frontier models it is watching for next. A release date here is the day the model went live as ArtificialWatch verified it, not an announcement. Updated as models ship.

Get the next launch alert, free →

This month so far: every model released in November 2026. This year: AI models released in 2026, also as a timeline.

This week

4 releases in the last 7 days.

• Nov 5Gemini Omni 1.2 Flash · Google
• Nov 5Inkling 2 Small · Thinking Machinesbench index 53
• Nov 4GPT Image 2.6 · OpenAI
• Nov 4Ripple-65B · Ricursive

Earlier in the last 30 days

18 releases since October 10, 2026, 14 of them before this week.

• Oct 30GPT-6 Rosalind Research · OpenAI
• Oct 26GPT-6.1 Terra · OpenAI$1 / $5 per 1M tokens · bench index 77.6
• Oct 23Kimi K3.1 Pro · Moonshotfrontier · $2 / $10 per 1M tokens · bench index 66.7
• Oct 21Flint · Recursive
• Oct 19Mercury 3 · Inception
• Oct 19MiniMax M3.1 Flash · MiniMax$0.15 / $0.60 per 1M tokens · bench index 38.5
• Oct 16Beam · Reflection AIbench index 7.1
• Oct 16Glimpse · Elorian
• Oct 15Muse Spark Open · Metabench index 14.6
• Oct 15Nemotron 3.5 Bolt · NVIDIA
• Oct 15Nemotron 3.5 Super VL · NVIDIA
• Oct 14Grok Voice 3.0 · xAI
• Oct 13Gemini 4 · Googlefrontier · $2 / $10 per 1M tokens · bench index 80
• Oct 12Muse Wonder · Metafrontier · bench index 74.1

Coming next: frontier models

Not out yet. Odds are live Polymarket prices where a market exists: sentiment, not a promise. ArtificialWatch alerts the moment one answers on a public API.

• Fable 5.2 · Anthropic96% by Dec 31
  Unannounced - armed on Anthropic's point-release cadence, not a leak
• GPT-6.1 · OpenAI97% by Dec 31
  SCRAPPED per the WSJ (Sep 28): OpenAI pulled its planned October release after the model fell short in internal alignment tests and showed more deception than GPT-6 (safety head Saachi Jain on record) · still armed, since a shelved id can come back · Kalshi runs a market on 'Astra 6.1'
• Grok 4.8 · xAI64% by Nov 30
  Unannounced, no date — xAI's own public grok-build repo lists grok-4.8 in its model picker and router tests (read Sep 29; the tests use 500K and 256K context windows) · Nov 4: 44 days since Grok 4.7 shipped on Sep 21, already longer than the 34 days from 4.5 to 4.6 and the 40 from 4.6 to 4.7 · xAI's API release notes (newest entry Nov 2) and OpenRouter's catalog list no grok-4.8 · Polymarket's "next Grok (4.8+) by Nov 30" market: 65% (Nov 4) · Nov 7: 47 days since Grok 4.7, still no grok-4.8 in xAI's API release notes (newest entry Nov 2, an image-model retirement) or in OpenRouter's catalog (newest xAI id grok-4.7), and xAI's public grok-build repo, where the 4.8 picker entry appeared, has had no push since Sep 29 · Musk's posts about Grok since Nov 4 cover Grok Bot features and give no 4.8 date (his X posts, read Nov 7) · the Polymarket market last traded at 64%, with the book at 59-61% (Nov 7)

