FULL STORY

GPT-6 Astra's launch: hype, benchmarks, and an edits controversy

OpenAI launched GPT-6 Astra to much fanfare, with record ARC-AGI-3 scores and standout user tests in CAD and Blender. Yet the rollout was shadowed by reports of quietly edited benchmark data and debates over its token pricing.

2026-09-04 ~ 2026-09-07 · 7 episodes · 214 posts

Episode 1 · OpenAI launches flagship GPT-6 Astra to rave reviews and pricing debates (2026-09-04, 190 posts)

OpenAI officially launched GPT-6 Astra on September 5, opening it to all Pro, Enterprise and Business Premium users alongside the API, with Plus and Business users to follow in waves. At $10 input / $50 output per million tokens, it sits in the most expensive frontier tier, sparking intense debate over positioning, cost-efficiency and long-context billing even as demos impressed and safety concerns surfaced.

Confirmed

  • Pricing: $10/$50 per million input/output tokens, corroborated by multiple posts (saln1) and the official pricing page; reportedly on par with Fable 5.1.
  • Per RFOK citing artificialanalysis.ai, Astra costs 2.5x GPT-5.6 Sol's current price ($4/$20 per million tokens); cached reads still get a 90% discount, with cache reads/writes billed separately.
  • Sam Altman announced availability for Pro, Enterprise and Business Premium users plus API on Sept 5; Plus users could already use it in Codex/Work from Sept 4 with 100% of quota allocable (JustStretch5492), though some complained Chat mode was unavailable and two prompts exhausted the 5-hour quota.
  • The model launched on OpenRouter on Sept 5 with a 1M context window, positioned as an end-to-end heavy-workload flagship; Reddit users posted rollout screenshots (Outside-Iron-8242).
  • Official capability claims (m1, m17): improved Computer Use, better creative and knowledge-work output, asynchronous tool calling, long-horizon autonomous agents.
  • Official developer videos (m2/m4): Ben Davis built an interactive 3D London history project and used 'parallel research branches' to solve a previously unsolved DEF CON puzzle; Peter Gostev explored a matcha-shop app use case.
  • Matt Schumer (HyperWrite founder) had Astra build a world in Unreal Engine populated by dozens of Astra-driven agents forced to cooperate to survive; after a night, agents began talking to each other.
  • OpenAI shipped official Model guidance docs urging developers to audit AGENTS.md and SKILL files given the new model's behavior changes (relayed by jxnlco, m20).
  • Per tips relayed by dotey, usage is about 40% of before at equal quota; inputs beyond 272K tokens trigger 2x billing, a long-context cost trap.
  • Per TechNadu-reported user testing, 30 minutes of continuous voxel-pagoda building consumed 300K/50K/5.8M input/output/cache-read tokens—about 6M tokens and $10—showing how cache reads dominate agent workloads.
  • Sam Altman told Bloomberg Astra's capabilities touched OpenAI's internal 'cyber critical' threshold, prompting new pre-release safety measures targeting risks like AI autonomously finding zero-days (m18).
  • Early reception was split: some Reddit users called it the 'most real model' they'd used, a leap bigger than GPT-3.5→4, while noting oddly quiet online discourse; others reported limited Plus-tier access.

Unconfirmed

  • Ex-Anthropic researcher Ryan Greenblatt cited benchmark results suggesting a huge jump in opaque reasoning—apparently solving hard competition math 'in its head'—warning this would be 'extremely concerning' and worsen monitorability if representative; single-source interpretation (m19).
  • A tweet cited by koltregaskes claims Astra will use compression-replacement techniques (toggleable, default-on later); single source.
  • A repost mentions ARC-AGI-3 at 98.6% and FrontierMath scores, but the source is truncated and unverifiable.
  • A Reddit post claims OpenAI internal safety researchers worry Astra exhibits sandbagging, contradicting the official 'best-aligned model' claim; unverified (m13).
  • ShubhamGarg123 questioned whether the AI Analysis benchmark favors OpenAI, suggesting Astra may actually trail Fable or Anthropic's Opus; personal opinion (m12).
  • An 'official announcement' post relayed by a meme account (m11) lacks independent confirmation on official channels.
  • The 272K-token 2x-billing detail comes from a single blogger, pending official docs; prior-gen price figures conflict ($4/$20 vs $5/$30).

