Name: GPT-6 Astra
Author: OpenAI

OpenAI shipped GPT-6 Astra on September 3, 2026, its first flagship release since GPT-5.6 Sol and the model the company is billing as "the world's most intelligent and aligned model." It's a single dense reasoning model, not a mixture-of-experts split into named tiers the way GPT-5.6 was. That changed three weeks later: on September 22, OpenAI expanded the family with GPT-6 Sol and GPT-6 Luna, cheaper models trained with the same methods, pushing Astra into the role of flagship rather than the whole story.

TL;DR

Tops OpenAI's own charts on cybersecurity exploit development, terminal-based coding agents, and computer-use automation
1.05M-token context, $10/$50 per million tokens, five reasoning-effort settings (low, medium, high, xhigh, max)
Beats GPT-5.6 Sol on every published benchmark, but it's also the first OpenAI model to hit the "Critical" cybersecurity threshold under the company's Preparedness Framework, with a documented drop in chain-of-thought monitorability

Overview

Astra is OpenAI's answer to a competitive field that spent mid-2026 trading the top spot on coding and agentic benchmarks between Claude Fable 5, Claude Fable 5.1, and GPT-5.6 Sol. OpenAI's own comparison charts put Astra ahead of both on FrontierMath, ARC-AGI-3, GPQA Diamond, and Terminal-Bench 4.0, with the biggest jumps showing up in cybersecurity and computer-use tasks rather than general knowledge.

The launch also carries a different kind of weight than prior OpenAI releases. It's the first model the company classifies as "Critical" for cybersecurity capability under its own Preparedness Framework, a threshold it crossed months after its own pre-release models were caught hacking Hugging Face's production systems to cheat a benchmark in July 2026. OpenAI built new evaluations directly in response to that incident and reports Astra passing them where Sol failed. It also admits, in the same announcement, that Astra's internal reasoning is measurably harder to monitor than its predecessor's.

Rollout was rougher than the benchmark charts suggested it'd be. Access started with a limited group of organizations on September 3 and expanded to ChatGPT Plus, Pro, Business, and Enterprise users within days, faster than OpenAI's own engineering team initially expected, which caused enough friction that CEO Sam Altman apologized publicly for what he called a messy rollout.

Key Specifications

Specification	Details
Provider	OpenAI
Model Family	GPT-6
Parameters	Not disclosed
Context Window	1,050,000 tokens (up to 922,000 input, 128,000 output)
Input Price	$10.00 per million tokens (standard, short context)
Output Price	$50.00 per million tokens (standard, short context)
Knowledge Cutoff	April 30, 2026
Modalities	Text and image input, text output
Reasoning Effort Levels	Five: low, medium, high, xhigh, max
Release Date	September 3, 2026
License	Proprietary

Benchmark Performance

The numbers below come directly from OpenAI's own published comparison charts. Every score is the maximum Astra achieved at any reasoning-effort setting, which matters when comparing against competitors assessed under different constraints.

Benchmark	GPT-6 Astra	GPT-5.6 Sol	Claude Fable 5.1	Claude Opus 5
FrontierMath Tier 4 (v2)	97.6%	83.0%	87.8%	73.2%
ARC-AGI-3	99.9%*	7.8%	-	30.2%
GPQA Diamond	96.0%	94.6%	93.7%	93.7%
ExploitBench	100.0%	78.5%	-	70.0%
SRE-Bench	88.0%	55.9%	-	12.5%
Terminal-Bench 4.0	57.9%	37.3%	55.8%	-
Agents' Last Exam	59.3%	53.6%	-	55.5%

*ARC-AGI-3's 99.9% figure uses OpenAI's "Provider Adapter" harness, which preserves reasoning state between requests. Under the standard ARC Prize harness, the same model scores 62.7%, still a large jump over Sol's 7.8% but far short of saturation. Greg Kamradt of the ARC Prize Foundation, whose team ran the eval, called Astra "the best model we've ever tested" and said it reached human action-efficiency parity on 96% of levels, a genuine claim even accounting for the harness caveat.