  Earlier reports (3)

  a Cursor server list reportedly carries grok-4.8-high, -high-fast and -xhigh-fast (single source) · Musk: ~2.5T params on a new C++ stack · Oct 7: Musk says most Grok Bot requests "will be handled by a lightning-fast version of Grok 4.8 when that comes out" (his X post) · Oct 30: Musk: "I know it's taken longer than expected, but Grok 4.8 will not be a disappointment", adding that xAI is not optimizing for benchmarks (his X post; still no date)
• Grok 5 · xAI46% by Mar 31
  No date · Musk on record: 6T parameters (Nov 17, 2025); on Sep 14 he put Grok 4.8 and 4.9 ahead of it
• Kimi K4 · Moonshot13% by Mar 31, 2027
  Not announced — next generation after K3 · Oct 15: @lyraxana says Moonshot's next flagship, which they say is codenamed x4 internally, has finished pre-training but is still in mid-to-late post-training, so not soon (one account, unconfirmed)
• Next Claude Opus (5.6+ / 6) · Anthropic47% by Nov 30
  Unannounced — armed on the cadence the moment 5.5 shipped, not on a leak · Oct 11: users on X (@Mr_Salio, @kimmonismus, @koltregaskes) report Opus 5.5 and Fable 5 on claude.ai answering from a newer checkpoint, with a 'VoxelBench' knowledge question going round as the tell; unconfirmed, Anthropic silent, no new Opus id on OpenRouter (Oct 11) · Oct 29: @legit_api and @Mr_Salio report a new Opus checkpoint in early testing alongside a newer Fable one; unconfirmed, and no new Opus id on OpenRouter or models.dev (read 15:00 UTC) · Nov 7: 46 days since Opus 5.5 (Sep 22); Anthropic's last two Opus gaps were 58 days (Opus 4.8 to Opus 5) and 60 (Opus 5 to Opus 5.5), so the same pace would put the next one around Nov 19-21 · Anthropic's models overview names no Opus 5.6 or Opus 6, its release notes list no model launch since Haiku 5.5 (Oct 7), and OpenRouter and models.dev list no Opus newer than 5.5 (read Nov 7) · Anthropic's next dated event is The Briefing: Work at the Frontier, an enterprise livestream on Nov 10 (9:30 am PST) whose page names no model
• Composer 3 · Cursor
  Cursor's from-scratch model, trained on SpaceX's Colossus
• DeepSeek V5 · DeepSeek
  No date · DeepSeek's named next model is V4.1-Pro (Sep 10); Liang told investors (Sep 21) its next models will train on Huawei chips from late 2026 · Oct 19: The Information reports DeepSeek is working closely with Huawei to shape the design of its future AI chips, per two people familiar with the matter · Nov 3: STAR Market Daily reports, citing several supply-chain sources, that DeepSeek plans to train V5 on Huawei Ascend chips (DeepSeek has not confirmed)
• Fable 5.5 · Anthropic
  Unannounced · Nov 5: METR's independent report on AI R&D acceleration at Anthropic charts internal-only models under codenames Anthropic gave METR: Model 2 (finished training July 2026), Model 3 (Aug-Sep) and Model 4 (Sep-Oct), with Model 4 at 178.6 on Anthropic's AECI capability index and over 50% on every AECI benchmark it was tested on; METR does not say which, if any, becomes Fable 5.5 or any other product · Nov 7: still no Fable 5.5 or Mythos 5.5 in Anthropic's models overview, on its Mythos page (Mythos 5.1 only) or in its newsroom (last model post: Haiku 5.5, Oct 7), nor on OpenRouter or models.dev; 40 days since Sonnet 5.5, Anthropic's longest wait between frontier releases since July · its next dated event is The Briefing: Work at the Frontier, an enterprise livestream on Tuesday, Nov 10 (9:30 am PST) with Chief Product Officer Ami Vora; the event page names no model · Polymarket's next-Fable (5.2+) market last traded at 29% by Nov 9, 45% by Nov 10, 58% by Nov 12 and 90% by Nov 30

  Earlier reports (12)

