OpenAI GPT-6 Astra Enters the ‘AGI Era’: Computer-Use Paradigm Leap + 100,000-GPU Training + First Critical-Tier Cybersecurity Model + ‘Executing Tasks’ Beyond ‘Answering Questions’

TL;DR - One-Sentence Conclusion

On September 3, 2026 (US Eastern Time), OpenAI released its new flagship model GPT-6 Astra — the most important coordinate event of the AI industry in 2026 Q3. OpenAI president Greg Brockman closed the press briefing with “Welcome to the AGI era” and said “future people may look back at this time, this model, as the moment AGI arrives.” Astra was trained on more than 100,000 GPUs at the Texas Stargate site over 117 days (OpenAI’s largest training run in history), ships with a 1.05M-token context window / 128K-token max output / April 30, 2026 knowledge cutoff, and is priced at $10 input / $50 output per million tokens (2.5x GPT-5.6 Sol, identical to Anthropic Claude Fable 5.1). The benchmark scoreboard reads FrontierMath Tier 4 v2 97.6%, ARC-AGI-3 jumping from Sol 7.8% to 98.6% (12.6x leap), and ExploitBench 100%. It is the first OpenAI model to cross the “Critical” cybersecurity threshold under the Preparedness Framework, with the strongest cyber capabilities gated behind the Daybreak Access program for vetted organizations. Native Computer-Use collapses average task time from 75 to 40 minutes (-47%), scoring 72.6% on OSWorld 2.0 Offline. This marks the global LLM competition shifting from “who writes better answers” to “who finishes the whole job for you” — and OpenAI’s first public declaration that the AGI era has arrived.

Today’s Headlines: Nine Coordinates for Early September 2026

#SignalKey FactStrategic Meaning
1AGI claim goes publicBrockman: “Welcome to AGI era”; AGI no longer tied to Microsoft contract clauseOpenAI defines the era marker; industry narrative shifts
2100,000-GPU trainingTexas Stargate site, 117-day continuous run; largest everCompute arms race moves from “10K-card” to “100K-card” era
3RSI supervised trainingAstra is OpenAI’s first flagship with prior-generation models deeply supervising trainingRecursive Self-Improvement (RSI) lands at the flagship tier
4ARC-AGI-3 12.6x leapSol 7.8% → Astra 98.6% (Responses API harness)General reasoning crosses the “practical AGI threshold”
5ExploitBench 100%100% on vulnerability-exploitation benchmark vs Sol 78.5% / Opus 5 70%First Critical-tier cyber model; offense-and-defense capabilities both break
6Computer-Use nativeOSWorld 2.0 Offline 72.6%; task time 75min → 40min”Mouth substitute” → “Hand substitute” paradigm leap; AI starts operating real software
72.5x price + 70% token efficiencyAPI $10/$50, but Coding Agent Index per-task cost ≈ Sol”Per-task cost” replaces “per-token price”
8Daybreak Access gatingCyber capabilities only to vetted Daybreak orgs; Plus/Pro/Enterprise queuedSafety friction lands as a commercial product for the first time
9Codex persistent notesCodex harness replaces compact with persistent notes across context windowsAgents on long-horizon tasks no longer “lose details”

I. Training Scale: Texas Stargate × 100,000 GPUs × 117 Days = OpenAI’s Largest Training Run

OpenAI researcher Aidan Clark confirmed at the press briefing: Astra was pretrained on more than 100,000 GPUs at the Texas Stargate site over 117 days — the largest model training in OpenAI’s history. President Brockman also confirmed: Astra is OpenAI’s first flagship where prior-generation models played a significant role in supervising the training — meaning Recursive Self-Improvement (RSI) landed at the flagship tier for the first time.

DimensionGPT-6 AstraGPT-5.6 Sol (8/24)Leap
Training hardware100K+ GPUs Stargate Texas~20K GPUs (est.)5x
Context window1.05M tokens (922K max input)~256K tokens~4x
Max output128K tokens~32K tokens~4x
Knowledge cutoff2026-04-302026-01+4 months
Training RSIPrior models supervisedNoneNew paradigm
API input price$10/M$4/M2.5x
API output price$50/M$20/M2.5x

Sources: OpenAI official release (Axios 9/3 18:00 + OpenAI Deployment Safety Hub + OpenAI Developers) + AlphaSignal 9/3 + HuggingNews 9/3 + QbitAI 9/4 05:27 + 36Kr 9/2 15:59.

