FULL STORY

GPT-6.2: From Rumor Leak to Gradual Rollout

Multiple leaks in late October revealed OpenAI's internal GPT-6.2, with conflicting codenames. Amid debunks, employee teases, and repeated delays, reports emerged that the model had begun gradual routing to some ChatGPT users.

2026-10-30 ~ 2026-11-08 · 8 episodes · 83 posts

Episode 1 · Multiple leaks point to OpenAI internal model GPT-6.2 "Betelgeuse" as "Bel" rumors swirl (2026-10-30, 10 posts)

Around October 30, multiple independent sources simultaneously leaked evidence of OpenAI's next-generation model GPT-6.2, with the internal codename "Betelgeuse." Leakers tied it to the long-rumored internal codename "Bel," spawning speculation that GPT-6.5 "Bel" could be OpenAI's first model to re-commit fully to large-scale pretraining scaling. All information comes from unofficial tracker screenshots and personal leaks; OpenAI has not responded.

Confirmed

  • Screenshots of an internal tracking bot posted by different users (@takerofwhizz, VraserX, ilkamoi's repost, etc.) show a new entry "gpt-6.2-betelgeuse" in OpenAI internal records, marked INTERNAL with regions: default and a do-not-share notice
  • A November 1 Reddit post by BrennusSokol compiling screenshots also shows the tracker capturing anomalous entries such as "gpt-5.2-beta-kremlin / biggame-nutshell," spread by kimmonismu and others
  • X user markk (60k followers) claimed on October 30 that GPT-6.2 is imminent and based on the long-rumored internal codename "Bel"
  • Investor Bindu Reddy leaked the same day that after Google's release of Carbon, OpenAI would also release its internally codenamed "Bel" model, welcoming the news and saying "AGI needs to be liberated"
  • Blogger haider1 traced the multiple versions of the "Bel" rumor: initially said to be a completed 10T+ parameter pretraining run and successor to "Doug," later linked to the Navier-Stokes breakthrough and hundreds of math results; haider1 notes the industry still cannot tell whether Bel is a larger internal pretrained general model for math breakthroughs or a codename for an already-released model
  • VraserX further speculated that the rumored GPT-6.5 "Bel" could be OpenAI's first genuine re-commitment to pretraining scaling, and if even part of the scale rumors hold, its performance could be notable

Unconfirmed

  • Whether "Betelgeuse" is based on the rumored "Bel" model lacks any evidence chain; it rests only on claims by markk and Bindu Reddy, and VraserX's GPT-6.5 version number is personal speculation
  • haider1's suggestion that gpt-6.2 might belong to a new Sol series is pure conjecture
  • Whether the model is near public release and the meaning of the "kremlin / biggame-nutshell" entries cannot be verified; how much of the circulating OpenAI leak information is genuine also remains unclear

Why it matters

  • If true, GPT-6.2 would be another major version iteration for OpenAI, with "Betelgeuse" continuing its astronomical naming tradition
  • Multiple independent tracker screenshots corroborate each other, making this more credible than single-source rumors, though fabricated screenshots and coincidental naming remain risks
  • If the "Bel" rumors hold (10T+ parameter pretraining, return to scaling, math breakthroughs), the significance could exceed an ordinary version iteration, but the evidence chain is currently very weak

Episode 2 · GPT-6.2 Astra surfaces in Codex usage data, OpenAI apparently gray-testing next flagship (2026-11-01, 17 posts)

Starting Nov 1, multiple bloggers nearly simultaneously reported that OpenAI is about to release GPT-6.2, internally codenamed "Astra." Independent sources observed the model identifier in Codex usage panels, API responses, and the CLI, with some accounts seeing nearly 20% of requests routed to the new model and one account as high as 98.5%. The converging evidence strongly suggests pre-release gray/shadow testing, but OpenAI has made no official confirmation and the release timing remains rumor.