  Opus and Sonnet both went 5 -> 5.5 (Sep 22, Sep 28); on Oct 1 users on X reported claude.ai routing to a 'Fable 5.5' (unconfirmed, no id in any catalog) · the same reports returned Oct 9 (@Mr_Salio) and Oct 11 (@kimmonismus, citing many users); Anthropic has said nothing, and OpenRouter still lists only Fable 5 and 5.1 (Oct 11, again Oct 15) · Oct 13 12:27 UTC LiteLLM's sync bot added azure_ai/claude-fable-5-5 from an Azure Models API row (eastus2, marked GA); LiteLLM pulled it Oct 14 16:17 UTC after an audit found no such model on Azure and a test call returned 404 DeploymentNotFound (LiteLLM PRs #46201, #46428) · Oct 15: Artificial Analysis's site code carried an unreleased Claude slug, claude-onigiri-eap-2, in low/medium/high/xhigh variants with the same Opus 4.8 fallback as Fable 5, Fable 5.1 and Opus 5.5 (read 10:50 UTC); @LuminaBench, who found it, thinks it is Fable 5.5, which is unconfirmed · AA removed it in a new build by 13:50 UTC (old file 404s) · the same afternoon @Mr_Salio and @LuminaBench relayed new reports of Fable 5.5 served in place of Opus 5.5 on some accounts and in Claude Code (unconfirmed) · Oct 15 14:01 UTC @lyraxana says partners got access to Fable 5.5, with no release date and a guess of next week (one account, unconfirmed) · Oct 18 18:45 UTC the same account says partners now have a second Fable 5.5 checkpoint that looks stronger than the first (unconfirmed) · Oct 19: The Information reports Anthropic is preparing to release a new model that is better in reasoning and intuition than its current ones; what we could read names no model and no date · Oct 20: two commits dated Oct 17 in the public pingdotgg/t3code repo (PR #20424, merged Oct 20 04:36 UTC) carry a 'Co-authored-by: Fable 5.5' trailer at Anthropic's noreply address, the one Anthropic's Claude Code adds, which fits the partner-access reports (read on GitHub; surfaced by @LuminaBench) · the same morning @kimmonismus, quoting another account, says it ships today (unconfirmed) · Oct 20 13:01 UTC: Anthropic's research post on scorpion-toxin binders credits 'an internal research Claude model', working in Claude Science for 64 hours (318,484 designs, 17 of 211 lab-tested binders met its criteria); the post names no model, and @Mr_Salio's reading that it is Fable 5.5 is unconfirmed · Oct 25 20:14 UTC: @ChrisGPT (an AI reporter account) says, citing sources, that Fable 5.5 has been rolled out to enterprise customers, naming JPMorgan and Meta among 'top clients'; one account, unconfirmed, and Anthropic has said nothing · @kimmonismus, relaying it, expects a consumer release in the coming week (his reading) · Oct 26 13:00 UTC: Anthropic's public S-1 names Mythos 5.1, Fable 5.1 and Opus 5.5 as its latest releases and does not mention Fable 5.5 · @Mr_Salio expects Fable 5.5 to ship before the IPO (his reading, no source given) · Oct 28 13:00 UTC: a Wayback Machine copy of Cognition's FrontierCode 2.0 results file (Oct 27 19:48 UTC) lists a hidden 'Obsidian 3' in the claude-code harness at 46.8% best, above Opus 5.5's 45.5%, and the row is gone from the current file · @kimmonismus reads it as Fable 5.5 (his reading); the file draws it in Muse Wonder 1.0's blue, not the Claude rows' orange, so who made it is open · Oct 29: @legit_api and @Mr_Salio report a newer Fable checkpoint and a new Opus checkpoint in early testing, going by SVG outputs shared on Discord; unconfirmed, Anthropic silent, OpenRouter and models.dev still list no Fable 5.5 (read 15:00 UTC)
• Fable 6 · Anthropic
  No date · Anthropic has said nothing on Fable 6; Fable 5.1 and Mythos 5.1 shipped Sep 1 · Nov 4: 64 days since Fable 5.1 and 37 since Anthropic's last frontier release (Sonnet 5.5, Sep 28; its longest gap since July is 39 days); no fable-6 id on OpenRouter or models.dev; the next name Anthropic has shown is 5.5, after its Mythos page read 'Claude Mythos 5.5' on Nov 1 before it was changed back · Nov 7: 67 days since Fable 5.1 and 40 since Sonnet 5.5, now Anthropic's longest wait between frontier releases since July (39 before); Anthropic's models overview, Mythos page and newsroom name no Fable 5.5, Fable 6 or Mythos 5.5, and OpenRouter and models.dev top out at Haiku 5.5 (Oct 7) · Anthropic's next dated event is The Briefing: Work at the Frontier, an enterprise livestream on Nov 10 (9:30 am PST) whose page names no model · Polymarket's next-Fable (5.2+) market last traded at 45% by Nov 10 and 90% by Nov 30
• Fable 6.1 · Anthropic
  No date · Fable 6 itself is unannounced (latest: 5.1, Sep 1); Reuters: Anthropic weighing a new model before its IPO (Sep 18)
• Gemini 3.5 Pro (gemdelta) · Google
  Internal id per an Aug 6 leak — not on the public arena board when checked
• Gemini 4 Carbon · Google
