FULL STORY

GPT-6.2 Launch Day: From Leak to Backlash

GPT-6.2 Astra and Sol leaked in Codex hours before OpenAI's official launch on Nov 18. Benchmark gains were overshadowed by coding stagnation, demo controversy, and a polarized, possibly coordinated backlash.

2026-11-18 ~ 2026-11-18 · 12 episodes · 217 posts

Episode 1 · GPT-6.2 Sol and Astra spotted unreleased in OpenAI Codex model picker (2026-11-18, 5 posts)

On November 18, multiple independent users discovered unreleased GPT-6.2-series models in OpenAI's Codex model selector, corroborating each other and suggesting OpenAI is running a staggered rollout, with an official announcement possibly imminent.

Confirmed

  • @markk reported that GPT-6.2 Sol and GPT-6.2 Astra appeared in the Codex model selector; a screenshot shows GPT-6.2 Sol checked as the default "coding and everyday" option.
  • @AliveStudios's screenshot shows the dropdown listing, besides GPT-6.2 Sol, a row of unreleased models including GPT-6.2 Astra, GPT-6.1 Sol (checked at the time), GPT-6 Astra, and GPT-6.1 Terra.
  • @JadedReception5216 subsequently said they also just saw "6.2 Sol" among the selectable models in Codex, corroborating earlier reports and suggesting the test entry may be opening to more users.
  • @zyyyyl9144 also posted a screenshot saying GPT-6.2 Astra and Sol are coming.

Not yet confirmed

  • Some forwarded messages claim GPT-6.2 Astra and Sol are rolling out, available to some users, with an official announcement imminent—but neither this nor any related information has been confirmed by OpenAI, and there is no official release post yet.

Why it matters

  • Multiple unrelated users seeing the same model names at different times suggests this may not be a one-off UI bug, but a deliberate staged rollout or pre-launch preparation by OpenAI.
  • The simultaneous appearance of multiple codenames—Sol, Astra, Luna, Terra—in the model list hints that the GPT-6.x series may adopt a multi-model parallel product strategy, worth watching for confirmation in future official announcements.

Episode 2 · OpenAI launches GPT-6.2 Astra and Sol: flagship power at a fraction of the cost (2026-11-18, 93 posts)

On November 18, OpenAI officially released two GPT-6.2-series models: flagship GPT-6.2 Astra and efficient all-rounder GPT-6.2 Sol, available immediately on ChatGPT Work, Codex, and API for Plus, Pro, Enterprise, and Edu users. Responding to community anticipation with the quip "some asked politely, some just demanded," President Greg Brockman reshared the announcement, saying both models excel at professional work and scientific research at significantly lower cost per task.

Confirmed

  • Astra's positioning: a new flagship for complex professional work—cleaner code, stronger frontend and 3D modeling, more efficient tool calling, and slide/spreadsheet/document outputs closer to industry templates with less of an "AI flavor."
  • Sol's positioning: an efficient everyday workhorse that handles multi-step agent workflows; OpenAI says many employees already use it as their daily driver (employee Thys Sottiaux echoed this in a repost).
  • Pricing: Astra's API input/output is about 20% cheaper than GPT-6.1 Astra; per @kimmonismus relaying official info, agent costs are only about 5% of before (see official pages for specifics).
  • Efficiency: both models achieve better results on hard tasks with fewer output tokens.
  • Benchmarks: per @danielmac8 citing the launch blog, Astra roughly matches Claude Fable 5.2 across multiple benchmarks including Agents' Last Exam at a much lower price; Sol has scores on the Artificial Intelligence Index (exact figures per the blog).
  • Ultrafast: launched alongside; the Pro 200 plan unlocks the ultrafast lane (details on Pro 50 in @pvncher's post).
  • Third parties: both models are live on LMArena with Battle Mode blind tests and Agent Mode trials.

Not yet confirmed

  • The official post included no full benchmark table or per-item pricing; exact figures await the launch blog.