Confirmed

  • Multiple independent sources spotted a "gpt-6.2-astra" entry in Codex API usage breakdowns: the model selector UI still shows GPT-6 Astra, but some accounts' usage breakdowns list gpt-6.2-astra separately (m2, m8, m11, m16).机器之心 noted identical screenshots across many accounts, suggesting broad gray-testing rather than a display glitch, though the official model catalog does not yet list it (m16).
  • User idg23's data: gpt-6.2-astra accounted for 18.3% of usage over a week, versus 79.3% for gpt-6-astra (m12); 新智元 reported developers' packet captures showing the new model absorbing 19.8% of real traffic in billing (m15); 量子位 reported some users (mostly $500 Pro plans) seeing it in their model lists, with 20% of requests routed to 6.2, and some users 100% (m17).
  • A Reddit-shared screenshot shows X user Sad AlbertX's usage history with gpt-6.2-astra at 98.5%, claiming use since Nov 1 (m9).
  • kimmonismus and markk both posted API response JSON screenshots with model field "gpt-6.2-astra" (m1, m3).
  • A Reddit user found the Codex CLI accepts codex --model gpt-6.2-astra; requests error out but the name is accepted, suggesting backend readiness (m11); however blogger Angaisb tested it and found it not usable (m7).
  • idg23 also found an unpublished entry in the Japan-region usage panel: a model named gpt-6-astra took 97.6% of usage over 7 days (gpt-6.1-sol 1.1%, gpt-5.6-terra 1%, etc.) before being rolled back, further evidence of shadow testing (m14).
  • 新智元 reported the Codex underlying model was quietly switched to gpt-6.2-astra, with background on the originally planned October 6.1 release (m15); 量子位 documented the buzz around unreleased models from both OpenAI and Anthropic surfacing (m17).

Unconfirmed

  • Screenshot authenticity is unverified by OpenAI; similar Reddit leaks are also unconfirmed rumors (m8, m9, m10).
  • imjustnewatai also claimed GPT-6.2 Astra launches next week, but the poster admits prior trolling, so credibility is doubtful (m6).
  • markk claimed GPT-6.2 could be announced as early as Monday, possibly as the finale of OpenAI's "28 Days of Improvements"—speculation (m3); koltregaskes also judged a next-week release likely (m13); 机器之心 mentioned the "skipping 6.1 or direct release" theory (m16).
  • haider1 offered two guesses on "Astra 6.2": a larger, stronger model possibly the shelved "Bel," or a regular iteration of the astra/sol/terra series (m4). He also argued the dense 28-day run of Codex upgrades (larger context windows, stronger subagents, more generous usage) may be part of a launch strategy targeting Anthropic's 5.5 lineup (m5)—blogger opinion.
  • kimmonismus suggested the new model may be a flagship response to competitor Fable 5.5—blogger opinion (m1).

Why it matters

  • Real model identifiers in usage panels, CLI acceptance of the model name, and the rolled-back 97.6% gpt-6-astra entry mutually corroborate that the product is in pre-release testing rather than mere rumor—the strongest evidence in this cluster.
  • If GPT-6.2 is real, it will reshape OpenAI's flagship lineup and competitor timing; 新智元 juxtaposed its leak with Claude Mythos 5.5's same-weekend leak, arguing the next flagship race has started early (m15), warranting close watch on official follow-ups.

Episode 3 · Rumor Debunked: Viral GPT-6.2 Astra Demos Were Just Regular GPT-6 (2026-11-03, 3 posts)

Community claims that OpenAI was secretly testing GPT-6.2 Astra in Codex were debunked: the viral Xbox controller SVG was actually drawn by regular GPT-6 Astra, and some viral demos show signs of fakery.

Episode 4 · OpenAI employees tease a bigger GPT model with cryptic memes (2026-11-04, 2 posts)

Multiple OpenAI employees, including researcher Thibault Sottiaux, have posted cryptic "Le GPT Fat" memes, widely interpreted as hints that a larger GPT model is coming soon.