  Announced Oct 28, not public yet · Google calls it its strongest model, state-of-the-art on cybersecurity and long-horizon tasks (Google blog) · access started the week of Oct 28 for a subset of cyber defenders, then more trusted partners through the Fairwind Program; developers, businesses and consumers get it 'as soon as possible', no date — that week is now OVER and nothing opened: re-searched Nov 3, still no public Gemini API entry, no model id and no Carbon row on Google's pricing page, so this stays watch · no gemini-4-carbon id in the Gemini API changelog, OpenRouter or models.dev (read Oct 28, 18:05 UTC) · price at launch: $2 / $10 per 1M introductory, the same as Argon, cached input 95% off; $4 / $20 after the introductory period · Artificial Analysis Intelligence Index v5.0 (read Oct 28): 59.4 at high, first, ahead of Claude Opus 5.5 (58.4) and GPT-6 Astra (56.9); AA marks it as having no publicly available API · Google's own figures: SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0%, #4 on WebDev Arena at 1769 (+89 over Argon) · Argon took the same path: announced Sep 30 for Fairwind testers, GA Oct 13 · before the announcement: Business Insider (Oct 9) reported Carbon as an internal Gemini 4 checkpoint, and @LuminaBench showed gemini-4-carbon references in Antigravity build 1.4.3 (Oct 24) · Oct 28 18:58 UTC: the launch alert was fired by hand on Mark's call while Carbon was still trusted-access only (event aw_mvsha1jm_when9huu1i, 374 notifications); the entry stays watch because it is not on a public API, and its primary one-shot is spent, so the API opening will not alert again · Oct 30: first on Artificial Analysis' new HLE-Diamond (read directly) at 69.1% at high, ahead of Opus 5.5 at max (67.9%); reasoning 79.6%, knowledge 58.6% · Cognition FrontierCode 2.0 (Carbon added Oct 28, read Oct 30): 40.0% at medium in Devin CLI, fourth, behind Opus 5.5 (45.5%), GPT-6 Astra (42.8%) and Sonnet 5.5 (41.0%), $4.04 a task · Nov 7: still not public, 10 days after the announcement (Argon took 13, Sep 30 to Oct 13); the Gemini API changelog (newest entry Nov 6), models list and pricing page have no Carbon entry, and neither OpenRouter nor models.dev lists a gemini-4-carbon id; Google has given no date for developer access
• Gemini 5 · Google
  Not announced — no Google word or date; Gemini 4 Argon shipped Oct 13
• Gemini experimental release · Google
  Unnamed early build · Google shipped these as gemini-exp-MMDD before Gemini 2; Gemini 4 then shipped Oct 13 under its own name, Argon
• Gemini Krypton · Google
  Rumored, not announced · Nov 9 06:12-07:55 UTC: @chetaslua posts outputs he says come from 'Gemini Krypton' ('krypton-a version first checkpoint'), compares it with Gemini 4 Carbon and Argon, and says it takes about twice as long on the same prompt; he does not say where he runs it; @Mr_Salio relayed it · one account, unconfirmed, and Google has not named a Krypton · Google's Gemini 4 builds use element codenames: Argon shipped as gemini-4-argon (Oct 13), Carbon was announced Oct 28, Barium was reported by Business Insider (Oct 9) · whether Krypton would ship as a Gemini 4.x or a Gemini 5 is unknown; this hunt fires on any gemini id containing krypton · Nov 9 10:16 UTC: a second account, @lyraxana, posts side-by-side videos labelled 'Gemini 4 Krypton vs Carbon vs Argon', the first use of a Gemini 4 prefix for it; still unconfirmed, Google has said nothing · Nov 9 11:27-11:28 UTC: @LuminaBench and @kimmonismus amplified the outputs; kimmonismus says he cannot verify them · no krypton id on OpenRouter or models.dev (11:49 UTC)
• Gemini Omni Pro · Google
  Leaked — expected to ship alongside Gemini 3.5 Pro, and that anchor is gone: our own 3.5 Pro row records it reported cancelled (9to5Google, Sep 30; no Google statement) and Google stepping back from 3.5 Pro to focus on Flash (Sep 24), with Gemini 4 taking the top slot as Argon on Oct 13 · Nov 8: the Omni line has shipped twice since the leak with no Pro tier (Omni 1.1 Flash GA Aug 27, Omni 1.2 Flash Nov 5), and our sweep's cumulative ledger of 627 catalog ids has never carried an Omni Pro id — its only Omni ids are gemini-omni-1.1-flash and gemini-omni-1.2-flash · hunt left armed, since a Pro tier can still arrive on its own
• GPT-5.6 Zinc / Magnesium · OpenAI
  Codenames sighted disabled in DesignArena — Zinc rumored an upgraded Sol, Magnesium a Terra replacement
• GPT-6.2 · OpenAI
  Not announced — no OpenAI word or date · Nov 5 02:23 UTC: @Mr_Salio says both ids have since been removed from the quota page; we could not check that either · Nov 7: Friday, Nov 6, passed without a new Astra or Sol; OpenAI's launch that week was GPT-6 Cyber (Nov 5, to approved Daybreak Red customers only) · no gpt-6.2 id on OpenRouter (its newest OpenAI ids are GPT-6.1 Terra and Terra Pro, Oct 26) or models.dev · Polymarket's 'When will GPT Astra 6.1+ be released?' market, which a GPT-6.2 Astra would also settle, last traded at 56% for Nov 9 to 15 and 2% for this week, and its 'Next GPT Sol (6.2+)' market at 65% by Nov 15

