GPT-6 Astra review, pricing and limits

GPT-6 Astra targets agentic coding and computer use, not everyday chat. On Artificial Analysis's current index it scores 53 with max reasoning, running at 51 tokens per second and costing $3.26 for each index task. OpenAI pitches its cheaper GPT-6.1 Sol as close in ability, so weigh the premium carefully.

GPT-6 Astra is OpenAI's flagship reasoning model, released September 3, 2026, with a 1.05M-token context window and a 128,000-token output limit. It scores 96.0% on GPQA Diamond and is the first OpenAI model rated Critical for cybersecurity. Since September 29 it also runs in an Ultrafast tier that OpenAI says is up to 6x faster in the API.

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 GPT-5.6 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 GPT-5.6 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; OpenAI's model page lists audio and video as unsupported, though the Responses API image-generation tool is listed as supported alongside web search, file search, a hosted shell, apply-patch code editing, skills, MCP, tool search and computer use. It improves on GPT-5.6 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 lists at five times the per-token rate of GPT-6.1 Sol, and OpenAI describes that cheaper model as delivering near-Astra intelligence at a fifth of the price (a vendor claim, not an independent result). Batch, Flex, Fast mode and, since September 29, 2026, an Ultrafast tier sit on top; the exact rates and worked cost examples are in the pricing FAQ and pricing table. OpenAI says Ultrafast generates up to 6x faster in the API and up to 8x (about 300 tokens per second) in OpenAI Codex, where it comes with the Pro 500 and Enterprise plans of ChatGPT Work. API access is open to all users at low default rate limits, from 500,000 tokens per minute on usage tiers 1 to 3 up to 5 million on tier 5, and Ultrafast supports only US data residency and global processing.

Astra is a closed, hosted model with no released weights. It began as a staged rollout to organizations in OpenAI's Trusted Access and Daybreak programs, and OpenAI's model documentation now lists standard rate limits for API usage tiers 1 to 5 (the free tier is not supported), which is why HokAI now lists it as generally available. It is reachable through the OpenAI API, ChatGPT plans (off by default for administrators at launch) and Codex, and OpenAI confirmed deployment on AWS Bedrock and Microsoft Azure. For agent builds, the OpenAI Agents API and Bedrock Managed Agents are the two routes OpenAI announced at DevDay 2026.

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 GPT-5.6 Sol, cutting the indirect prompt-injection success rate to 8.5% (versus 27.0% for GPT-5.6 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 (Codex is the first-party surface; see also best AI coding assistants), computer-use automation, or authorized offensive security research through the Daybreak program, where OpenAI's own evaluations show it leading GPT-5.6 Sol; the defensive side of that work sits with Codex Security. It is a weaker pick for everyday chat or cost-sensitive work: Artificial Analysis calls it notably slow at 51 tokens per second at max effort and expensive at $3.26 per Intelligence Index task, and OpenAI's own GPT-6.1 Sol targets most of the same work for a fifth of the standard token price. Teams that need a fully disclosed, industry-standard benchmark suite to compare against Gemini 3.5 Pro or Claude before committing will not find one in OpenAI's launch materials. For a head-to-head read see Grok 4.7 vs Opus 5 vs Astra vs Gemini 3.8.

Astra's training data runs through late April 2026. By policy, API prompts and completions do not feed back into OpenAI's training data by default, and qualifying enterprise accounts can opt into Zero Data Retention. On September 29, 2026, OpenAI packaged Zero Data Retention with Private Safety Processing as Private Intelligence, which runs automated safety reviews without giving OpenAI personnel access to the content, and said a Private Inference preview built on confidential computing is due this fall. The model operates under OpenAI's usage policies, which prohibit unauthorized offensive cyber activity and CBRN uplift content regardless of a user's access tier.

Pricing

Standard is $10 per 1M input tokens and $50 per 1M output. Cached input reads bill at a tenth of the input rate and cache writes at 1.25x. Requests over 272K input tokens bill the whole request at 2x input/cache and 1.5x output. Batch and Flex are half of Standard and Fast mode is 2x. Ultrafast, live since DevDay 2026, is $60 input and $300 output ($120 and $450 above 272K). Regional processing adds 10%. For scale, [GPT-6.1 Sol](/hub/models/gpt-6.1-sol) lists at $2 and $10 and [GPT-6 Luna](/hub/models/gpt-6-luna) at $0.10 and $0.50. Checked on OpenAI's pricing page on 2026-09-30.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• 1.05M-Token Context Window: Reaches 1.05M tokens, over 5x its predecessor'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 its predecessor 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.
• Ultrafast Speed Tier: OpenAI's fastest service tier, added at DevDay 2026: generation runs up to 6x faster through the API and up to 8x inside Codex, at a premium over Fast mode.

Pros

• Posts a 99.9% ARC-AGI-3 score under OpenAI's disclosed evaluation and leads its predecessor on graduate-level science and advanced math reasoning (exact scores in the benchmark FAQ).
• Artificial Analysis rates it 53 on its current Intelligence Index at max effort and describes it as fairly concise, at 60M output tokens for the full suite.
• Indirect prompt-injection attack success rate fell to 8.5% from 27.0% on its predecessor, per OpenAI's own system card.
• Context window is over 5x larger than its predecessor's, holding accurate recall deep into the full span (see Key Features).

Cons

• At max reasoning effort Artificial Analysis measures 51 output tokens per second, and OpenAI's own cheaper Sol model is pitched as a near-equal at a much lower price.
• 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% vendor-reported · 30 Sep 2026 — Abstract visual puzzles built to resist memorisation, % solved.
• ARC-AGI 3: 99.9% vendor-reported · 30 Sep 2026 — Interactive abstract puzzles, the third ARC series, % solved.
• BrowseComp: 91.5 vendor-reported · 30 Sep 2026
• OSWorld v2: 72.6 vendor-reported · 30 Sep 2026
• DeepSWE v1.1: 74.1 vendor-reported · 30 Sep 2026
• Exploitbench: 100 vendor-reported · 30 Sep 2026
• GPQA Diamond: 96% vendor-reported · 30 Sep 2026 — PhD-level science questions that are hard to search for, % correct.
• ScreenSpot Pro: 92.7 vendor-reported · 30 Sep 2026
• Terminal Bench 4.0: 57.7 vendor-reported · 30 Sep 2026
• Humanity's Last Exam: 57.2% vendor-reported · 30 Sep 2026 — Expert-written questions across many fields, % correct.
• Benchcad Vision2code: 95.9 vendor-reported · 30 Sep 2026
• Frontier Math Tier4 v2: 97.6 vendor-reported · 30 Sep 2026
• AA Intelligence Index: 53 cited: Artificial Analysis · 30 Sep 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
• Output speed: 51 tok/s cited: Artificial Analysis · 30 Sep 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit GPT-6 Astra Official Page