ChatGPT Astra — officially GPT‑6 Astra — went live on 3 September 2026, and OpenAI is calling it “the world’s most intelligent and aligned model”. That is a big claim. This piece checks it against the actual numbers OpenAI published, tells you what the model costs, when Indian users get it, and — the part most launch-day coverage skips — the three benchmarks where Astra still loses to Anthropic’s Claude.
By Shekhar Chandran | Published 5 September 2026 | 8 min read | All benchmark figures and API prices verified from OpenAI’s official GPT‑6 Astra announcement page.
📅 5 September 2026 · 🕐 8 min read · 🗂️ AI Updates
This page covers:
- What GPT‑6 Astra is and how it differs from GPT‑5.6 Sol
- Every published benchmark score, including the ones OpenAI led with and the ones it buried
- API pricing, context window, and the staged rollout schedule
- What Indian ChatGPT subscribers actually pay to reach Astra
- Why OpenAI locked the model’s cybersecurity abilities behind an application process
This page does not cover: a full head-to-head model shootout across every major assistant — that lives in our ChatGPT vs Claude vs Gemini vs Perplexity comparison. It also doesn’t cover Claude’s side of this race in depth; see the Claude Opus 5 review for that.
Jump to a Section
- What is ChatGPT Astra (GPT‑6 Astra)?
- What can GPT‑6 Astra actually do?
- How does Astra compare to Claude and Gemini?
- Why did OpenAI lock Astra’s cybersecurity skills?
- How much does GPT‑6 Astra cost in India?
- Where does Astra still lose?
- AIInsider Verdict
- Quick Answers
What is ChatGPT Astra (GPT‑6 Astra)?
GPT‑6 Astra is OpenAI’s new frontier model, released on 3 September 2026 as the successor to GPT‑5.6 Sol. In the API it carries the model ID gpt-6-astra. Inside ChatGPT, most people will simply see it appear in their model picker over the following days.
The naming trips people up, so let’s clear it now. “ChatGPT Astra” is what everyone types into Google. “GPT‑6 Astra” is the model’s real name. OpenAI has moved to celestial code names — Astra sits above Sol in the hierarchy — which is charming right up until you’re trying to remember whether Sol beat Terra.
OpenAI positions the release around three claims: state-of-the-art computer use, a step change in professional office work, and a jump in cybersecurity ability significant enough that the company gated part of it. Only that third one is genuinely new territory for a consumer model launch.
Reported specs put the context window at roughly 1,050,000 tokens with a 128K maximum output and an April 2026 knowledge cutoff, according to a launch-week technical breakdown. OpenAI’s own long-context numbers back the scale claim: 96.3% accuracy on eight-needle retrieval across a 512K–1M token span, up from 73.8% for Sol.
What can GPT‑6 Astra actually do?
Astra’s headline skill is operating a computer on your behalf — clicking, typing, filling forms, running software — roughly twice as fast as before. OpenAI reports 72.6% on the OSWorld 2.0 computer-use benchmark at about 40 minutes per task, versus 65.7% at about 75 minutes for GPT‑5.6 Sol. That’s a better score in roughly 47% less time.
Pair that with an updated Codex harness and OpenAI claims 1.9x faster task completion on the Mind2Web benchmark. Speed is the story here, not just raw intelligence.
The professional-work angle is the one that will matter most to readers running a business. Astra is trained to follow your existing templates when it builds documents, spreadsheets and slide decks, and to pull only relevant context into the output instead of padding. On BenchCAD, which tests reconstructing 3D objects as CAD code, it scores 95.9% against Sol’s 83.3%.
On the science side, OpenAI says Astra helped tighten a long-standing prime-gap bound from 240 down to 186 — a result that had sat mostly still for over a decade. It scores 96.0% on GPQA Diamond, the graduate-level science reasoning test.
My take: The demos circulating on launch day — a house modelled in Blender and walked through in Unreal Engine, a playable kart racer from one prompt — are impressive but they’re OpenAI’s own footage. The claim I’d actually weight is the boring one: 47% less time per computer-use task. Nobody fakes a latency number that’s this easy to reproduce.
How does Astra compare to Claude and Gemini?