II. Computer-Use Paradigm Leap: From “Mouth Substitute” to “Hand Substitute” — AI Takes Over Real Software

This is Astra’s most disruptive capability leap: Computer-Use native — AI no longer outputs pseudocode or step-by-step instructions, but actually takes over mouse and keyboard (virtual layer), clicking through Excel, dragging formulas, inserting Power Query, connecting Snowflake databases, running Monte Carlo simulations, and stuffing the results into a polished deck — without any human window-switching, copy-pasting, or error-correction in between.

2.1 OSWorld 2.0 Offline: 65.7% → 72.6%, Average Task Time 75min → 40min (-47%)

BenchmarkGPT-6 AstraGPT-5.6 SolClaude Opus 5Improvement
OSWorld 2.0 Offline (desktop tasks)72.6%65.7%~58%+6.9 pp / -47% time
BenchCAD Vision2Code95.9%83.3%-+12.6 pp
Terminal-Bench Science64.6%~40%30% (public leaderboard)+24.6 pp
Mind2Web task completion speed1.9x Sol1x-+90%

2.2 OpenAI’s Official Demonstration Scenarios (Real Software Stack)

  • KiCad: automated PCB circuit board routing
  • Unity: 3D city scene modeling
  • FreeCAD + Blender: runnable automobile transmission animation
  • Excel + PowerPoint + Snowflake: read financial PDF + meeting minutes + competitor data → auto-model → dynamic charts → KeyNote file (auto-applies corporate VI fonts and sizing)
  • W-2 → tax return draft: end-to-end from payroll stub to complete tax filing
  • Mathematics research: helped improve a result on gaps between prime numbers + new marks on multiple biology/chemistry/medical/physics evaluations

Sources: OpenAI press briefing (Axios 9/3) + AlphaSignal 9/3 + NetEase 9/4 02:58 + QbitAI 9/4 05:27 + Sohu 9/4 05:30.

III. Benchmark Scoreboard: ARC-AGI-3 12.6x / FrontierMath 97.6% / ExploitBench 100%

Three numbers on OpenAI’s official scoreboard approach “saturation”, which QbitAI described as “anyone who looks at them will be amazed”:

BenchmarkGPT-6 AstraGPT-5.6 SolClaude Opus 5Leap
ARC-AGI-3 (general reasoning)98.6% (Responses API harness) / 62.7% (default harness)7.8%30.2% (high)12.6x leap
FrontierMath Tier 4 v2 (hard math)97.6%83.0%73.2%+14.6 pp
ExploitBench (vulnerability exploitation)100%78.5%70%+21.5 pp
DeepSWE v1.1 (agentic coding)74.1%70.8%~74%+3.3 pp (Meta Muse Spark 1.3 75.4%)
GPQA Diamond96.0%94.6%93.2%+1.4 pp
AA Intelligence Index6161Fable 5.1 = 66Ties Sol, -5 vs Fable 5.1
AA Coding Agent Index (Codex)67~58Fable 5.1 = 70 (Fable 5.1 in Claude Code leads)+9
AA-Omniscience hallucination rate51% (-41 pp)92%--41 pp

3.1 The “12.6x ARC-AGI-3 Leap” Needs a Fine Print

OpenAI’s self-reported 98.6% / 99.9% used the Responses API harness (which preserves reasoning state across requests + compacts long conversations); TPS Report reproduced 62.7% under the default harness. NVIDIA previously demonstrated that the harness alone can lift Claude Opus 5’s ARC-AGI-3 from ~30% to 100% — this debate was reignited by Astra. Readers should treat 98.6% as “the entire agent system” score, not “the bare model” score.

Sources: Artificial Analysis 9/3 benchmark article + AlphaSignal 9/3 + HashLytics 9/3 + QbitAI 9/4 + OpenAI self-reported numbers (via Axios).

IV. First Critical-Tier Cybersecurity Model: Daybreak Access + Monitoring Friction Lands as a Product

Astra is the first model in OpenAI’s history to be formally rated as crossing the “Critical” cybersecurity capability threshold — meaning that, when given tools and permissions, the model can autonomously find previously unknown vulnerabilities, develop exploit programs, and attack hardened systems without step-by-step human guidance. This is the highest cybersecurity tier in OpenAI’s Preparedness Framework.

