GPT-6 Astra review: pricing, Daybreak access, Critical cyber

Key takeaways

• GPT-6 Astra launched September 3, 2026 as OpenAI’s new flagship after the GPT-5.6 family (Sol / Terra / Luna).
• API model id listed as gpt-6-astra. Short-context list price: $10 / 1M input, $50 / 1M output (cached input $1.00; cache writes $12.50). Long-context rates: $20 / $75 (cached $2.00; cache writes $25.00).
• Access is staged: Trusted Access / Daybreak first; Plus, Pro, Business, Enterprise, API, and AWS “in the coming days.” Not instantly GA for every account on day one.
• First OpenAI model designated at the Critical cybersecurity capability threshold. Advanced cyber workflows stay gated (alpha testers, then Daybreak Blue). Default product access is more restrictive on exploit-style work.
• GPT-6 Astra Pro called out for Pro / Business / Enterprise; eligible API customers get Zero Data Retention options (per launch coverage).
• Vendor benches are strong on agentic coding, science, and computer use. Treat them as OpenAI harness + max effort unless you reproduce them. Peer coding numbers overlap with Muse Spark 1.3 and Gemini 3.8 Flash on some public boards.

Table of contents

1. What GPT-6 Astra is
2. Pricing
3. Access and rollout
4. Cyber Critical and Daybreak
5. What OpenAI claims improved
6. Who should switch from Sol
7. FAQ

---

What GPT-6 Astra is

The naming fight is over for now. OpenAI’s ChatGPT release notes for September 3, 2026 introduce GPT-6 Astra by that full name.

For practitioners:

GPT-6 Astra is OpenAI’s new flagship model for coding, research, computer use, and long multi-step agent work. It sits above the GPT-5.6 Sol / Terra / Luna ladder, with Critical-tier cyber capability and staged access.

That separates:

• ChatGPT / Codex / API product surface: where most people will eventually pick Astra.
• GPT-5.6 Sol (previous flagship): still the live workhorse many orgs run at promotional $4/$20 until they migrate.
• Daybreak / Daybreak Blue: access programs for trusted defensive cyber workflows, not a separate “smarter Astra” you toggle in the consumer picker.
• Astra Pro: higher-tier product access for Pro / Business / Enterprise (confirm in your plan UI as it rolls out).

OpenAI and press briefings also framed Astra as its largest training run to date (100k+ GPUs at Stargate, per company remarks covered by The New Stack). Treat that as vendor color, not a buying criterion.

Pricing

Verified September 3, 2026 against OpenAI’s API pricing table.

Meter	Short context	Long context
Input	$10.00 / 1M	$20.00 / 1M
Cached input	$1.00 / 1M	$2.00 / 1M
Cache writes	$12.50 / 1M	$25.00 / 1M
Output	$50.00 / 1M	$75.00 / 1M

Same sticker shape as Claude Fable 5.1 ($10/$50). About 2.5× GPT-5.6 Sol’s promotional $4/$20 API rates. Business/Enterprise ChatGPT credit cards list Astra at 250 / 25 / 1,250 credits per 1M input / cached / output (Help Center rate card). Fast mode on Astra in Work/Codex is billed at 2.5× Standard credit rates per that card.

OpenAI’s public line: price per task matters more than price per token if Astra finishes with fewer retries. Prove that on your harness before you rewrite budgets.

Access and rollout

Per OpenAI Help Center (Sep 3, 2026):

• Rolling out first to a limited set of organizations.
• Not yet generally available on day one.
• Broader availability planned over the coming days for Plus, Pro, Business, and Enterprise, plus API and AWS (press briefing coverage).

Extra agent safety monitoring may pause or stop a conversation if the system thinks the agent misread instructions, so you can review. Expect more friction than Sol on some agent jobs, especially anything that looks like cyber work.

Cyber Critical and Daybreak

On September 1, OpenAI’s Path to Astra said Astra meets the Critical cybersecurity capability threshold under the Preparedness Framework: with the right tools and access, it can find previously unknown flaws and develop exploits across well-protected systems without step-by-step human guidance. First OpenAI model designated at that level.

Launch posture:

• Standard Astra access refuses or blocks some advanced cybersecurity workflows.
• Less restricted defensive access starts with a small alpha group, then expands via Daybreak Blue.
• API cybersecurity safety checks may stop a task outright rather than pause for approval (per launch coverage).
• Users outside trusted programs should expect slowdowns, pauses, or blocks on cyber-adjacent and sometimes unrelated work at launch.

This is the same industry pattern as Anthropic’s Mythos / Fable split and Google’s Flash Cyber Fairwind gate: flagship intelligence with a gated cyber twin or mode.

What OpenAI claims improved

Vendor-reported (max effort unless noted; harness-dependent):

• DeepSWE v1.1: 74.1% vs Sol 70.8% (OpenAI chart). Public/peer numbers the same week put Muse Spark 1.3 max around 75.4% and Gemini 3.8 Flash / Opus 5 near 74% on overlapping boards. Uncertainty bands overlap. Do not crown a coding king from one table.
• Strong showings cited on ARC-AGI-3, FrontierMath Tier 4, BenchCAD Vision2Code, Terminal-Bench Science, OSWorld V2-Offline (72.6% vs Sol 65.7%, faster wall time).
• Codex harness changes: notes that persist across context windows (experimental config.toml), ask-without-blocking for independent workstreams, faster Mind2Web with the new harness.
• Computer use demos across desktop apps and browser QA.
• Alignment narrative: fewer unauthorized target exits on an internal impossible-task test versus Sol without production safeguards (comparison caveats apply). Written reasoning may be harder to monitor than Sol on some evasion evals.

Always run your own Codex / Responses / Cursor evals.

Who should switch from Sol

Move when:

• You already pay for high-end ChatGPT / Codex and need the new agent memory and computer-use loop.
• Your Sol failures are long-horizon coding or research agents, not cheap classification.
• You can absorb ~2.5× Sol promo token rates or prove lower tokens-per-task.

Wait when:

• Astra is not in your picker yet (normal during staged rollout).
• Sol at $4/$20 still clears your evals.
• You need open cyber tooling without Daybreak Blue.

Peer shopping: Claude Fable 5.1, Muse Spark, Gemini 3.8 Flash. Roundup: GPT-6 Astra vs recent LLMs.