  Earlier reports (10)

  Oct 29: one X account (@Josh67Joshua) said OpenAI's Bel codename is short for 'GPT-6.2 Betelgeuse' · Oct 29: a second account (@taker_of_whizz) posted a tracker screenshot listing gpt-6.2-betelgeuse as INTERNAL; both unconfirmed, and OpenAI has named neither · Nov 1: @TokenGremlin, @Mr_Salio and @LuminaBench posted screenshots of a gpt-6.2-astra id in Codex usage data and responses (picker still says GPT-6 Astra); unverified, none in OpenRouter or models.dev · Nov 1 19:28 UTC: our own `codex --model gpt-6.2-astra` on a ChatGPT account returned 400 model_not_found ('not supported when using Codex with a ChatGPT account'), despite posts claiming it runs in the CLI · Nov 3: @hakmgpt, @kimmonismus, @Mr_Salio and @LuminaBench say ChatGPT's GPT-6 Astra is routing some users to GPT-6.2 and posted sample outputs; unconfirmed, OpenAI has said nothing and no gpt-6.2 id is in OpenRouter or models.dev · Nov 3 later: @LuminaBench called the circulating GPT-6.2 Astra demos fake, and @Mr_Salio apologized, saying the SVG clip he posted had been enhanced beyond the model's original output; treat the samples as unreliable · Nov 3: two OpenAI API ids, rc-ada-or-vm and rc-ada-sh-vm, sat on OpenAI's model list from about 18:41 to 18:46 UTC (github.com/spullara/models history; our radar saw both, @LuminaBench flagged them); what they are is unknown, and nothing ties them to GPT-6.2 · Nov 3: OpenAI's own blog scored an unnamed 'newer internal model' at 0.875 on its BenchCAD 2.0 test against 0.594 for GPT-6 Astra; OpenAI did not call it GPT-6.2 · Nov 4: The Information reported that new OpenAI models could come as soon as Friday, Nov 6, including the next generation of its Astra, Sol and cyber-focused models, citing people with knowledge of them; OpenAI has not confirmed, The Information did not use the GPT-6.2 name, and no gpt-6.2 id is in OpenRouter or models.dev (16:31 UTC) · Nov 4 23:28 UTC: @lyraxana and, 16 seconds later, @AiBattle_ posted screenshots of Google Cloud's quota page listing base_model values gpt-6.2-astra and gpt-6.2-sol, and @AiBattle_ says a release is likely Nov 5; screenshots only, we could not read the page ourselves, and OpenAI has not announced either
• GPT-6.5 · OpenAI
  Not announced — no OpenAI word or date · Nov 4: OpenAI's API changelog (newest entry: GPT-6 Rosalind, Oct 30) and OpenRouter's catalog list no gpt-6.5 id · OpenAI's Nov 3 blog scored an unnamed 'newer internal model' above GPT-6 Astra on its BenchCAD 2.0 test without naming it; that signal is tracked on the GPT-6.2 page · Nov 7: OpenAI's API changelog still has no gpt-6.5 id; its newest entries are GPT-6 Cyber (Nov 5, for approved Daybreak customers) and two image models (Nov 4) · OpenRouter's newest OpenAI ids are still GPT-6.1 Terra and Terra Pro (Oct 26), and models.dev lists no gpt-6.5 either · OpenAI's next dated event is its 'Cyber at the Frontier' livestream on Nov 12 at 10am PT