Why it matters

  • The two-tier model lineup (power + savings) plus the 20% price cut directly targets the value proposition of Anthropic's Claude Fable 5.2; @danielmac8 cited mattparlmer's view that with Gemini back at the frontier plus this release, the earlier "OpenAI monopoly/behind" narrative is being questioned.
  • Polymarket simultaneously opened a prediction market on GPT-7's release timing, showing the market's continued attention to OpenAI's cadence.

73 more related posts →

Episode 3 · GPT-6.2 Astra and Sol launch: gains in intelligence and cost, but coding stalls and reviews sour (2026-11-18, 89 posts)

On November 18, OpenAI released GPT-6.2 Astra and GPT-6.2 Sol, followed by results from Artificial Analysis, Epoch AI, Vals AI, Mercor, and multiple community reviewers. The overall picture is consistent: significant gains in cost-effectiveness and general intelligence, but coding agent capability stagnant or even regressing.

Confirmed

  • Pricing: Per Artificial Analysis, GPT-6.2 Astra is priced at $8/$40, 20% cheaper than GPT-6 Astra, with cached input dropping from $1 to $0.
  • Artificial Analysis Intelligence Index v5.0.1: Claude Fable 5.2 (max) leads with 65, GPT-6.2 Astra (max) second at 62, and Sol at 57. The cost-score curve shows Astra reaching a high score at around $5 or less; user zephyrz9 believes OpenAI has moved far ahead of the previous generation in "performance per dollar."
  • Coding Agent Index (Codex harness, max effort): GPT-6.2 Astra ties GPT-6 Astra (both 53), while GPT-6.2 Sol dropped 3 points versus GPT-6.1 Sol. Yuchen Jin and multiple Reddit users questioned whether the new models improved coding at all.
  • Epoch ECI: GPT-6.2 Astra scored 174, level with Fable 5.2, and above what the prior reasoning-model trend line predicted.
  • Vals AI: GPT-6.2 Astra scored 29.01% on the Vals RSI-Index — the highest of any OpenAI model it has tested, though still below Opus 5; on the Vals Index leaderboard, Astra and Sol rank 7th (67.66%) and 9th (66.06%) respectively, only 1.6 points apart, yet Sol's per-task cost is roughly one-third. Bindu Reddy's evaluation showed Astra's Agentic score fell from 75.6 to 70.8 generation-over-generation, while it led in browser use with 93.0.
  • Mercor OpenWorldEngineer benchmark: GPT-6.2 Astra topped the chart with a 67.2% average, beating Fable 5.2 (54.6%) and GPT-6 Astra.

Unconfirmed

  • ChrisGPT claimed GPT-6.2 Astra scored about 170 on the OpenAI ECI, making it the new leader, and said frontier model trend lines gain roughly 15.6 points per year; Epoch's independent test gave 174 — the two figures need reconciliation.
  • Detailed benchmark data from some Reddit users (Electronic-Weekend66 and others) is flagged as unverified leaks; conclusions like "Astra uniquely owns the most attractive value-for-money segment" align in direction with public AA/Epoch data, but details lack official confirmation.

Why it matters

  • Tansu Yegen pointed out that the bigger story of this launch may be Sol: achieving GPT-6 Astra's intelligence index score at roughly 15% of the per-task cost, democratizing last generation's flagship performance.
  • ValsAI's hands-on testing echoes OpenAI's September claim of an "automated research intern," concluding GPT-6.2 Astra can already complete tasks under human guidance that would take a skilled researcher days.
  • The stalled coding agent capability and declining Agentic scores (Bindu Reddy assessed it as "trading performance for cost") have raised questions about OpenAI's trade-off priorities; Arena announced that an Agent Arena rating based on real usage sessions is coming soon, offering a window for follow-up observation.

69 more related posts →

Episode 4 · OpenAI Accused of Showcasing Misleading GPT-6.2 Demos (2026-11-18, 2 posts)

Teknium accused OpenAI of using an older GPT-6.0 Blender demo when launching GPT-6.2, while the GPT-6 Astra release showed stronger results, fueling speculation that OpenAI deliberately degraded the model's performance.

