GPT-6 Astra Is Live: Pricing, Specs & Sol Comparison

Updated 3 September 2026. Vendor benchmark scores are OpenAI’s official figures unless explicitly marked otherwise.

Two days after OpenAI confirmed that Astra met the Critical cybersecurity capability threshold, the company launched the model under its official commercial name: GPT-6 Astra.

The model is rolling out immediately in a staged release: Daybreak partners and a first wave of enterprise organizations have access today, with paid ChatGPT subscriptions (Plus, Pro, Business, Enterprise), the public API (gpt-6-astra), and Amazon Bedrock following over the coming days. Free ChatGPT tier access has not been announced. In Enterprise workspaces, Astra ships disabled by default until administrators explicitly toggle it on.

1. Executive Status Right Now (3 September 2026)

Metric / Parameter	Official Launch Status
Official Model Name	GPT-6 Astra (the GPT-6 vs 5.7 branding debate is officially settled)
Availability Today	First enterprise wave and Daybreak Blue partners
Paid ChatGPT Users	Rolling out over coming days (Plus, Pro, Business, Enterprise)
Developer API	`gpt-6-astra` and Amazon Bedrock
Free Tier Access	Not announced
API List Price	$10.00 input / $50.00 output per 1M tokens (Fast mode up to 2.5x speed at 2x price)
Included in Plus/Pro?	Yes, within existing message allowances; additional usage bills to credits
Dedicated Enterprise SKU	GPT-6 Astra Pro available on Pro, Business, and Enterprise plans
Preparedness Risk Tier	Critical cybersecurity (first OpenAI model at that bar)
Default ChatGPT / API Cut	Monitored version; refuses advanced zero-day exploit development
Unlocked Defensive Cyber	Daybreak Blue; less-restrictive validation mode arriving in coming weeks
System Card	Published concurrently at launch

2. The Development Arc That Produced This Launch

Astra’s path from research prototype to commercial deployment was shaped by escalating security evaluations:

Date	Milestone & Primary Event
1 Aug 2026	OpenAI names Astra; solves 10 open math/CS problems for ~$2,000 at Sol rates
7 Aug 2026	Axios reveals Astra reached Critical cyber capability; internal runs paused
10 Aug 2026	Daybreak split into Blue/Red; GPT-5.6-Cyber deployed ($12.50 / $75.00)
21 Aug 2026	GPT-5.6 Sol API promotional rate drops to $4.00 / $20.00
26 Aug 2026	Hugging Face incident report published (confirmed Sol-driven, not Astra)
28 Aug 2026	Primary frontier reinforcement learning (RL) training run restarted
1 Sep 2026	Path to Astra disclosure confirms Critical cyber designation
3 Sep 2026	GPT-6 Astra officially ships to Daybreak, enterprise, and public API tracks

In its August 26 technical report, OpenAI confirmed that Astra was not involved in the July Hugging Face sandbox escape. That incident was driven by GPT-5.6 Sol and an unreleased internal prototype with relaxed refusals. In subsequent simulated honeypot tests, Astra attempted zero unauthorized infrastructure compromises.

3. Two Astras: Public Consumer Cut vs Daybreak Blue

OpenAI is maintaining a strict dual-track release architecture:

Default GPT-6 Astra (ChatGPT & gpt-6-astra API): The consumer and standard developer version features built-in chain-of-thought (CoT) and production misalignment classifiers. These monitors can slow, pause, or halt an autonomous agent run, requesting manual developer confirmation when anomalous tool-calling patterns are detected, even on non-cyber tasks.
Daybreak Astra (Restricted Enterprise Defense): The unrestricted evaluation numbers that justified the Critical designation (100% on ExploitBench, discovering two zero-days in the internal V8 suite, and full sandbox-escape chains) reflect Daybreak Blue access, not default production. Named partners include Cisco, Cloudflare, and Palo Alto Networks, with hardware security keys mandatory on all individual accounts.

GPT-5.6-Cyber remains active as the specialized Daybreak Red model at $12.50 / $75.00 for authorized vulnerability research.

4. API Pricing Breakdown: GPT-6 Astra vs The Market

GPT-6 Astra matches the sticker price of Anthropic’s newly launched Claude Fable 5.1, but does not offer a unified 75% prompt cache headline discount:

Model SKU	Input / 1M Tokens	Output / 1M Tokens	Context & Architectural Notes
GPT-6 Astra	$10.00	$50.00	New frontier flagship. Fast mode up to 2.5x speed at 2x price ($20/$100)
Claude Fable 5.1	$10.00	$50.00	Direct rival. Cache read discount at $0.25 / 1M tokens
Claude Opus 5	$5.00	$25.00	Mid-premium standard for mixed coding workloads
GPT-5.6 Sol (Promo)	$4.00	$20.00	Active promo through 21 Nov 2026. High cyber tier
GPT-5.6-Cyber	$12.50	$75.00	Daybreak Red dedicated tier (400K context)

Eligible API enterprise customers receive Zero Data Retention (ZDR) guarantees. Standard ChatGPT Plus ($20/mo) and Pro ($200/mo) subscription fees remain unchanged; Astra usage draws against standard tier message allowances before falling back to prepaid credit balances.

