FULL STORY

GPT-6 Sol/Luna: Launch, Backlash and Rapid Iteration

OpenAI launched GPT-6 Sol and Luna on Sept 23 at half price, drawing mixed reviews and benchmark debates, then quickly followed with GPT-6.1 Sol on Sept 30, offering Astra-level intelligence at roughly a fifth of the price.

2026-09-23 ~ 2026-10-02 · 15 episodes · 184 posts

Episode 1 · OpenAI Launches GPT-6 Sol and Luna at Half Price; Big Hallucination Gains but Not a Clean Sweep (2026-09-23, 92 posts)

On September 23, OpenAI officially launched GPT-6 Sol and GPT-6 Luna, available the same day in ChatGPT Work and Codex (covering Plus/Pro/Business/Enterprise/Edu, with Free and Go users getting Luna on desktop) plus API access (gpt-6-sol). Sam Altman claimed across-the-board gains in intelligence, alignment, coding and computer use at half the per-token price of GPT-5.6. Artificial Analysis benchmarks and community tests, however, show the new models are not a clean sweep: cheaper, more efficient and far less hallucination-prone, but regressing on some software-engineering and knowledge-work benchmarks. Rival releases from Anthropic and Qwen the same day sharpened the coding-API price war, which became a community meme (ns123abc's "Sir" template).

Confirmed

  • Halved pricing: Sol dropped from $4/$20 to $2/$10 per million tokens, Luna from $0.20/$1.20 to $0.10/$0.50. OpenAI claims some coding-eval task costs are up to 93% cheaper than Claude Opus 5 (an official figure relayed by Polymarket, pending third-party verification); Altman called per-task pricing unbeaten in the market. The official positioning is "build with Sol, scale with Luna," both models building on GPT-6 Astra with stronger coding/computer use, higher factuality and clearer, jargon-light answers; Sol targets Claude Opus 5.5 while Luna emphasizes cost efficiency.
  • Coding agent results: In Codex, GPT-6 Sol (max) scores 57 on the Coding Agent Index, up 2 points over GPT-5.6, at roughly half the per-task cost; Luna (max) matches Sol High and Astra Low on DeepSWE at about one-third of Sol's price (relayed by Internet--Sensation).
  • Lower hallucination: Artificial Analysis found AA-Omniscience hallucination rates falling from 92% to 60% for Sol and 93% to 77% for Luna (max effort).
  • Knowledge-work regression: On GDPval-AA v2.1 (adapted from OpenAI's 44-occupation dataset), both Sol and Luna regress at max effort, mainly due to shorter answers omitting key points.
  • DeepSWE decline: User tests (power97992, Angaisb and others) show GPT-6 Sol below GPT-5.6 Sol on DeepSWE; Luna improves slightly over 5.6 Luna at the same reduced price. kimmonismus's benchmark roundup (FrontierCode 48.4% xhigh) suggests Sol matches Fable 5.1 at 85% lower cost; krizzalicious49 posted benchmark screenshot collections versus the 5.6 series.
  • AutomationBench highlight: reachvb relayed that Sol (xhigh) beats Opus 5 on AutomationBench at 9% of the cost.
  • Positive hands-on: Dan Shipper of Every says GPT-6 Sol is now his daily Codex driver—writing near Astra quality, faster and 50% cheaper than 5.6 Sol.

Unconfirmed

  • Rumors relayed by ns123abc claim Sol also trails 5.6 Sol on computer use and research debugging with no cybersecurity gains—unverified; commenters in kristoph's thread mocked it as a Terra rename, also rumor.
  • sytelus claims Sol underperforms the previous generation on HealthBench and some bio/chem safety evals, without sources.
  • weswinder and adonissingh suspect vendors deliberately "buried" the DeepSWE benchmark—speculation with no official response.
  • power97992 reports Sol's xhigh losing to 5.6 Sol max on some coding tasks (small personal sample); FalconsArentReal relayed similar early user comparisons.

Why it matters

A new generation failing to beat its predecessor across the board is unusual, fueling debate about slowing generational progress and selective benchmark disclosure. Meanwhile, halved prices plus sharply reduced hallucination (Reddit users like indiantruely argue the pricing could squeeze many open- and closed-source models) make the models highly attractive for cost-sensitive production. With Anthropic and Qwen releasing the same day, the coding-API market is intensifying. Diverging signals between third-party benchmarks, user tests and daily-use impressions suggest choosing models by testing your own workloads rather than trusting "newer is better."

Timeline

  • September 23: OpenAI announced GPT-6 Sol/Luna, live same day in ChatGPT Work, Codex and API; Altman posted multiple follow-ups on pricing and positioning; Artificial Analysis published its evaluation; Anthropic and Qwen also launched new products.

