GPT-6 Astra

Overview

GPT-6 Astra is OpenAI's flagship model, opened as a limited preview to trusted partners on September 3, 2026 and released publicly to paid users on September 4, 2026. It replaces GPT-5.6 Sol at the top of the lineup and collapses that release's three-tier Sol/Terra/Luna structure back into a single frontier model.

OpenAI positions it as the model for demanding end-to-end work — advanced analysis, software engineering, deep research, scientific work and document creation — with the emphasis on long-horizon agentic tasks involving computer and browser use. That last part is the product change: Astra ships with computer use as a first-class capability rather than a research preview, which means driving browsers, spreadsheets and desktop applications through the same interfaces a person uses instead of a bespoke API integration per application.

The training run is the largest OpenAI has done. VP of research Aidan Clark described it as "by far" the company's biggest, and the first time OpenAI pretrained on more than 100,000 GPUs, at the Stargate site in Texas. Company president Greg Brockman framed the launch as the arrival of the "AGI era" — a marketing claim, and one the benchmark authors quoted at launch declined to endorse.

Two things about this release should shape how you evaluate it. The first is a pricing cliff at 272K input tokens that is easy to fall off (see Pricing & Access). The second is that Astra reasons using a technique called recurrent depth, which makes its internal reasoning materially harder to inspect than GPT-5.6's — OpenAI's own evaluations say so. For a full write-up of the launch and the benchmark disputes, see our launch coverage.

Capabilities

• Computer and browser use: Native operation of browsers, spreadsheets and desktop applications through human-facing interfaces. This is the capability OpenAI leads with, and the one that changes what an AI agent can be pointed at.
• Long-horizon agentic work: Tasks that run for many steps against real systems, rather than single-turn question answering. Positioned as the primary use case, not a side benefit.
• Recurrent depth reasoning: The model loops the same transformer layers over a hidden internal state before producing output tokens, spending compute in latent space instead of in visible chain-of-thought text.
• Long-context ingestion: 1,050,000 tokens total, up to 922,000 of them input — a large repository, a document set, or a long agent trajectory in a single request.
• Vision input: Images alongside text, for charts, scans, screenshots and UI captures.
• Hosted tools: Web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, tool search and MCP servers.
• Structured work: Streaming, function calling, structured outputs and prompt caching.

Technical Specifications

• API model ID: gpt-6-astra
• Context window: 1,050,000 tokens
• Maximum input: 922,000 tokens
• Max output: 128,000 tokens (reasoning tokens count toward this)
• Knowledge cutoff: April 30, 2026
• Input modalities: Text, images
• Output modalities: Text
• Audio / video: Not supported
• APIs: Chat Completions, Responses, Batch. Realtime is not supported — and this is not a regression: gpt-5.6-sol and gpt-5.6-luna mark Realtime "Not supported" too. OpenAI serves v1/realtime from a separate gpt-realtime-* family, which the flagship line has never been part of.
• Features: Streaming, function calling, structured outputs, prompt caching, file search, image input, web search
• Fine-tuning: Not offered
• Rate limits (standard, Tier 1): 500 requests per minute, 500,000 tokens per minute
• Long-prompt billing threshold: 272,000 input tokens

The GPT-5.6 Family (Previous Generation)

GPT-5.6 is not retired. All three tiers remain listed and callable, and OpenAI's deprecations page still names them as the recommended replacements for older models being shut down — GPT-6 Astra does not appear there at all. Their prices were cut at the Astra launch, and OpenAI's model reference notes that the reduced rates are promotional pricing available at least through November 21, 2026, so treat them as provisional.

Sol, Terra and Luna share the same 1,050,000-token context window, the same 128,000-token output cap and the same February 16, 2026 cutoff; what separates them is speed, reasoning depth and price. The bare gpt-5.6 alias routes to Sol. Note that GPT-5.5 is now more expensive than Sol on both input and output while carrying a cutoff eleven weeks earlier — there is no reason to start new work on it.