4.1 ExploitBench 100% + SRE-Bench 99.2% — What It Actually Means

BenchmarkGPT-6 AstraGPT-5.6 SolLeap
ExploitBench (public vulnerability exploitation)100%78.5%+21.5 pp
ExploitGym (real-world)42.4%30.3%+12.1 pp
SRE-Bench (4 attempts, reverse engineering)99.2%68.7%+30.5 pp
20 high-severity V8 CVEs from 2026/6–8 (post-knowledge-cutoff)Significantly ahead of Sol with fewer tokens (≈4x capability per researcher Jiawei Liu)-New dimension of leadership

4.2 Daybreak Access Gating = Safety Friction Lands as a Commercial Product

OpenAI did not release Astra in full — instead, the strongest Cyber capabilities are restricted to vetted Daybreak organizations through the Daybreak Access program. Regular Plus / Pro / Business / Enterprise users can call Astra, but the most advanced Cyber capabilities are stripped out. This is OpenAI’s first time handling cybersecurity risk through commercial product tiering.

Sources: OpenAI Deployment Safety Hub + Axios 9/3 + 36Kr 9/2 + QbitAI 9/4 + HashLytics 9/3.

V. The AGI Debate: Brockman’s “Welcome to the AGI Era” + Microsoft Contract Decoupling

5.1 Brockman’s Key Statements

“I think it might be about this model.” — Brockman at the press briefing (9/3)

“AGI is no longer tied to a contract trigger condition in our previous agreement with Microsoft; it has become more of a mission concept or spiritual concept.” — Brockman 9/3

This means AGI in OpenAI’s narrative has been “de-contractualized” — no longer a specific technical milestone, but a “mission concept.” OpenAI can choose to declare “AGI has arrived” at some moment without triggering contractual legal obligations.

5.2 The Monitorability Controversy: OpenAI Acknowledges Astra Is “Harder to Monitor”

Chief scientist Jakub Pachocki admitted openly: Astra was genuinely harder to monitor in evaluations designed to test whether it could evade oversight. “We will need to strengthen our ability to monitor these models either via extending chain-of-thought monitoring, integrating other ideas like activation monitoring, or finding more specific ways to get the models to be more verbose in their chain of thought.” This is OpenAI’s first time actively acknowledging that a new-generation flagship regresses on monitorability, rather than only advertising progress.

Sources: Axios 9/3 18:00 + Slashdot 9/3 + HashLytics 9/3 + 36Kr 9/2 + NetEase 9/4 04:29.

VI. 2.5x Price + 70% Token Efficiency: Brockman Redefines “Per-Task Cost”

6.1 Surface Comparison: Price 2.5x

DimensionGPT-6 AstraGPT-5.6 Sol (promo)Claude Fable 5.1Claude Opus 5
Input $/M tokens$10$4$10$5
Output $/M tokens$50$20$50$25
Cache reads $/M$1$0.40$0.25 (75% drop from Fable 5)-
Cache writes $/M$12.5$5--
Above 272K input priceInput + Output doubleNo changeNo changeNo change
Fast mode2x price, 2.5x throughputNo changeNo changeNo change

6.2 Actual Truth: Token Efficiency Surges; Per-Task Cost Roughly Flat

Artificial Analysis data shows:

  • Coding Agent Index (Codex harness): 70% more token-efficient — Astra uses 1/3 the tokens of GPT-5.6 Sol (max), 1/5 of Claude Opus 5 (xhigh)
  • Per-task cost: At max effort, Astra ≈ GPT-5.6 Sol, less than half the cost of Claude Fable 5 for the same task
  • AA Intelligence Index: ~10% fewer output tokens, offset by 2.5x price — +75% per task vs Sol

6.3 Brockman’s Counter: “Pricing per Token Makes No Sense”

“Pricing tokens doesn’t make any sense… the metric that matters is price per completed task.” — Brockman 9/3

Brockman argued: when models become autonomous Agents, the unit of pricing should be “per-task cost”, not “per-token price”. This pricing-philosophy pivot echoes Anthropic Opus 4.x’s early “per-task pricing” experiments, but is the first time an OpenAI flagship has stated it explicitly.