72 more related posts →

Episode 2 · GPT-6 Sol/Luna launch splits opinion: half-price but dumber? (2026-09-23, 9 posts)

Early reactions to OpenAI's GPT-6 (codenamed Sol/Luna) are sharply divided. Multiple developers and long-time subscribers report that Sol regresses on non-coding tasks, lacks depth, and has lost the personality of its predecessors, while others see roughly equal quality at half the price. OpenAI has not responded, and the core debate is whether price cuts can offset perceived capability shrinkage.

Confirmed

  • Pricing at about half of the previous generation is the main confirmed benefit, noted by users such as Gigaslavx, who felt overall quality may be a slight net gain, with mild regression in Luna.
  • Multiple users report Sol is worse outside coding: tacomaster05 says it "barely tries to think"; rickasaurus calls the regression unacceptable despite lower prices, joking it should have been named Terra.
  • Jarr11 finds GPT-6 Sol clearly below GPT-5.6 Sol, closer to 5.6-Terra, yet priced as premium, and sees a capability gap: 6-Sol insufficient, 6-Astra too compute-hungry.
  • AirportEither2456 tested across reasoning tiers: 6-Sol frequently asks for clarification and underperforms GPT-5.6 Sol on High; Luna also lags its predecessor; recommends Opus 5.5.
  • On tone, WolframRvnwlf and developer migtissera complain GPT-6 Sol feels soulless and machine-like compared with 5.5/5.6, with migtissera saying he finally understands the "Bring Back 4o" crowd.
  • ObviousResort8887 speculates the cost cut aims to recover losses from the previous Astra's heavy token consumption.
  • Per xeophon's digest, impressions aren't one-sided: Luna has surprises, and Claude Opus 5.5 shows a crushing usage comeback.

Unconfirmed

  • Whether Sol is a true net regression versus 5.6 rests on early subjective vibes/evals; users like sstreichsbier are wait-and-see and have switched back to GPT-5.x.
  • The "6-Astra is too compute-hungry" claim comes from individual reports and awaits broader testing.

Why it matters

  • If the "cheaper for worse quality" trade-off holds, it directly affects retention of heavy subscribers and developers.
  • The alleged capability gap (6-Sol insufficient, 6-Astra too costly) leaves the lineup without a balanced option, pushing users to older versions or competitors.
  • Divided word-of-mouth, competitor usage surges, and nostalgia echo previous "Bring Back 4o" dynamics, showing model competition now hinges on capability, price, and personality together.

Episode 3 · LMArena Opens GPT-6 Sol and Luna for Public Testing (2026-09-23, 2 posts)

LMArena has made GPT-6 Sol and Luna available for public testing via Battle Mode and Agent Mode, with leaderboard results and comparisons to GPT-5.6 coming soon.

Episode 4 · Four New Models Reshape the Cost-Performance Frontier, Opus 5.5 Tops Intelligence Index (2026-09-23, 2 posts)

Four new models—MiMo-V2.6-Pro, Claude Opus 5.5, GPT-6 Luna, and GPT-6 Sol—reshaped the intelligence-versus-price frontier this week. Opus 5.5 leads with a score of 58 but costs $5.98 per task, versus just $0.07 for GPT-6 Luna.

Episode 5 · OpenAI and Anthropic Slash Prices Within 90 Minutes, Kicking Off a New Model Price War (2026-09-23, 7 posts)

On September 23, OpenAI and Anthropic released cheaper models within about 90 minutes of each other, kicking off a new API price war. OpenAI cut GPT-6 Sol and Luna API prices by 50% versus GPT-5.6 promotional pricing, while Anthropic launched Claude Opus 5.5 at roughly 40% lower prices. Observers see model competition shifting from intelligence to cost efficiency.

Confirmed

  • GPT-6 Sol is priced at $2/M input and $10/M output, about one-fifth of 5.6-Sol pricing, in line with Terra 5.6 and Sonnet 5 (@TraditionalHome8852, @haider1)
  • GPT-6 Luna is priced at $0.10/M input and $0.50/M output; even the already-cheap 5.6-Luna looks expensive by comparison (@altryne, @haider1)
  • Anthropic released Claude Opus 5.5, claiming it is about 40% cheaper (@krishnan)
  • Simon Willison reviewed Claude Opus 5.5 and GPT-6 Sol/Luna the same day, flagging the new price war (via @altryne)

Interpretations

  • @TansuYegen argues that both companies suddenly competing hard on cost, not just intelligence, is usually a sign of technology moving from impressive to genuinely useful at scale
  • @krishnan observes model pricing shifting toward "per-task accounting"
  • @bindureddy says the simultaneous price cuts show open-source models are putting real pressure on closed vendors
  • @TraditionalHome8852 reasons that since OpenAI did not release GPT-6 Terra, Sol's real target is the upcoming Sonnet 5.5 rather than Opus
  • @haider1 calls the pricing "ridiculously good" and says OpenAI has changed the game on price-performance