  Earlier reports (3)

  Oct 19: Tibo Sottiaux, who works on Codex and ChatGPT at OpenAI, posted that he had bought a new domain and was 'very excited for it', without naming it (his X post); X users tied it to opensi.lol · our read of opensi.lol: registered Sep 22, and the Sep 23 web archive copy calls it 'an independent parody' of OpenAI with the domain for sale; the Oct 19 copy is a paid message board that quotes Tibo's post and hides a '6.5' tag next to a countdown to Nov 1, 10 a.m. Pacific · Oct 28: Tibo's next domain post was 'poteto.com has landed', a potato game (his X post); neither he nor OpenAI has tied him to opensi.lol · Nov 1: the countdown ran out with no GPT-6.5; opensi.lol now shows a single video
• GPT-6o · OpenAI
  Unannounced · a DevDay reveal was speculated by one X account (Sep 29); DevDay came and went without one, and no gpt-6o id is among the 469 our sweep has on record · still armed so a gpt-6o id alerts instead of filing under GPT-6 Astra
• GPT-7 · OpenAI
  Not announced — no OpenAI word or date · Kalshi (read Oct 8): 12% before March 2027, 35% before June, 60% before September, 81% before 2028
• Grok 4.9 · xAI
  Unannounced — Musk (Sep 14) put 4.8 and 4.9 ahead of Grok 5 · no id, price or date · Sep 14: Musk on X: 'Grok 4.9 is probably Astra/Fable class', with Grok 4.8 'a noticeable improvement' on 4.7 (his post as quoted by TeslaNorth and Wccftech) · Nov 4: Grok 4.8 itself has not shipped, 44 days after Grok 4.7; no grok-4.8 or grok-4.9 id in OpenRouter's catalog; Polymarket's 'next Grok (4.8+) by Nov 30' market last traded at 67% · Nov 7: Grok 4.8 is still unshipped, 47 days after 4.7, and that market last traded at 64%
• Mythos 5.2 · Anthropic
  Unannounced - the Mythos-tier twin of the 5.2 hunt
• Mythos 5.5 · Anthropic
  Unreleased · Nov 1: Anthropic's own Mythos page (anthropic.com/claude/mythos) reads 'Claude Mythos 5.5' in its page header and wordmark alt text, read directly 15:3x UTC, while its body and pricing still describe Mythos 5.1 · first flagged by @kimmonismus on a tip from @SergioOSINT · no announcement, date or price; no Mythos 5.5 id on OpenRouter or models.dev · removed again by 17:30 UTC Nov 1 (header back to 'Claude Mythos 5.1', re-read directly; @Mr_Salio spotted it at 17:28)
• Mythos 6 · Anthropic
  No date · unannounced; Mythos 5.1 (Sep 1) is still limited to vetted organizations
• Next Claude Sonnet (5.6+ / 6) · Anthropic
  Unannounced — armed on the cadence the moment 5.5 shipped, not on a leak
• Next Gemini 4 (4.1+, Barium) · Google
  No date · Oct 22: @lyraxana says the Carbon checkpoint on LMArena and in the Gemini web app has been updated (unconfirmed) · Oct 24: @LuminaBench's tracker shows gemini-4-carbon and gemini-4-carbon-teamfood among the model references in Google Antigravity build 1.4.3 (one account's screenshot, unconfirmed; no Carbon id in OpenRouter or models.dev)