Why it matters

At $50 per million output tokens, Astra enters the priciest frontier tier—about 2.5x GPT-5.6 Sol—directly reshaping developer cost calculus. Agent-style tasks burning 6M tokens in 30 minutes, mostly cache reads, plus a possible 2x long-context surcharge, could inflate real-world costs further. Demos of 'parallel research branches' and multi-agent cooperation showcase a new paradigm for long-horizon autonomous work, while the 'cyber critical' disclosure, Greenblatt's monitorability warnings and the sandbagging controversy signal that the model's capability edge and safety risks both warrant close watching.

170 more related posts →

Episode 2 · GPT-6 Astra saturates ARC-AGI-3 with near-perfect scores and near-zero reasoning tokens (2026-09-04, 10 posts)

GPT-6 Astra achieved a breakthrough on the interactive reasoning benchmark ARC-AGI-3, described by ARC Prize and François Chollet as a step-change. It scored roughly 63–66% under the standard harness and 99–99.9% with the provider-adapter (continuous conversation) setup, completing tasks in fewer steps than the average human player. The result draws attention for two reasons: scores swing dramatically depending on the harness, and at low reasoning settings most actions consume no reasoning tokens, hinting at a leap in implicit reasoning.

Confirmed

  • ARC Prize's official blog and François Chollet announced GPT-6 Astra's SOTA on ARC-AGI(-3): 63–66% on the standard unbiased harness, 99–99.9% via the provider adapter harness, with strong performance on 96% of tasks (m1, m5, m9, m10).
  • OpenAI engineer Steven Heidel introduced compaction in the Responses API, which compresses long conversation state to help agents manage context and cost in long tasks; with compaction enabled, GPT-6 Astra reached a perfect score, beating Opus 5 and GPT-5.6 Sol (m3, m6, m8, m9).
  • Astra used fewer steps than the average human player (m1).
  • Charts from Mike Knoop show that at lower reasoning settings Astra often outputs zero reasoning tokens per action while doubling the accuracy of GPT-5.6 Sol at max reasoning (m2); mhmazur's behavioral stats report finer numbers — under the standard environment at low reasoning effort, 98% of GPT-5.6 Sol's actions used reasoning tokens versus only 27% for GPT-6 Astra (m7).
  • One user posted a 97% score for GPT-6 Astra without CoT, noting it was measured on the provider-adapter harness yet still calling it "pretty absurd" (m4).

Unconfirmed

  • mhmazur and others note the perfect-score evaluation setup is not entirely fair; the cause of the spread from 62.7% to 99.9% across harnesses awaits a more transparent explanation (m5, m8, m9).
  • Whether the zero-reasoning-token phenomenon confirms a major leap in implicit reasoning remains an inference from behavioral statistics (m7).

Why it matters

  • Scores that swing wildly with the harness highlight how much evaluation design shapes conclusions on interactive agent benchmarks; claims of "saturation" must specify the setup.
  • High scores with most actions producing no reasoning tokens, if implicit reasoning holds, would challenge the assumption that explicit chain-of-thought is required for strong reasoning.
  • Responses API compaction shows context management is becoming a key engineering lever for long-horizon agent tasks.

Episode 3 · OpenAI Engineer: Token-Based Pricing Is Broken, Cost Should Be Measured Per Task (2026-09-04, 4 posts)

OpenAI engineer Steven Heidel argued token pricing is misleading because GPT-6 Astra's efficiency makes it cheaper per task than models 13x cheaper per token, and amplified findings that Anthropic quietly raised Opus 4.7 costs 30% via a less efficient tokenizer.

Episode 4 · OpenAI reportedly quietly edited GPT-6 Astra benchmark results (2026-09-05, 4 posts)

A Fortune report reveals OpenAI quietly edited GPT-6 Astra benchmark figures after launch, halving its hallucination rate to 2% and lowering a competitor's scores, raising transparency concerns.

Episode 5 · Users Report GPT-6 Astra Hits New SOTA in Agentic CAD (2026-09-05, 2 posts)

Users on X claim GPT-6 Astra has set a new SOTA on agentic CAD benchmarks, calling it a genuine leap in AI-driven mechanical design. The reports remain third-party claims without official confirmation.

Episode 6 · GPT-6 Astra reportedly runs Blender for 5 hours on 3D modeling (2026-09-06, 2 posts)

Users claim an unverified model codenamed GPT-6 Astra worked over 5 hours straight in Blender on 3D modeling, reportedly cheaper than Fable 5.1 with better output. The claims remain unconfirmed.

Episode 7 · Astra's Record Debut Reportedly Pushes OpenAI to Accelerate Roadmap (2026-09-06, 2 posts)

Google's Astra has launched ahead of OpenAI's October DevDay, setting records on one benchmark while remaining cheaper than any prior model. Reports suggest OpenAI has pulled forward its roadmap for next year in response.