Unconfirmed

  • Reddit user @JPBartley asks whether, given both companies were reportedly losing money before the cuts, further price cuts are about market share or a sustainable business—an open question

Why it matters

  • The price war shifts competition from pure intelligence to cost efficiency and could accelerate large-scale AI adoption, though sustainability of the losses remains uncertain

Episode 6 · Tests Show GPT-6 Sol Scales Nearly Linearly With Effort (2026-09-23, 3 posts)

Developer tests show GPT-6 Sol scores rise nearly linearly with effort levels, but Bug Hunt Bench data still places GPT-6 Sol(max) behind Opus 5.5(medium) on real bug-finding.

Episode 7 · GPT-6 Astra Tops Bug Hunt Bench While Sol Version Regresses (2026-09-24, 2 posts)

A real-world benchmark of 105 hidden bugs shows GPT-6 Astra leading with 45 fixes, while GPT-6 Sol regressed sharply, fixing only 29 bugs compared to 43.5 by its predecessor GPT-5.6 Sol.

Episode 8 · Leaks Map GPT-6 Codenames to GPT-5.6 Versions (2026-09-24, 3 posts)

Leaker Pawel Huryn mapped rumored GPT-6 codenames to GPT-5.6 versions, with Sol matching Terra. He also claims GPT-6 Luna (max) scored just 18.3 at n=3, confirming a performance regression that barely beats local Qwen3.8.

Episode 9 · Bug Hunt Bench Ranks Frontier Coding Models on 105 Real Bugs with 200x Cost Gap (2026-09-24, 3 posts)

Pawel Huryn launched Bug Hunt Bench, a blind-graded leaderboard testing frontier coding models on 105 deliberately planted bugs across two repos, revealing a 200x cost gap between models; the author says his code-review skill significantly boosts model scores.

Episode 10 · GPT-6 Sol and Luna Launch With Aggressive API Pricing (2026-09-25, 2 posts)

OpenAI launched GPT-6 Sol and Luna, built on GPT-6 Astra with faster speeds and lower costs. API pricing spans from Astra at $10/$50 per million tokens down to Luna at half of GPT-5.6's promo price, with routing cutting costs further.

Episode 11 · OpenAI's efficient GPT-6 models seen as freeing compute for bigger frontier models (2026-09-25, 2 posts)

Community speculation suggests OpenAI's efficient GPT-6 'Luna/Sol' models are designed to pair with Cerebras' ultra-fast inference while freeing data center capacity for larger frontier models reportedly already in the pipeline.

Episode 12 · GPT-6 Sol Underestimated: 1.6x Robot Task Gains at 47% Lower Cost (2026-09-25, 2 posts)

Leaked benchmarks suggest GPT-6 Sol scores 1.6x GPT-5.6 Sol on robot tasks at 47% lower cost, though below Opus 5.5's Pareto frontier; users argue the model is underrated as cheaper and much faster.

Episode 13 · GPT-6.1 Sol Launches with Astra-Level Intelligence at a Fifth of the Price (2026-09-29, 32 posts)

OpenAI released GPT-6.1 Sol on 09-30, promising Astra-level intelligence at roughly a fifth of the price. Third-party evaluator Artificial Analysis reports the model replaced GPT-6 Sol within 7 days, scoring just 1 point below GPT-6 Astra on Intelligence Index v4.3.2 and nearly tying GPT-6 Astra and Fable 5.1. The consensus: it dramatically extends the intelligence/cost Pareto frontier, though Anthropic's Opus 5.5 still leads on absolute intelligence. Its impact on developer model selection and cost structures makes it worth watching.

Confirmed

  • Artificial Analysis published a GPT-6.1 Sol eval page: five variants (max/xhigh/high/medium/low) scored 42–52, up across all reasoning tiers (low 42 vs 34, medium 48 vs 40), with per-question cost differing up to 5.5x between tiers.
  • Pricing is $2/$10 per million tokens with cache-read discounts raised above the prior 90%; per Reddit user that90sguy's cited data, GPT-6 Sol costs about half of Sonnet 5.5, a third of 5.6 Sol, and a fifth of Opus 5.5.
  • Artificial Analysis also detailed Intelligence Index v4.3.2 with ten benchmarks including AA-Briefcase, GDPval-AA, AutomationBench and Terminal-B; Claude Sonnet 5.5 climbed to second in the same update.