  Earlier reports (3)

  Oct 28: Google announced Gemini 4 Carbon, which now has its own hunt (gemini-4-carbon); this one keeps 4.1+ and Barium · Gemini 4 Argon shipped Oct 13; Business Insider reported Oct 9 that Google staff were testing newer Gemini 4 checkpoints codenamed Barium and Carbon · Oct 19: The Information's Stephanie Palazzolo says its newsletter has a scoop on Gemini 4 Carbon, one of the releases she expects in the next few weeks (she corrected 'this month' to that) · Oct 21: @Lentils80, quoted by @lyraxana, says Gemini 4 Carbon early access has started for partners and that it is being tested on LMArena (unconfirmed; Google's Gemini API Early Access Program page names no model, and no Carbon id is in Google's docs, OpenRouter or models.dev)
• Next Gemini Pro · Google
  No date · Google stepped back on 3.5 Pro to focus on Flash (Sep 24); Kalshi runs a 'next Gemini Pro' market · armed for a 3.6-3.9 Pro id
• Next GPT Sol (6.2+ / 7) · OpenAI
  Unannounced — armed on the cadence the moment GPT-6.1 Sol shipped, not on a leak · Nov 7: Friday, Nov 6, the soonest date The Information gave, passed without a new Sol or Astra; OpenAI's launch that week was GPT-6 Cyber (Nov 5, to approved Daybreak Red customers only), and its Nov 6 posts were a wet-lab benchmark (BioBasicsBench) and a safeguards framework with the Frontier Model Forum · OpenRouter's newest OpenAI ids are still GPT-6.1 Terra and Terra Pro (Oct 26), and neither OpenRouter nor models.dev lists a gpt-6.2 id · OpenAI has a 'Cyber at the Frontier' livestream on Nov 12 at 10am PT, billed for more cyber announcements · Nov 8: @Mr_Salio and @kimmonismus say GPT-6.1 Sol in ChatGPT now routes some requests to a newer model they take for GPT-6.2 Sol, judging by their own test outputs; unconfirmed, and OpenAI's ChatGPT and model release notes (read Nov 9) say nothing about it · Polymarket's 'Next GPT Sol (6.2+)' market last traded at 65% by Nov 15, 79% by Nov 30 and 98% by Dec 31