Episode 5 · Mystery rc-ada model IDs flash in OpenAI API as GPT-6.2 rumors swirl (2026-11-04, 20 posts)

Starting Nov 4, users and third-party model trackers repeatedly caught two unrevealed OpenAI API model IDs, rc-ada-or-vm and rc-ada-sh-vm, which were quickly removed. Combined with a report by The Information's Stephanie Palazzolo, an X UI screenshot showing gpt-6.2-astra and gpt-6.2-sol entries, and GPT-6.2 Astra appearing in Codex usage stats, the community widely expects GPT-6.2 to ship in Astra, Sol and Cyber variants as soon as Friday, Nov 6 — but OpenAI has said nothing official, so all of this remains rumor.

Confirmed

  • Models Tracker bot and third-party tracker zAI's "New OpenAI Models" panel both captured the two IDs, with mutually corroborating screenshots captioned "gpt-6.2 astra loading."
  • Per Reddit user @SoggyPerspective776, the IDs were removed shortly after appearing.
  • Per The Information (broken by @stephpalazzolo; relayed by @kimmonismus, @testingcatalog, @ChrisGPT, @haider1), OpenAI may release new models as early as Friday, Nov 6, spanning Astra, Sol and Cyber (successor to its cybersecurity-focused model); @haider1 suggested names like sol 6.2 and astra 6.x. The report itself cautions that timelines may change — it is journalism, not an official announcement.
  • X user @VraserX posted a screenshot showing gpt-6.2-astra and gpt-6.2-sol in a basemodel field, speculating a same-day dual release.
  • Per @haider1, GPT-6.2 Astra appeared days earlier in Codex usage stats and API responses; the same post claims "Mythos 5.5" was found in Anthropic's website HTML, describing a mutual hide-and-seek between the two labs.
  • 新智元 rounded up the leak signals, including a claimed 87.5% benchmark beating full GPT-6 (source unclear, unofficial).
  • OpenAI has made no statement on the IDs or any release plan.

Unconfirmed

  • The purpose, capability and positioning of the two models are unknown; @WorldofAI's weekly also cited a "HUGE LEAK" screenshot of a mysterious GPT-ADA candidate, identity unverified.
  • @WorldofAI claims GPT-6.2 Astra is strong at coding and SVG generation, comparable to Fable 5.5 — unverifiable.
  • @kimmonismus said the GPT-6.2 rumors are real and could land "as soon as tomorrow"; some speculate a same-day drop (@udmrzn's reposted screenshot showing "astra loading"), while @heypearlai joked about hoping it's not reheated leftovers.
  • @imjustnewatai claims an internally trained model arrives Friday, priced "very expensive" yet far smarter than GPT 6 Astra and GPT 6.1 Sol, with even the low tier beating GPT 6 Astra's top config — no sourcing.
  • @kimmonismus relayed a Sam Altman interview remark about releasing a model that internally solved Navier-Stokes; unclear and unverifiable.
  • Community members note "ada" echoes OpenAI's retired legacy naming; the 87.5% benchmark is also unverified.

Why it matters

  • The "rc-" prefix typically means release candidate; if that reading holds, the new models are in the final pre-release stage — a strong launch signal.
  • The Information's "three variants at once" report aligns with @VraserX's dual-entry screenshot; if true, OpenAI would for the first time split its lineup by use case (Astra, Sol, Cyber) with a possible multi-model same-day launch. Combined with the @imjustnewatai pricing claims, the API ID leaks, Codex stats leaks and rumored benchmarks, these signals form a useful trail for tracking OpenAI's roadmap — but treat everything as rumor until officially confirmed.

