GPT-6 Astra review, pricing and limits

GPT-6 Astra launched September 3, 2026 with a 1.05 million token context window, more than five times Sol's 200K, and a 128,000-token output limit. It targets agentic coding and computer-use workflows rather than everyday chat, where OpenAI's own benchmarks show it barely ahead of Sol.

GPT-6 Astra is OpenAI's flagship reasoning model, released September 3, 2026, scoring 96.0% on GPQA Diamond and 97.6% on FrontierMath Tier 4 v2. It is the first OpenAI model to reach the Preparedness Framework's Critical cybersecurity threshold, gating its exploit-creation capability behind an invite-only access program.

Provider: OpenAI · Family: GPT-6

More about OpenAI on HokAI

Context window: 1,050,000 tokens · Max output: 128,000

Input modalities: text, image · Output: text

About GPT-6 Astra

GPT-6 Astra is OpenAI's flagship large language model, released in a staged rollout starting September 3, 2026. It succeeds GPT-5.6 Sol, which launched two months earlier as part of the Sol, Terra, and Luna family. OpenAI has not disclosed Astra's architecture or parameter count, describing it only as trained on its largest run to date at the company's Stargate site in Texas. OpenAI president Greg Brockman called the release a generational leap and said he personally believes it may mark the arrival of artificial general intelligence, a claim that is unverified and sits oddly next to OpenAI's own benchmark disclosures, which paint a more mixed picture on general-purpose tasks (see the benchmark FAQ below). On the reasoning and knowledge evaluations OpenAI chose to publish, Astra leads Sol; on a broad open-ended knowledge exam it trails a leading rival model. Its headline score on a general-intelligence puzzle benchmark holds only inside a persistent test configuration that carries memory across calls, not in an ordinary single-shot API request (see Quirks). OpenAI skipped several of the benchmarks the industry typically uses to compare frontier models at launch, an unusual gap for a flagship release that leaves direct comparison to Claude and Gemini incomplete; the exact scores that are and are not published are broken out in the benchmark FAQ rather than repeated here. The context window is Astra's biggest architectural jump over Sol, large enough that OpenAI added a separate pricing tier once a single request crosses a fixed token threshold. Long-context recall stays strong deep into the window on OpenAI's own retrieval test. That size is also a trap: stuffing the whole window into one call is dramatically more expensive than retrieving just the relevant slice through a smaller pipeline, a cost dynamic covered in Quirks and the cost examples rather than restated here with exact figures. The model reads text and images but writes text only; it cannot generate or edit images in the same call. It carries a wider tool belt than Sol, including a hosted code-execution environment, direct patch-based code editing, standardized external-tool connectivity, and the ability to pick the right tool from a large registry on its own, alongside computer-use control that operates a desktop environment directly. It improves on Sol across the computer-use and browser-automation evaluations OpenAI published, with exact scores broken out in Key Features and Pros. Astra does not support fine-tuning. Astra costs meaningfully more per token than Sol at launch, under a tiered structure covering standard use, high-context requests, batch processing, and a faster paid mode. The exact rates, thresholds, and worked cost examples are covered in the pricing FAQ and pricing table, not repeated here. Astra is a closed, hosted model with no released weights. At launch, access went first to select organizations enrolled in OpenAI's Trusted Access and Daybreak programs, with ChatGPT Plus, Pro, Business, and Enterprise access rolling out afterward and turned off by default for administrators to enable. OpenAI has confirmed API access alongside deployment on AWS Bedrock and Microsoft Azure, with broader third-party platform support expected as the rollout widens. Wikipedia's coverage of the launch describes it as a limited preview at release, with a fuller public release expected roughly a week later. Astra is OpenAI's first model rated at the Preparedness Framework's Critical level for cybersecurity, capable of finding previously unknown vulnerabilities and building working exploits across hardened systems with little step-by-step guidance from a human operator. During testing it discovered and chained two previously unreported vulnerabilities into a working exploit. Full exploit-creation capability stays gated behind the invite-only Daybreak program. OpenAI reports a meaningful drop in high-severity misalignment flags versus Sol, cutting the indirect prompt-injection success rate to 8.5% (versus 27.0% for Sol). Independent evaluation by the UK AI Security Institute found Astra's chain-of-thought reasoning harder to monitor for evasive behavior under adversarial prompting than prior models, a regression OpenAI itself flags as a research priority. Astra fits teams building agentic coding tools, computer-use automation, or authorized offensive security research, where it leads Sol by a wide margin on OpenAI's own published evaluations. It is a weaker pick for everyday chat or cost-sensitive work: its general-purpose intelligence score sits flat against Sol's own, despite launching at a substantially higher effective price at maximum reasoning effort (see Cons). Teams that need a fully disclosed, industry-standard benchmark suite to compare against Claude or Gemini before committing will not find one in OpenAI's launch materials. Astra's training data runs through late April 2026. By policy, prompts and completions sent to Astra through the API do not feed back into OpenAI's training data by default, and qualifying enterprise accounts can opt into zero-retention handling that discards each request right after it completes, with no OpenAI staff review; the company says it is also testing a Private Safety Processing mode specific to Astra. The model operates under OpenAI's standard usage policies, which prohibit unauthorized offensive cyber activity and CBRN uplift content regardless of a user's access tier.