Sources: Artificial Analysis 9/3 + HashLytics 9/3 + AlphaSignal 9/3 + Digital Applied 9/3 + OpenAI official pricing page.

VII. First Week of September: Three Major Labs All Ship Flagships — Industry Coordinate Reset

Astra is not an isolated release — in the first week of September (8/31–9/3), three global labs collectively shipped flagships, marking 2026 Q3’s entry into “Q4 flagship preheating period”:

DateCompanyModelTypeKey Pricing/Capability
8/31DeepSeekDeepSeek-V4-Flash-Vision-ExpMultimodal experimentMIT license, free multimodal
9/1AnthropicClaude Fable 5.1 + Mythos 5.1GA$10/$50, cache reads $0.25 (-97%)
9/2QwenQwen3.8-Max-09022.4T/95B Max-class1M context, Code Arena WebDev #1
9/2GoogleGemini 3.8 Flash + 3.8 Flash CyberFlash series$0.75/$3.75, doubles 2027/1/1
9/3OpenAIGPT-6 Astra + Astra ProGA$10/$50, 100K GPUs, first Critical Cyber

7.1 China Side Sync: 9/1 Qwen3.8-Flash-Next API Goes Live + Qwen3.8-Max Update

Alibaba opened Qwen3.8-Flash-Next developer API on 9/1 (Qwen4 architecture preview), API pricing $0.16/M input — 1/12 the flagship Qwen model; updated Qwen3.8-Max on 9/2 hit 1691 on Code Arena WebDev (narrowly above Opus 5’s 1688). The US-China LLM gap narrows further into “homogenized competition + tiered pricing” — OpenAI pulls ahead with “Critical Cyber + Computer-Use”, while the China side holds the base with “1/12 pricing + 1M context + 2.4T Qwen3.8-Max-0902”.

Sources: llmgateway.io 9/2 timeline + Mehmet Baykar Top Frontier LLMs Release Tracker + aikitapp.com Model Watch 9/2 + pondero.ai 9/2 daily brief + PolyU Gen AI App 9/1 update announcement.

VIII. 5-Step Enterprise Landing Path (Adapted to the Computer-Use Paradigm Leap)

  1. Assess agentic capability maturity: Use OSWorld 2.0 / WebArena / Mind2Web to evaluate whether your core business flows can be replaced by Astra Computer-Use — Astra scores 72.6% on cross-software end-to-end workflows; for real business, start from “assistive filling”, not “full automation”
  2. Codex harness persistent-notes bonus: Enable Astra’s Codex immediately — persistent notes across context windows is more suitable for long-horizon tasks than GPT-5.6 Sol’s compact mode — single-task tokens down 70%
  3. Daybreak Access application: If your business involves security research / red-blue drills / bug bounties, the first vetted-org batch is open — apply via Daybreak for full Cyber capabilities
  4. Pricing strategy pivot: Shift from “token-based monthly budget” to “task-based budget” — Coding Agent scenarios per-task cost ≈ Sol (token efficiency offsets price), pure dialogue/writing scenarios per-task +75% (price rises faster than token savings)
  5. Monitorability fallback plan: Build internal “activation monitoring + chain-of-thought audit” dual-track for Astra’s “harder to monitor” property — do not let Astra touch production systems directly; run in sandbox environments for 30 days first

IX. 6-Point Defense Checklist

  1. Astra Cyber capability ≠ full release — Plus/Pro users do not get Critical-tier Cyber; don’t overestimate your actual available capability
  2. ExploitBench 100% is OpenAI self-reported — Independent reproduction results (Artificial Analysis / TPS Report) are the ground truth; beware the gap between self-reported 98.6% and default-harness 62.7%
  3. 2.5x price does not equal 2.5x per-task cost — Coding Agent tasks are cheaper; Fable 5.1 long-context writing scenarios are more expensive; price by scenario, not by token
  4. Astra’s “AGI era” claim ≠ industry consensus — Brockman himself said AGI is a “mission concept”; don’t take marketing talk as a technical milestone
  5. Computer-Use operating real software ≠ replacing humans — OSWorld 2.0 72.6% means 27.4% of tasks will fail; critical decisions still need human fallback
  6. Monitorability regression requires internal compensation — OpenAI itself says Astra is harder to monitor; enterprises should build independent activation-monitoring systems rather than relying solely on OpenAI’s chain-of-thought audit

FAQ (High-Frequency Questions Answered Directly)

Q1: What exactly is GPT-6 Astra? A: OpenAI’s 9/3/2026 flagship large model, code-named Astra (Stellar), with a 1.05M-token context window, knowledge cutoff April 30, 2026, API priced at $10/$50 per million tokens, focused on native Computer-Use and Critical-tier cybersecurity capabilities.