  Earlier reports (4)

  Nov 3: no GPT-6.2 Sol id has turned up in any catalog our radar sweeps (this hunt has not fired), and OpenAI has said nothing about one · the GPT-6.2 signals so far name other lines, all unconfirmed: a 'GPT-6.2 Betelgeuse' codename (Oct 29) and a gpt-6.2-astra id in Codex usage screenshots (Nov 1) · OpenAI's Nov 3 blog scored an unnamed 'newer internal model' at 0.875 on its BenchCAD 2.0 test against 0.594 for GPT-6 Astra, without calling it GPT-6.2 · those signals are tracked on the GPT-6.2 page · Nov 4: The Information reported a next-generation Sol could come as soon as Friday, Nov 6, alongside new Astra and cyber models (people with knowledge of them); OpenAI has not confirmed · Nov 4 23:28 UTC: @lyraxana and @AiBattle_ posted screenshots of Google Cloud's quota page listing a gpt-6.2-sol base_model value beside gpt-6.2-astra; screenshots only, unconfirmed, and OpenAI has not announced it · Nov 5 02:23 UTC: @Mr_Salio says both ids have since been removed from the quota page; we could not check that either
• Next Kimi K3 (3.2+) · Moonshot
  No date · any K3 point release past 3.1 (K2 went to K2.5) · none in the catalogs we sweep · Oct 23: K3.1 Pro shipped, so this hunt is now the next K3 and alerts SOTA subscribers, as K3.1 did
• Next Muse Wonder (1.1+ / 2) · Meta
  No date · Muse Wonder 1.0 shipped Oct 12; Muse Spark took point releases about a month apart (1.1 Jul 9, 1.2 Aug 5, 1.3 Sep 2)
• OpenAI Bel · OpenAI
  Codename, not announced · an X leak (@synthwavedd, late Aug) said OpenAI had finished pretraining a model called Bel with over 10T parameters; OpenAI has never named it and no benchmark exists · on Oct 10 @kimmonismus, citing dinner talk, said Bel would overtake Anthropic by year end · on Oct 29 another account (@Josh67Joshua) said Bel is short for GPT-6.2 Betelgeuse, unconfirmed · it may ship under another name, such as GPT-6.2, GPT-6.5 or GPT-7

Also on the radar, each with its own page: GLM-6 (Zhipu AI) · MiniMax M4 (MiniMax) · Qwen 4 Max (Alibaba)

ArtificialWatch watches 101 unreleased models in all. The full list is on the watchlist.

Every AI release market, live · Next by lab: Claude · OpenAI · Gemini · Grok · DeepSeek · Kimi · Qwen · Meta · Mistral · MiniMax · GLM

Earlier releases

Month by month: November 2026 · October 2026 · September 2026 · August 2026 · July 2026

By lab: Google · OpenAI · Anthropic · Meta · xAI · Alibaba · Ant Group · Inception · MiniMax · NVIDIA · Thinking Machines · Zhipu AI

Use this chart

Writing about new AI model releases? Paste this where you want the chart. It redraws when the table updates, and the credit line links back to this page.

<figure><a href="https://artificialwatch.com/releases"><img src="https://artificialwatch.com/api/aw/og/releases" alt="The newest AI models to go live and how many shipped in the last 30 days, by ArtificialWatch" width="600" height="315"></a><figcaption>Chart: <a href="https://artificialwatch.com/releases">ArtificialWatch</a>, CC BY 4.0</figcaption></figure>Copy embed

How they compare: the LLM leaderboard, head-to-heads and API prices. To hear when the next one goes live, watch free.

Follow the log: RSS · subscribe in Apple or Outlook Calendar · add to Google Calendar. One all-day event per release, on the day it went live.

The other end of a model's life: the AI model retirement calendar, every shutdown date the labs have published.

How early this log sees them: live before it's announced, the timed launches our sweep found on a public API before the lab posted (10 of 22 so far).