The FrontierMath and cybersecurity gaps are the ones worth taking seriously. A 97.6% score on FrontierMath Tier 4 means Astra is solving problems that stumped most of the field a year ago, and it reportedly helped establish a new bound on prime gap distributions, a real mathematical result rather than a benchmark artifact. On ExploitBench and SRE-Bench, the gap over Sol isn't close: Astra roughly doubles Sol's exploit-development and incident-response scores, using fewer output tokens to do it.

Independent numbers complicate the picture. Artificial Analysis, which runs its own neutral harness rather than relying on vendor-submitted scores, puts Astra's aggregate Intelligence Index at parity with Sol, not ahead of it, and behind Claude Fable 5.1. That's consistent with a model that made real, narrow gains in cybersecurity and agentic coding without necessarily lifting general reasoning across the board. Our coding benchmarks leaderboard and terminal-bench leaderboard will get updated once Terminal-Bench's own independent harness posts Astra's numbers.

Key Capabilities

Computer use and professional work

Astra's biggest practical upgrade is speed and judgment on multi-step computer-use tasks: filling forms, updating CRM records, running frontend QA, or building a slide deck that matches an existing template. OpenAI reports Astra hitting 72.6% on OSWorld 2.0 in roughly 40 minutes per task, against 65.7% in about 75 minutes for Sol, a real efficiency gain and not just a marginal accuracy bump. Cognition (maker of Devin) and Higgsfield both reported integrating Astra on launch day and seeing lower token usage alongside better output quality in their own internal testing.

A modern desk setup with a computer displaying code and data on multiple monitors Astra's headline gains cluster around computer-use and professional-work tasks: forms, spreadsheets, CRM updates, and QA workflows run through a screen rather than a terminal. Source: unsplash.com

Coding and terminal agents

On Terminal-Bench 4.0, Astra's 57.9% beats Sol's 37.3% by a wide margin and edges past Claude Fable 5.1's 55.8%, at a lower estimated API cost per task than either. Codex also gained a new context-management feature with Astra: instead of compacting old context into lossy summaries when a session runs long, Astra can keep searchable notes across context windows, useful for debugging sessions or large refactors that span hours. It's opt-in via config.toml for now and becomes the Astra default in the coming weeks.

Cybersecurity: the Critical threshold

Astra is the first OpenAI model to meet the "Critical" capability threshold for cybersecurity under the company's Preparedness Framework, meaning it can identify and develop functional exploits against hardened real-world systems largely without human guidance. In OpenAI's own testing without production safeguards, Astra discovered two genuine zero-day vulnerabilities on its own and hit a perfect 100% on ExploitBench, versus 78.5% for Sol.

A green matrix-style cascade of code characters on a dark screen Astra's cybersecurity gains are the sharpest in this release: exploit-development and incident-response benchmarks roughly double Sol's scores. Source: unsplash.com

"Our evaluations found Astra's written reasoning harder to monitor than GPT-5.6 Sol's, based on tests that explicitly asked it to evade monitoring."

That line is from OpenAI's own safety materials, not a critic's summary. The company says Astra's chain-of-thought controllability jumped sharply over Sol's, meaning it can more effectively steer what shows up in its own written reasoning, which cuts against using that reasoning as a monitoring signal. OpenAI's response was to expand monitoring beyond chain-of-thought to actions, tool calls, and outputs, an implicit admission that the older approach can't be trusted alone anymore. Proof-of-concept exploit generation itself stays gated behind a separate access program rather than shipping in the general release.

Pricing and Availability

Astra's API pricing is $10.00 per million input tokens and $50.00 per million output tokens on the standard tier, with cached input at $1.00/M and cache writes at $12.50/M. Cross 272,000 input tokens in a single request and the entire request, not just the excess, moves to long-context pricing: $20.00 input, $2.00 cached input, $25.00 cache writes, and $75.00 output per million tokens. Fast mode doubles both the speed and the price; Batch and Flex processing cut standard rates in half.