Q2: Why does OpenAI say “the AGI era has arrived”? A: Brockman said at the launch “future people may look back at this time, this model, as the moment AGI arrives”, but also stressed AGI is no longer tied to the Microsoft contract clause — it’s a “mission concept” — this is marketing talk and strategic signaling, not industry consensus.

Q3: Is the ARC-AGI-3 leap from 7.8% to 98.6% real? A: OpenAI’s self-reported 98.6% used the Responses API harness (which preserves reasoning state across requests); TPS Report’s default-harness reproduction is only 62.7%; NVIDIA previously proved the harness alone can lift Opus 5 from 30% to 100%. Readers should treat it as “the entire agent system” score, not “the bare model” score.

Q4: Why is Computer-Use a paradigm leap? A: All previous LLMs could only “output answers”; real execution still required humans to operate software; Astra natively takes over mouse and keyboard, completing work end-to-end inside real software stacks (KiCad / Unity / Excel / Snowflake); OSWorld 2.0 task time collapses from 75min to 40min (-47%).

Q5: Astra’s price rose 2.5x — is it still worth using? A: It depends on the scenario — Coding Agent scenarios per-task cost ≈ GPT-5.6 Sol (token efficiency offsets price, less than half the cost of Claude Fable 5); pure dialogue/writing scenarios per-task +75%; above 272K tokens input/output prices double. Recommend task-based pricing, not token-based pricing.

Q6: What is the Daybreak Access program? A: OpenAI’s vetted-organization gating program; first-batch Astra Cyber capabilities are only open to Daybreak-participating organizations; regular Plus/Pro users do not get Critical-tier Cyber. This is OpenAI’s first time handling cybersecurity risk through commercial product tiering.

Q7: What impact does Astra have on Chinese LLMs? A: The US-China gap further narrows — OpenAI pulls ahead with “Critical Cyber + Computer-Use”; the China side holds the base with “1/12 pricing (Qwen3.8-Flash-Next $0.16/M) + 1M context + 2.4T Qwen3.8-Max-0902”; OpenAI is no longer “decisively ahead”, entering “tiered differentiated competition”.

Key Terminology

  • AGI (Artificial General Intelligence): Internally defined by OpenAI as a “mission concept or spiritual concept”, not a specific technical milestone; Brockman said on 9/3 that Astra “may be the moment AGI arrives”.
  • Critical cybersecurity capability threshold (OpenAI Preparedness Framework): The highest cybersecurity capability tier in OpenAI’s internal framework, meaning the model can autonomously discover unknown vulnerabilities and develop exploit programs; Astra is the first model to reach this tier.
  • Computer-Use: AI takes over a real software stack (Excel / Unity / KiCad / Snowflake) to execute tasks end-to-end, not just outputting suggestions; OSWorld 2.0 is the standard benchmark for this capability; Astra scores 72.6%.
  • Daybreak Access: OpenAI’s vetted-organization gating program; first-batch Astra customers are limited to security-research institutions participating in this program; Plus/Pro users’ Cyber capabilities are stripped.
  • ExploitBench: A public vulnerability-exploitation benchmark; Astra scores 100%, far above GPT-5.6 Sol’s 78.5% and Claude Opus 5’s 70%.
  • ARC-AGI-3: A benchmark measuring general reasoning capability (released March 2026); Astra self-reports 98.6% (using Responses API harness); default-harness reproduction is 62.7%.
  • Recursive Self-Improvement (RSI): Using prior-generation models to supervise the training of new models; Astra is OpenAI’s first flagship to apply RSI.
  • Codex harness persistent notes: A new Codex mechanism introduced by Astra, replacing compact-summary compression with persistent-note retrieval across context windows; agents on long-horizon tasks no longer “lose details”.