Unconfirmed

  • HLE-Diamond scores shared by haider1 (53.2% total, 67.4% on pure-reasoning subset with only 4k tokens; vs Fable 5.1 at 50.7% and Opus 5.5 at 54.6%) are attributed to GPT-6.1 sol but not officially confirmed.
  • blader (via charliermarsh) reported cfo.ai internal evals where 6.1-sol flipped a roughly 2x deficit versus Opus 5.5 into roughly 3x superiority; personal internal data.
  • Gabriel, engineering lead at Energy (per npew), claimed GPT 6.1 greatly outperformed Opus 5.5 in internal enterprise evals at 40% lower price and 2x speed, and the company adopted it; one-sided account.
  • DeryaTR claimed early access, finding capability on par with GPT-6 Astra but faster and cheaper, with major implications for Codex-based apps; personal early impressions.
  • nicdunz (via pvncher) claimed gpt-6.1 sol max is priced below astra low with intelligence near astra max; unverified.

Why it matters

  • Reviewers (rhiever, randalolson, danielmac8) agree on Artificial Analysis data that Opus 5.5 still leads on intelligence, but GPT-6.1 Sol pushes the intelligence/cost frontier far out; danielmac8 advises tier selection by use case. bindureddy noted it matches Astra at 1/5 the price, marveling that OpenAI "fully bounced back in under a week".
  • dkundel relayed Theo's Terminal Bench 4 self-test: GPT-6.1 Sol beat Opus 5.5 at roughly 1/30 the price.
  • Sentiment isn't uniform: kimmonismus found it slightly slow but efficient, incredibly cheap and solid enough to compete with Sonnet 5.5, yet was disappointed OpenAI offered no Opus-class rival, worried about subscription plan changes, still prefers Opus 5.5, and predicted the next major releases may come from China.
  • BorisMPower noted the Sol release arrived on schedule, in contrast to earlier reports that GPT-6.1 Astra was halted over safety concerns.

12 more related posts →

Episode 14 · GPT-6 Astra Ultrafast Reportedly 8x Faster in Codex (2026-09-30, 2 posts)

Unverified reports say OpenAI's GPT-6 Astra Ultrafast boosts Codex speeds up to 8x, reaching about 300 tok/s, alongside a newly revealed ChatGPT Pro tier.

Episode 15 · GPT-6.1 Sol third-party reviews: near-Astra performance at a fraction of the cost (2026-09-30, 21 posts)

After OpenAI released GPT-6.1 Sol, third-party reviews converged on a common verdict: performance close to the flagship GPT-6 Astra at a fraction of the cost. Artificial Analysis found Sol extends OpenAI's cost-efficiency frontier — $0.72 per Intelligence Index task at max effort — with per-task cost about 30% lower than GPT-6 Sol. Independent developers largely agree, though some benchmark results diverge.

Confirmed

  • Pawel Huryn planted 105 bugs in 2 real code repositories (Bug Hunt Bench, blind, single-prompt): at max/xhigh, GPT-6.1 Sol fixed 44 bugs for $6.56 vs GPT-6 Astra's 45 for $33; all effort levels published (n=3 for max/xhigh, n=2 otherwise, more reruns ongoing); he sees no reason to use Astra given the cost
  • adonissingh's eyebench-v3: Sol ranked second, near Astra at 3.8x cheaper, 1/8 the price of Opus-5.5; Pareto-optimal on cost, but output token efficiency below Astra at every effort level
  • cedricchee found 3D generation near Astra quality with 30% faster speed, +4 points on AA Index over GPT-6 Sol; codestantine reported Sol beating Astra on a self-built low-resource translation benchmark (relayed by nickbaumann)
  • danielmckinn0n on GamowLabs RareBench: Sol nearly matches Astra at 1/4 the price, on the cost Pareto frontier
  • Artificial Analysis published evaluations for all effort levels (Low to Max) and updated Intelligence Index to v4.3 (AutomationBench-AA replaces old items)
  • Reddit user therealjerseytom says it "lives up to the hype": handles challenging work like 6 Astra at much lower token cost, especially when rework is needed

Unconfirmed

  • BridgeBench shows Sol only 3 points above GPT-6 Sol and 82 behind GPT-6 Astra, sharply diverging from OpenAI's "near Astra-level intelligence" claim and other reviews; awaits independent verification
  • Wsz2020 cautions the "88% cheaper than Opus 5.5 max" headline is misleading given Opus's different pricing structure
  • A joke Runescape Bench (relayed by banteg) ranks Sol second at 10% of Astra's price — limited reference value
  • User chogku claims a GPT-6.1 Sol + Opus 5.5 combo approaches Astra at far lower cost (relayed by rudrank); personal opinion

Why it matters

Pawel Huryn stresses that unlike GPT-6 Sol, seen as a nerfed GPT-5.6 Terra, GPT-6.1 Sol is a genuine capability gain. If the low-cost high-performance result holds across more scenarios, it could reshape the price-performance landscape of frontier models and squeeze flagships like Astra out of everyday coding and reasoning workloads.

1 more related posts →