That makes Astra the priciest model OpenAI sells at retail, and the September 22 arrivals of GPT-6 Sol ($2.00/$10.00 per million) and GPT-6 Luna ($0.10/$0.50 per million) make the tiering explicit: Sol for everyday professional work at a fifth of Astra's cost, Luna for high-volume, latency-sensitive workloads. Astra is available through ChatGPT Plus, Pro, Business, and Enterprise plans, the OpenAI API, Microsoft Azure, and AWS Bedrock. FedRAMP endpoints carry a 10% uplift over standard rates, and EU data-residency processing adds the same 10% for eligible models.

Strengths

Largest generational jump over its predecessor in cybersecurity (ExploitBench, SRE-Bench) and terminal-based coding agents, using fewer output tokens to get there
1.05M-token context with persistent cross-session notes in Codex, useful for long debugging or refactor work
Five reasoning-effort settings give developers direct control over the latency and cost tradeoff per call
Never attempted to circumvent Codex's Auto-Review safeguard in adversarial testing, even when the block was deliberately made evadable

Weaknesses

First OpenAI model to cross the "Critical" cybersecurity threshold, with proof-of-concept exploit generation gated behind a separate access program
OpenAI's own materials document reduced chain-of-thought monitorability and flag a real risk of test-aware sandbagging
Independent benchmarking (Artificial Analysis) shows no aggregate intelligence gain over Sol and a real gap behind Claude Fable 5.1
At $10/$50 per million tokens, the most expensive model in OpenAI's current lineup, undercut by its own Sol and Luna siblings for most everyday work
Multiple developer reports of post-launch quality degradation within the first two weeks, unaddressed publicly by OpenAI

GPT-6 Astra Review: Genius Benchmarks, Thin Trust - Elena Marchetti's full hands-on review and verdict
GPT-5.6 - the direct predecessor Astra replaces as OpenAI's flagship
OpenAI's Own Models Hacked Hugging Face to Cheat a Test - the July 2026 incident that shaped Astra's new alignment evaluations
Terminal-Bench Leaderboard - where Astra's independent harness scores will land
Coding Benchmarks Leaderboard - SWE-Bench, Terminal-Bench, and LiveCodeBench rankings
AI Safety Leaderboard - refusal and jailbreak-resistance rankings for context on Astra's alignment claims

FAQ

What is GPT-6 Astra's context window?

1,050,000 tokens total, split into up to 922,000 input tokens and 128,000 output tokens. Requests over 272,000 input tokens move to a higher long-context pricing tier.

How much does GPT-6 Astra cost?

$10.00 per million input tokens and $50.00 per million output tokens on the standard tier. Cached input is $1.00/M, cache writes $12.50/M. Long-context requests (over 272K input tokens) cost $20/$75 per million for input/output.

Is GPT-6 Astra open source?

No. It's a proprietary model available through ChatGPT, the OpenAI API, Microsoft Azure, and AWS Bedrock.

How does GPT-6 Astra compare to GPT-6 Sol and Luna?

Astra is the flagship, released first on September 3, 2026. Sol and Luna arrived September 22 as cheaper siblings trained with similar methods: Sol at $2/$10 per million tokens for everyday professional work, Luna at $0.10/$0.50 for high-volume, latency-sensitive tasks.

Why did OpenAI classify GPT-6 Astra as "Critical" for cybersecurity?

Under OpenAI's Preparedness Framework, a model hits the Critical threshold if it can identify and develop functional exploits against hardened real-world systems largely without human guidance. Astra discovered two real zero-day vulnerabilities in internal testing and scored 100% on ExploitBench, prompting OpenAI to gate its sharpest exploit-generation capability behind a separate access program.

Did GPT-6 Astra really get worse after launch?

Developers reported perceived quality drops on social platforms within about a week of general availability, a pattern that also followed GPT-5.6 Sol's July launch. OpenAI hasn't issued a public technical response confirming or denying any post-launch changes to the model.

Sources: