GPT-6 Astra Explained: Price, Benchmarks, Limits

OpenAI named Astra publicly on 1 August 2026, inside a research post about mathematics rather than a product launch. An internal build had closed ten long-standing open problems in mathematics and theoretical computer science — among them an explicit construction of a non-sofic group and a disproof of Connes's rigidity conjecture, with the proofs formally checked in Lean 4. A month later, on 3 September, the same family shipped as GPT-6 Astra.

That gap between the research framing and the product framing is the useful thing to hold onto. Astra is a genuinely different model at the top of the difficulty curve. It is not a uniform upgrade over what you were using last month, and for a lot of Malaysian teams it will be the wrong default.

1. What actually shipped

The model card is short and worth reading in full. The numbers that matter:

• Model ID: gpt-6-astra. One snapshot, no dated variants yet.
• Context window: 1,050,000 tokens — 922,000 of which can be input.
• Max output: 128,000 tokens per call.
• Knowledge cutoff: 30 April 2026.
• Pricing: US$10 per million input tokens, US$50 per million output. Cached input drops to US$1; cache writes cost US$12.50. Anything over 272K input tokens is billed at 2x input and 1.5x output.
• Reasoning effort: low, medium, high, xhigh, max.
• Tools: web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search.

Rollout was staged. A limited set of organisations got it on day one; Business and Pro subscribers followed within roughly a day, Plus users hours after that, then the API and AWS. Enterprise administrators control workspace access and the model launches disabled by default — a detail worth checking if your team says "we don't have Astra yet."

2. The headline benchmarks

OpenAI's own numbers are strong: 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, 100% on ExploitBench. On computer and browser use it scores 72.6% on OSWorld 2.0 while taking roughly 47% less time per task than GPT-5.6 Sol. Computer-use performance close to doubled, and the underlying optimisations gave existing Sol deployments something like a 60% speedup for free.

The cybersecurity result is the one that carries a policy tail. OpenAI designated Astra as crossing the "Critical" cybersecurity capability threshold under its Preparedness Framework — the highest severity level the company defines. That is not marketing. It changes what safeguards ship with the model and how enterprise access is gated.

3. Where the independent numbers disagree

Here is the part the launch coverage mostly skipped. Artificial Analysis scored Astra at 61 on its Intelligence Index — tied with GPT-5.6 Sol, and five points behind Claude Fable 5.1. On GDPval-AA v2, a benchmark adapted from OpenAI's own dataset of economically valuable tasks across 44 occupations, Astra dropped roughly 80 Elo points against its predecessor.

Token efficiency improved by about 10% at max effort. Price went up 2.5x. Net result on that index: roughly 75% more expensive per task than GPT-5.6 Sol, for no measured gain on generic knowledge work.

Both things are true at once. Astra is a step change on long-horizon agentic work, computer use, formal mathematics and security research. It is roughly flat on the kind of drafting, summarising and analysis that makes up most corporate AI usage. If your evaluation set looks like the second category, upgrading is a price increase with extra steps. We make the same argument at more length in our frontier AI models comparison.

4. The four things Astra changes in practice

Filtering out the noise, four capability shifts are real enough to plan around.

5. What it costs on a real workload

List price is misleading for agentic work because agents burn output tokens. One complex early-access build reported by testers consumed roughly 8 million tokens — about US$17 for a single task. Multiply that across a team of ten engineers running long agent sessions daily and you are into five-figure ringgit monthly spend before anyone has shipped a feature.

Three levers keep that sane. Prompt caching cuts repeated input from US$10 to US$1 per million, which matters enormously for agents that re-read the same codebase. Batch processing runs at half price for anything that does not need a live answer. And reasoning effort is a dial, not a switch — most tasks do not need max, and the gap between medium and max on routine work is small relative to the token bill.

The discipline that saves money is boring: measure cost per solved task, not cost per token. A cheaper model that fails 30% of the time is not cheaper.

6. Who should actually switch

Our read for Malaysian teams, stated plainly so you can argue with it:

• Switch now if you run long-horizon coding agents, browser automation, or security research. The computer-use and Codex improvements are the strongest part of this release, and the context-preservation change — searchable notes instead of lossy summarisation across context windows — is the first fix we have seen that meaningfully reduces agent drift on multi-hour tasks.
• Stay on GPT-5.6 Terra or Luna if your workload is document drafting, meeting summaries, customer replies, or retrieval-augmented Q&A. Terra sits at US$2/US$12 per million tokens. Astra is five times the input price for work it does not measurably do better.
• Run your own evaluation before deciding either way. Public benchmarks are proxies for someone else's job, not yours. We cover how to build a defensible internal eval set in our piece on LLM evaluation in production.

7. The governance question nobody is asking

A model classified "Critical" for cybersecurity capability, with hosted shell access, computer use and MCP tool calling enabled by default in the API, is a different governance object than a chat assistant. If you are in a BNM RMiT-regulated environment, or handling PDPA-covered personal data, the questions your risk committee should be asking changed on 3 September — specifically around what the model is permitted to execute, not just what it is permitted to read. Our AI governance guide for Malaysian organisations covers the control set; the short version is that agentic tool access needs its own approval path, separate from your existing LLM policy.

Astra is the most capable model OpenAI has shipped and simultaneously the easiest one to waste money on. Both statements come from the same source: it is optimised for the hardest 5% of tasks, and most organisations spend their token budget on the other 95%.