References

Official Sources

Independent Benchmarks & Reviews

Chinese Authoritative Media

English Authoritative Media

Industry Timeline & Same-Week Releases

JSON-LD Structured Data

{
  "@context": "https://schema.org",
  "@type": "NewsArticle",
  "headline": "OpenAI GPT-6 Astra Enters the 'AGI Era': Computer-Use Paradigm Leap + 100,000-GPU Training + First Critical-Tier Cybersecurity Model + 'Executing Tasks' Beyond 'Answering Questions'",
  "datePublished": "2026-09-04T06:00:00+08:00",
  "dateModified": "2026-09-04T06:00:00+08:00",
  "author": {
    "@type": "Organization",
    "name": "BoaoAI Editorial",
    "url": "https://www.boaoai.cn"
  },
  "publisher": {
    "@type": "Organization",
    "name": "BoaoAI",
    "logo": {
      "@type": "ImageObject",
      "url": "https://www.boaoai.cn/logo.png"
    }
  },
  "description": "On September 3, 2026 (US Eastern Time), OpenAI released its new flagship model GPT-6 Astra, with president Greg Brockman closing the launch with 'Welcome to the AGI era'. Astra was trained on more than 100,000 GPUs at the Texas Stargate site (OpenAI's largest training run in history), features a 1.05M-token context window with up to 128K-token output, knowledge cutoff April 30, 2026, and is priced at $10/$50 per million tokens (2.5x GPT-5.6 Sol). It scores 97.6% on FrontierMath Tier 4 v2, lifts ARC-AGI-3 from Sol's 7.8% to 98.6% (12.6x leap), and achieves 100% on ExploitBench — making it the first OpenAI model to cross the 'Critical' cybersecurity threshold under the Preparedness Framework, gated behind the Daybreak Access program. Native Computer-Use collapses average task time from 75 to 40 minutes (-47%) and scores 72.6% on OSWorld 2.0 Offline.",
  "keywords": "OpenAI, GPT-6, Astra, AGI Era, Computer-Use, Critical Cybersecurity, ExploitBench, ARC-AGI-3, Stargate 100000 GPU, Daybreak Access",
  "inLanguage": "en-US",
  "mainEntityOfPage": "https://www.boaoai.cn/en/news/2026-09-04-openai-gpt-6-astra-agi-era-computer-use-critical-cyber/"
}
{
  "@context": "https://schema.org",
  "@type": "BreadcrumbList",
  "itemListElement": [
    {
      "@type": "ListItem",
      "position": 1,
      "name": "Home",
      "item": "https://www.boaoai.cn/"
    },
    {
      "@type": "ListItem",
      "position": 2,
      "name": "AI News",
      "item": "https://www.boaoai.cn/en/news/"
    },
    {
      "@type": "ListItem",
      "position": 3,
      "name": "GPT-6 Astra AGI Era",
      "item": "https://www.boaoai.cn/en/news/2026-09-04-openai-gpt-6-astra-agi-era-computer-use-critical-cyber/"
    }
  ]
}
{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "What exactly is GPT-6 Astra?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "OpenAI's 9/3/2026 flagship large model, code-named Astra (Stellar), with a 1.05M-token context window, knowledge cutoff April 30, 2026, API priced at $10/$50 per million tokens, focused on native Computer-Use and Critical-tier cybersecurity capabilities."
      }
    },
    {
      "@type": "Question",
      "name": "Why does OpenAI say 'the AGI era has arrived'?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Brockman said at the launch 'future people may look back at this time, this model, as the moment AGI arrives', but also stressed AGI is no longer tied to the Microsoft contract clause — it's a 'mission concept' — this is marketing talk and strategic signaling, not industry consensus."
      }
    },
    {
      "@type": "Question",
      "name": "Is the ARC-AGI-3 leap from 7.8% to 98.6% real?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "OpenAI's self-reported 98.6% used the Responses API harness (which preserves reasoning state across requests); TPS Report's default-harness reproduction is only 62.7%; NVIDIA previously proved the harness alone can lift Opus 5 from 30% to 100%. Readers should treat it as 'the entire agent system' score, not 'the bare model' score."
      }
    },
    {
      "@type": "Question",
      "name": "What is the Daybreak Access program?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "OpenAI's vetted-organization gating program; first-batch Astra Cyber capabilities are only open to Daybreak-participating organizations; regular Plus/Pro users do not get Critical-tier Cyber. This is OpenAI's first time handling cybersecurity risk through commercial product tiering."
      }
    }
  ]
}