Pricing

$10 per 1M input tokens, $50 per 1M output. Cached input reads bill at $1 per 1M, cache writes at $12.50 per 1M. Requests over 272K input tokens bill the entire request at 2x input/cache and 1.5x output. Batch/Flex processing halves standard rates; Fast mode runs up to 2.5x faster at 2x the price. Roughly double GPT-5.6 Sol's $5/$30 launch pricing.

Key Features

• 1.05M-Token Context Window: Reaches 1.05M tokens, over 5x Sol's context, with 96.3% recall verified across the full window on OpenAI's own retrieval test.
• Reasoning Effort Tiers: Five effort levels (low, medium, high, xhigh, max) trade latency and cost for reasoning depth; the API defaults to low.
• Native Computer Use: Controls a desktop environment directly, scoring 92.7% on ScreenSpot-Pro and 72.6% on OSWorld 2.0, both ahead of Sol on the same evaluations.
• Critical-Tier Cybersecurity Capability: The first OpenAI model to reach the Critical threshold under the Preparedness Framework, scoring 100% on ExploitBench; full exploit-creation access is gated behind the invite-only Daybreak program.
• Extended Tool Use: Ships with a hosted execution shell, an apply-patch tool for direct code edits, native MCP support, and a tool-search capability for large tool registries.

Pros

• Posts a 99.9% ARC-AGI-3 score under OpenAI's disclosed evaluation and leads Sol on graduate-level science and advanced math reasoning (exact scores in the benchmark FAQ).
• On Artificial Analysis's coding-specific index, Astra matches Claude Opus 5 and Fable 5 while costing under half as much per task.
• Indirect prompt-injection attack success rate fell to 8.5% from 27.0% on Sol, per OpenAI's own system card.
• Context window is over 5x larger than Sol's, holding accurate recall deep into the full span (see Key Features).

Cons

• Its 61 score on Artificial Analysis's Intelligence Index is flat against Sol and trails Claude Fable 5.1's 65.7 by five points, despite launching at roughly 75% higher effective price per task at max reasoning effort.
• The marketed 99.9% ARC-AGI-3 score only holds in a persistent test configuration; stateless single-call usage scores roughly 17% to 63% on the same test.
• No SWE-bench Verified, AIME 2025, MMLU-Pro, or HumanEval scores were disclosed at launch.
• Closed weights, no fine-tuning support, and exploit-creation capability gated behind the invite-only Daybreak program.

Benchmarks

• arc agi 2: 95
• arc agi 3: 99.9
• browsecomp: 91.5
• osworld v2: 72.6
• deepswe v1 1: 74.1
• exploitbench: 100
• gpqa diamond: 96
• screenspot pro: 92.7
• terminal bench 4 0: 57.7
• humanitys last exam: 57.2
• benchcad vision2code: 95.9
• frontier math tier4 v2: 97.6
• artificial analysis intelligence index: 61
• artificial analysis price blended per m: 7.7

Visit GPT-6 Astra Official Page