At $0.20 / $1.20 per million tokens, Luna keeps the full million-token context and is now the cheapest sensible default for high-volume extraction. Astra costs 50x its input rate; that spread, not the benchmark table, should decide routing.

Use Cases

• Computer-use agents: Workflows that touch applications with no usable API — legacy internal tools, vendor portals, spreadsheet-driven processes. This is the reason to pick Astra over Sol.
• Long-horizon software engineering: Multi-hour work in live repositories where the model must plan, edit, run and re-plan. Astra's DeepSWE result is the benchmark with the most headroom left, and the one most predictive of real-repo performance.
• Deep research and scientific work: Extended literature and data synthesis where the 922,000-token input budget removes the retrieval step for a bounded corpus.
• Whole-codebase analysis: Ingesting a repository and its history in one request — but watch the 272K threshold, which most repositories of interest will cross.
• Tiered agent architectures: GPT-5.6 Luna or Terra as subagents doing extraction, classification and routing; Astra reserved for the steps where a wrong answer is expensive. At a 50x input price ratio, this is the deployment shape that makes Astra affordable.
• Document and screenshot understanding: Image input for charts, scans and UI captures inside an agent loop.
• Defensive security research: Astra is the strongest published model on offensive-security evaluations, but the capability is gated — see Limitations.

Performance / Benchmarks

Figures OpenAI reported at launch, vendor-run at the time of publication unless noted:

Most of these numbers no longer discriminate between frontier models. On GPQA Diamond, Claude Opus 5 sits at 94.2% and Gemini 3.1 Pro at 94.3%; 96% is a rounding error on a converged benchmark. FrontierMath Tier 4 is the exception — GPT-5.5 scored 35.4% on the earlier cut — so 97.6% is a genuine discontinuity, with the caveat that Astra ran the v2 revision rather than the cut those models took.

DeepSWE v1.1 at 74.1% is the honest number. It is nowhere near saturation and its tasks come from live repositories, which makes it the score that predicts whether the model helps on your codebase.

The ARC-AGI-3 result needs its asterisk. ARC Prize ran its own evaluation on the Semi-Private set on September 3, 2026 and published two figures: 62.7% using its standard harness — the configuration every provider runs identically, at roughly $26,000 of compute — and 99.9% using a provider adapter that preserves OpenAI's opaque reasoning state between requests. Same model, same tasks, a 37-point gap. The 98.6% in OpenAI's launch materials matches neither. ARC Prize also stated plainly that it is not claiming Astra is AGI, and that saturating the benchmark would not constitute proof of it.

Third-party aggregate: Artificial Analysis places GPT-6 Astra (max reasoning effort) at 55 on the Artificial Analysis Intelligence Index v4.2, against 51 for GPT-5.6 Sol on the same index version, with measured output around 73 tokens per second. Index versions are not comparable across revisions; both figures here are v4.2.

GDPval is missing. OpenAI's own measure of performance on real occupational tasks does not appear in the launch materials. For a release marketed on economic generality, that is the gap worth noting.

Safety & Preparedness

Astra is the first OpenAI model to reach the "Critical" capability threshold for cybersecurity under the Preparedness Framework, which triggered additional deployment restrictions. OpenAI's stated concern is straightforward: a model that can find previously unknown security flaws and develop new ways to exploit them helps defenders and attackers with equal enthusiasm.

The published safeguards:

• The public model is a restricted configuration that refuses advanced offensive tasks, including generating proof-of-concept exploits.
• Advanced cybersecurity capability was limited to a small group of alpha testers at launch, with expanded access for vetted defenders routed through OpenAI's Daybreak and Daybreak Blue programs. These are access tiers, not separately named models — there is no public unrestricted model ID.
• Enterprise workspaces have Astra off by default; an administrator must enable it.
• Misalignment monitoring runs across all tool-using inference in the public deployment, expanded from internal-only use previously.