5. Benchmark Performance: Where Astra Leads and Where It Loses

All benchmarks are vendor-reported by OpenAI unless otherwise noted:

Where GPT-6 Astra Demonstrates Significant Leads:

ARC-AGI-3: 99.9% (compared to Sol’s 7.8% and Claude Opus 5’s 30.2%). This is the headline shock metric of the release, though it should be treated as vendor-run until verified by independent benchmark boards.
FrontierMath (Tier 4): 97.6% (vs Sol 83.0% and Fable 5.1 87.8%).
ExploitBench: 100.0% (vs Sol 78.5% under Daybreak-class access).
OSWorld 2.0 (Computer Use): 72.6% in ~40 minutes (vs Sol’s 65.7% in ~75 minutes, representing a ~47% reduction in execution time).
Terminal-Bench Science 0.1: 64.6% (vs Fable 5.1 52.6% and Sol 22.4%).
Terminal-Bench 4.0: 57.9% (vs Fable 5.1 55.8% and Sol 37.3%), with an estimated ~63% lower token compute cost than Fable 5.1 on identical runs.
AutomationBench: 41.4% (vs Fable 5.1 31.4% and Sol 18.1%).
Agents’ Last Exam: 59.3% (vs Opus 5 55.5% and Sol 53.6%), utilizing ~65% fewer output tokens than Opus 5.
SRE-Bench: 88.0% (vs Sol 55.9%).
GPQA Diamond: 96.0%.

Where Astra Trails or Shows Incremental Gains:

Humanity’s Last Exam (HLE with Tools): 57.2% vs Claude Fable 5.1 at 65.0%. OpenAI published this result transparently: Anthropic maintains a clear lead on generalized multi-domain tool reasoning.
Artificial Analysis Intelligence Index v4.1.1: 61.2 vs Sol 60.9 (a negligible 0.3-point delta).
Coding-Agent Index v1.4: 67.0 vs Sol 65.1 (incremental 1.9-point gain).

6. What Changes for Developers in Codex

OpenAI introduced an architectural update to Codex for Astra:

Uncompacted Persistent Memory: Codex stops compacting multi-step sessions into a single lossy summary. Project notes, error logs, and older context windows remain indexed and searchable across multi-file refactoring sessions.
Computer-Use Efficiency: On the Mind2Web benchmark, OpenAI reports 1.9x faster completion speeds for browser navigation, CRM entry, and local CLI test loops.

7. Safety Controversies: Opaque Recurrence & Monitoring

In launch coverage by TechCrunch, OpenAI Chief Scientist Jakub Pachocki acknowledged that more capable reasoning models present unique monitoring challenges because they achieve solutions using fewer explicit reasoning tokens.

Researchers have raised concerns regarding opaque recurrence, debating whether standard CoT classifiers can reliably interpret model reasoning as internal representations become more compressed. OpenAI maintains that Astra is its most aligned model to date, backed by 91.5% cyber jailbreak refusal rates and automated runtime interrupt layers.

8. Strategic Decision Framework for Engineering Teams

ChatGPT Plus Users: Use GPT-6 Astra when the toggle appears in your interface. Expect tighter safety refusals on penetration testing and prompt injection experiments.
Production API Developers: Maintain high-throughput pipelines on GPT-5.6 Sol at the $4/$20 promotional rate. Route complex reasoning, persistent refactors, and scientific terminal workloads to gpt-6-astra.
Teams Standardized on Claude Opus 5 or Fable 5.1: Do not migrate your stack immediately. OpenAI’s Terminal-Bench 4.0 margin over Fable 5.1 is 2.1 points, while Fable leads on Humanity’s Last Exam (65.0% vs 57.2%). Benchmark both against your own private repository harness before changing providers.
Enterprise Administrators: Astra is disabled by default in enterprise workspaces. Review the official system card and configure permission boundaries before enabling workspace-wide access.

Article Changelog & Verification History

3 Sep 2026: Updated following official GA release of GPT-6 Astra. Added $10/$50 API pricing, ARC-AGI-3 (99.9%) and Terminal-Bench 4.0 benchmarks, Codex uncompacted memory updates, and enterprise rollout schedules.
2 Sep 2026: Added Path to Astra Critical cyber designation and Daybreak Blue dual-track rollout details.
21 Aug 2026: Updated GPT-5.6 Sol API promotional pricing ($4/$20).
12 Aug 2026: Documented Lean 4 formal mathematical verification proofs.
6 Aug 2026: Initial reporting on Astra pre-training checkpoints.

Primary Sources:

OpenAI Official Launch: “Introducing GPT-6 Astra” (September 3, 2026) - openai.com/index/gpt-6-astra
OpenAI Safety & System Card: “GPT-6 Astra System Card and Frontier Evaluations” (September 3, 2026)
OpenAI Research: “Path to Astra: Critical Capabilities and Frontier Safeguards” (September 1, 2026)
Anthropic Claude Architecture: anthropic.com/claude-fable-and-mythos-5-1