Episode 6 · GPT-6.2 Astra/Sol Surface on Google Cloud, but Launch Rumors Fall Flat (2026-11-05, 18 posts)

Starting November 5, multiple accounts reported traces of OpenAI's next-generation models on Google Cloud: quota page filters showed basemodel:gpt-6.2-astra and basemodel:gpt-6.2-sol, suggesting GPT-6.2 may ship in two variants. The leak originated with @testingcatalog and was amplified by @lyraxana, @scaling01, @Angaisb and @koltregaskes, who noted the entries were quickly removed — a pattern often read as signaling an imminent launch. On November 6, @kimmonismus, @danielmac8 and @1411337 relayed an The Information report: no GPT-6.2 on Thursday, earliest possible launch Friday, potentially including next-gen Astra, Sol and a cybersecurity-focused model (codenamed Cyber per @新智元, which claimed it 'eliminated 948 kernel vulnerabilities'), with the caveat that timing could change. Separately, Codex CLI screenshots shared by @ChrisGPT and @VraserX (original tweet by Layton Gott) showed gpt-6.2-astra xhigh, an unreleased model; OpenAI's Thibault Sottiaux responded with a cryptic 'basement full of treasures' remark, read as hinting at internal testing. By November 7, @haider1 noted Astra 6.2 had not launched Friday, consistent with the community belief that OpenAI never ships major models on Fridays; he argued this was a schedule slip, not a sign of a weak model. @ns123abc documented @kimmonismus's streak of missed predictions, and Reddit user @Bitter-College8786 lamented that only GPT-6.1-Terra shipped this week, with a rumored Qwen 4 launch also failing to materialize.

Confirmed

  • Google Cloud Console quota pages showed basemodel:gpt-6.2-astra and basemodel:gpt-6.2-sol, corroborated by screenshots from multiple accounts
  • Per @koltregaskes, the entries were removed after appearing, which he read as a sign of imminent launch
  • Per The Information (relayed by @kimmonismus, @danielmac8, @1411337): no Thursday launch, earliest Friday, possibly including Astra, Sol and a cybersecurity model, with timing subject to change
  • Per @新智元, the cybersecurity model is codenamed Cyber and claimed to have 'eliminated 948 kernel vulnerabilities'
  • Per @ChrisGPT and @VraserX, Codex CLI screenshots showed unreleased gpt-6.2-astra xhigh; Thibault Sottiaux gave a cryptic response
  • Per @haider1, the rumored Friday launch did not happen; he attributed it to schedule slip rather than model quality
  • Per @Bitter-College8786, only GPT-6.1-Terra shipped this week; the rumored Qwen 4 launch also failed to materialize
  • Per @ns123abc, @kimmonismus repeatedly missed his own launch predictions while claiming his info is 'always correct'

Unconfirmed

  • Whether GPT-6.2 truly has Astra and Sol variants, their positioning differences, and what the xhigh suffix means
  • Whether the Cyber model exists and its claimed '948 kernel vulnerabilities' figure
  • Actual launch timing: Friday passed with no announcement, the week ended with GPT-6.1-Terra, and all evidence rests on screenshots and media relays; neither OpenAI nor Google has officially confirmed anything

Why it matters

  • If genuine, this would be the first appearance of concrete codenames for OpenAI's next flagship, exposed unusually via a third-party cloud (Google Cloud) and internal tooling (Codex CLI), hinting at deepening OpenAI–Google Cloud distribution cooperation
  • A dual-variant Astra/Sol structure could signal distinct product lines; a Cyber model would extend OpenAI's footprint in cybersecurity
  • The leaked entries, cryptic employee replies, the failed Friday launch and leakers' broken track records show such timing rumors are unreliable; official announcements remain the only dependable signal

Timeline

  • Nov 5: gpt-6.2-astra and gpt-6.2-sol appear on Google Cloud Console then get removed; screenshots spread
  • Nov 6: The Information reports earliest Friday launch possibly including Cyber; Codex CLI screenshot reveals gpt-6.2-astra xhigh; OpenAI employee responds cryptically
  • Nov 7: @haider1 notes no Friday launch; @Bitter-College8786 says only GPT-6.1-Terra shipped; @ns123abc documents @kimmonismus's missed predictions