The system card also reports a decline in chain-of-thought monitorability relative to GPT-5.6 Sol: the model is better at controlling its own reasoning traces, and simply telling it that a monitor is present can cause it to shorten its explanations. OpenAI chief scientist Jakub Pachocki has said the company would withhold further scaling rather than accept more loss of monitoring confidence; outside researchers at Redwood Research have argued that pushing recurrent depth further moves reasoning out of observable channels altogether.

Limitations

• The 272K pricing cliff. Requests above 272,000 input tokens are billed at 2x input and cache rates and 1.5x output — for the entire request, not the overage. A 273K-token prompt costs roughly twice what a 271K one does. If you are near the line, trim or split.
• Reasoning you cannot inspect. Recurrent depth spends compute in latent space, so debugging a wrong answer with the reasoning trace is less useful than it was on GPT-5.6, and text-based safety monitors see less. This is a known, documented regression in monitorability, not a speculation.
• The public model refuses cybersecurity work. If your use case is security research, the model you can call is not the model in the benchmark table; you need Daybreak Blue access.
• No Realtime API. Astra is not on v1/realtime, and neither is GPT-5.6 — that endpoint is served by a separate gpt-realtime-* family. A voice or low-latency streaming architecture does not migrate to Astra; it stays on the realtime models, which trail the flagship line on context and knowledge cutoff.
• Text output only. No audio or video, as input or output.
• No fine-tuning.
• Price. 2.5x GPT-5.6 Sol's output rate and 50x Luna's input rate. OpenAI argues cost-per-task is the right metric; that is a claim you have to verify on your own workload, not a given.
• Cache writes cost 1.25x the uncached input rate, so caching a prefix used only once is a net loss — and above 272K those write rates double too.
• Reasoning tokens are billed as output and count against the 128,000-token output cap, so a high-effort request can spend most of its budget before any visible answer.
• Knowledge cutoff. Nothing after April 30, 2026 is in the weights.
• Benchmark headlines are harness-dependent. The ARC-AGI-3 spread between 62.7% and 99.9% is the clearest case; assume the same is possible elsewhere.

Pricing & Access

API pricing (per 1M tokens)

The long-prompt rule is the one to internalise. It is not a marginal rate on the tokens above 272,000 — the whole request reprices. Input and cache rates double, output goes to 1.5x. GPT-5.6 carries the same rule at the same threshold.

Other modifiers:

• Web search: $10.00 per 1,000 calls, plus search content tokens billed at model rates.
• Fast mode: launch coverage reports an API option running at roughly twice the throughput for twice the standard price (about $20 / $100 per million tokens). It is not broken out on OpenAI's main pricing table.

Availability

• ChatGPT: Plus, Pro, Business and Enterprise. No free tier. Enterprise workspaces must enable it explicitly — it is off by default. Pro, Business and Enterprise plans also get a higher-usage GPT-6 Astra Pro variant.
• OpenAI API: gpt-6-astra on Chat Completions, Responses and Batch.
• Clouds: AWS, alongside the direct API.
• Rollout order: a limited set of organizations on September 3–4, 2026, then the paid ChatGPT tiers over the following days.

Ecosystem & Tools

Official SDKs

• Python SDK
• Node SDK
• REST API via the Responses and Chat Completions endpoints

Hosted tools

Web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, tool search and Model Context Protocol servers are available without extra plumbing.

Products and platforms

• ChatGPT — the consumer and enterprise surface
• OpenAI Codex — the coding agent
• OpenAI Developer Platform — models, pricing and API reference
• OpenAI Cookbook — worked examples

Community & Resources

• GPT-6 Astra announcement - OpenAI's launch post
• GPT-6 Astra model reference - Context, pricing, modalities, endpoints, rate limits
• GPT-6 Astra system card - Preparedness evaluations and monitorability findings
• Path to Astra: critical capabilities and frontier safeguards - OpenAI on the Critical cybersecurity designation
• ARC Prize evaluation of GPT-6 Astra - The independent ARC-AGI-3 numbers and both harnesses
• GPT-5.6 Sol model reference - The previous flagship, still callable
• API pricing
• Model deprecations
• Developer forum