Episode 5 · OpenAI Releases GPT-6.2 Astra with Major Cybersecurity Gains (2026-11-18, 4 posts)

OpenAI's new GPT-6.2 Astra and Sol models lead cybersecurity benchmarks, topping Vercel's DeepsecBench with a 20-point gain over the prior generation and finding vulnerabilities at roughly $12 each on DeepBinary, prompting OpenAI to tighten safety guardrails for high-risk cyber tasks.

Episode 6 · Theo's Open Letter Slams GPT-6.2 Astra Developer Experience (2026-11-18, 3 posts)

Developer Theo published an open letter calling GPT-6.2 Astra's developer experience too poor to recommend, amid unverified claims OpenAI engineers themselves prefer 6.2 Sol Ultrafast and complaints about overly strict instructions.

Episode 7 · GPT-6.2 hit by wave of negative posts suspected coordinated (2026-11-18, 4 posts)

After GPT-6.2's release, a wave of seemingly coordinated negative posts drew suspicion of paid PR, with some tying it to Anthropic's IPO, while developers testing the Astra and Sol variants reported solid performance comparable to version 5.6.

Episode 8 · GPT-6.2 Models Top VoxelBench 2 Error Rates (2026-11-18, 2 posts)

Third-party evaluator VoxelBench reports that OpenAI's GPT-6.2 Astra and GPT-6.2 Sol record the highest error rates ever on VoxelBench 2, with Astra committing 19 errors and performing poorly in voxel generation.

Episode 9 · Mixed Reactions to GPT-6.2 as Hands-On Tests Push Back on Criticism (2026-11-18, 2 posts)

Amid widespread criticism on X, hands-on testers pushed back: Alex Finn found GPT-6.2 a highly capable computer-use model, and CtrlAltDwayne reported noticeably faster speed and more agile context compaction.

Episode 10 · Researchers: GPT-6.2 Astra Looks Safer but Is Harder to Monitor (2026-11-18, 8 posts)

On November 18, researcher tomekkorbak posted a Twitter thread breaking down the GPT-6.2 Astra system card, concluding that the model "looks safer but is actually harder to monitor."

Confirmed

  • GPT-6.2 Astra shows capability jumps without CoT: it can complete harder tasks without producing a monitorable chain of thought, including no-CoT performance on math competition tests. The author's concern is less any single data point than the pace of improvement in CoT-free capabilities—a break from the old trend, but a continuation of the new one started by GPT-6 Astra.
  • CoT controllability is slightly lower than the previous-generation 6 Astra; monitorability depends heavily on CoT length, and at equal CoT lengths 6.2 Astra is close to 6 Astra. However, OpenAI stresses that 6.2 is more token-efficient, meaning shorter CoT for the same performance—so at their respective median CoT lengths, 6.2's controllability may actually be worse.
  • Action-only monitoring (where the monitor detects misconduct from tool calls alone, without seeing the chain of thought) keeps degrading across multiple evaluations; as CoT gets shorter and models rely on it less, tool calls are becoming nearly the only signal monitors have, and 6.2 Astra performs worse against older action-only monitoring.

Why it matters

  • Taken together, these trends mean overall capability and safety metrics may improve, while the monitoring signals available to safety researchers are shrinking in exactly the scenarios they care most about—monitorable reasoning is fading further.
  • Former OpenAI safety researcher balesni reshared the thread, commenting that the model may be better overall, yet in the scenarios that really matter GPT-6.2 Astra is harder to monitor than GPT-6 Astra, and voicing concern that the safety discussion in system cards is disappearing.

Episode 11 · GPT-6.2 Astra Scores 174 on Epoch Capabilities Index (2026-11-18, 3 posts)

Epoch AI measured GPT-6.2 Astra at 174 on its Capabilities Index, topping a 241-result leaderboard and exceeding the reasoning-era trend line, with commentators calling it a breakthrough.

Episode 12 · Developer Tests GPT-6.2 Astra: Better, But Still on the Wrong Track (2026-11-18, 2 posts)

Developer theo's hands-on retest of OpenAI's GPT-6.2 Astra found some improvement over his earlier tests, but concluded the model remains fundamentally on the wrong track in scenarios he tracks, such as diagnosing VPS issues.