Episode 7 · GPT-6.2 Delays Spark Community Backlash as Safety Rumors Swirl (2026-11-07, 8 posts)

On November 7–8, as rumors of repeated delays to OpenAI's rumored GPT-6.2 circulated, the AI community erupted in complaints and memes. Markk, a well-known user with over 60,000 followers, tweeted that he was exhausted by GPT-6.2's constant slipping and openly questioned what is going on at OpenAI. Developer heypearlai was even sharper, complaining that OpenAI had been hinting for weeks that GPT-6.2 was about to launch, yet 'every day it's promised for tomorrow and every day nothing happens,' calling it pure rage bait designed to farm engagement.

Confirmed

  • Community frustration and mockery did indeed erupt in concentrated fashion: markk publicly complained about the delays, and heypearlai criticized officials for days of teasing without shipping.
  • None of the rumored new versions have been released: besides GPT-6.2, other rumored releases such as Fable 5.5 (a rumored Anthropic model codename), Grok 4.8, and the Carbon and Hydrogen variants of Gemini 4 are similarly 'missing.'
  • Drawing on experience with multiple trusted tester programs, koltregaskes noted that labs postponing releases after finding last-minute issues is quite common, and also shows labs react quickly to testing feedback.

Unconfirmed

  • Whether GPT-6.2 and the Astra codename actually exist, and any concrete release timing, remain unverified; OpenAI has made no response.
  • The real reason for the delay is unknown; koltregaskes's 'last-minute problems found' theory is merely personal speculation.

Why it matters

  • This wave of complaints reflects the community's anxiety over the vacuum in heavyweight model releases: markk joked that 'all the new version numbers have gone missing' and quipped about a 'new AI Winter,' while a meme shared by m5 claimed GPT-6.2 is waiting on Fable 5.5, and Gemini 4 is waiting on both of them — mocking an entire industry waiting for competitors to move first.
  • If heypearlai's accusation of days of official teasing is accurate, it raises questions about vendors' release cadence and traffic-driven engagement strategies, making any official response or actual release worth watching.

Episode 8 · GPT-6.2 Sol reportedly rolling out to some ChatGPT users, launch expected next week (2026-11-07, 5 posts)

On November 7–8, multiple X users reported that OpenAI's unreleased model GPT-6.2 (codenames Sol/Astra) has begun routing to a subset of ChatGPT users via gradual rollout, with an official launch expected as soon as next week.

Confirmed

  • User TokenGremlin said they had been routed to GPT-6.2 Sol and showed off a Three.js black hole visualization generated from a single prompt, with all the code in one HTML file. They felt the model is "insanely strong" at design and 3D work, already beating Opus 5.5 in single-shot tests.
  • @kimmonismus relayed the same observation, saying GPT-6.2 Sol is being gradually opened up to some users.
  • @imjustnewatai said 6.2 Astra/sol is being routed out to a large number of test users and performs noticeably better.
  • @m5 noted that The Information had previously reported GPT 6.2 was slated for release last Friday but didn't ship, and that GPT 6.2 screenshots have already surfaced in Google Docs.

Unconfirmed

  • @danielmac8 predicted GPT-6.2 Sol would launch on November 9 (Monday) with capabilities surpassing Opus 5.5, while GPT-6.2 Astra would arrive on November 12, on par with Fable 5.5, with both priced roughly 5 lower than Claude (original text truncated). This prediction has no official backing.
  • @imjustnewatai asserted it will "definitely launch next week," but both the model naming and release timing await official confirmation from OpenAI; @m3 and @m5 themselves also flagged the info as rumor.

Why it matters

  • If real-world tests bear out, GPT-6.2 Sol could be significantly ahead of competitors in single-shot 3D/design generation, directly challenging the positioning of rivals like Opus 5.5; if the rumored pricing holds, it would also intensify price competition among models. As of posting, OpenAI has not issued an official response.