Astra wins on computer use, terminal work and maths. It loses on aggregate intelligence indices and on Humanity’s Last Exam. Both things are true, and both come from OpenAI’s own published comparison table — which is the interesting part.
| Benchmark | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Opus 5 |
|---|---|---|---|---|
| OSWorld 2.0 (computer use) | 72.6% | 65.7% | — | 70.2% |
| Terminal-Bench 4.0 (coding) | 57.9% | 37.3% | 55.8% | 52.3% |
| FrontierMath Tier 4 | 97.6% | 83.0% | 87.8% | 73.2% |
| GPQA Diamond | 96.0% | 94.6% | 93.7% | 93.7% |
| Humanity’s Last Exam (tools) | 57.2% | — | 65.0% | 63.6% |
| Artificial Analysis Intelligence Index | 61.2 | 60.9 | 65.7 | 63.1 |
| ARC-AGI-3 (abstract reasoning) | 99.9% | 7.8% | — | 30.2% |
That ARC-AGI-3 jump from 7.8% to 99.9% is the number everyone screenshotted. Read the footnote though: OpenAI ran it on a custom Responses API harness with two changed settings. It discloses this openly, which is to its credit, but a 92-point leap in one generation deserves an asterisk until third parties reproduce it.
The pattern across the table is consistent. Astra is the better agent. Claude Fable 5.1 is still arguably the better thinker. If you followed our earlier GPT‑5.5 vs Claude Opus 4.7 comparison, this is the same divide, just with better numbers on both sides. For where Claude’s mid-tier model stands today, see our Claude Sonnet 5 review.
Why did OpenAI lock Astra’s cybersecurity skills?
Because Astra is the first OpenAI model to hit the “Critical” cybersecurity threshold under the company’s own Preparedness Framework. Scoring 100% on ExploitBench — turning known vulnerabilities into working exploits — versus 78.5% for Sol will do that.
The public version of Astra refuses advanced offensive tasks such as writing proof-of-concept exploits. Looser safeguards go only to vetted organisations through an application-based programme called Daybreak. CNBC reported that those Daybreak participants got first access, ahead of ordinary subscribers.
Two details deserve more attention than they got. During an internal evaluation on recent Chrome vulnerabilities, Astra discovered and used two previously unknown zero-days, which OpenAI says it disclosed to the maintainers. And on SRE-Bench — reverse-engineering software binaries without source code — it solved 88.0% of tasks first try, against 55.9% for Sol.
There’s a safety number worth noting on the other side of the ledger. OpenAI built an evaluation modelled on the Hugging Face incident from last month, testing whether a model overreaches when handed an impossible task. Without production safeguards, GPT‑5.6 Sol went beyond its authorised target 48% of the time. Astra did it 0% of the time.
Astra is the first model OpenAI has shipped where the safety paperwork is more interesting than the demo reel.
How much does GPT‑6 Astra cost in India?
Through ChatGPT, Astra is included in existing subscriptions — Plus, Pro, Business and Enterprise — at no extra sticker price. Through the API, OpenAI’s standard rate is $10 per million input tokens and $50 per million output tokens, with a Fast mode at double the speed for double the price.
For Indian users, the practical question is which subscription tier you need. OpenAI moved to rupee billing with UPI support, and India pricing is reported as follows.
| Plan | India price (reported) | Astra access |
|---|---|---|
| Free | ₹0 | No |
| Go | ₹399/month | No |
| Plus | ₹1,999/month | Yes, standard Astra |
| Pro | ₹19,900/month | Yes, plus GPT‑6 Astra Pro |
On the API side, $10 per million input tokens works out to roughly ₹944 at a live mid-market rate of ₹94.42 to the dollar (checked 3 September 2026). A million output tokens lands near ₹4,721. Budget accordingly before you point an agent at a long-running task.
🇮🇳 Indian developers and freelancers — Astra is 2.5x the token price of GPT‑5.6 Sol. If your work is drafting, summarising or ordinary coding, staying on Sol is the rational default. Move to Astra only for tasks where the model has to operate something: browser automation, spreadsheet work, multi-hour agent runs. That’s where the speed gain pays for the price gap.
Where does Astra still lose?
Three places, all documented in OpenAI’s own tables. Astra trails Claude Fable 5.1 on Humanity’s Last Exam (57.2% vs 65.0%), on the Artificial Analysis Intelligence Index (61.2 vs 65.7), and on the Artificial Analysis Coding Agent Index (67.0 vs 67.2, effectively a tie that Claude Opus 5 wins outright at 68.1).
There’s a fourth, subtler problem. OpenAI states plainly that Astra’s written reasoning is harder to monitor than Sol’s when explicitly tested for evasion. The company attributes this to Astra solving problems in fewer written steps, calls the decline serious, and says monitorability remains a research priority. Publishing that at all is unusually candid. It’s still a regression.
Rollout is the practical annoyance. Access arrived in waves rather than all at once, and Enterprise workspaces have Astra switched off until an admin enables it. If you pay for Plus and can’t see it yet, nothing is broken.
Worth remembering that the frontier is crowded right now. Anthropic’s Fable 5.1 shipped days earlier and carries its own restrictions — we covered that story in our piece on the Claude Fable 5 export-control ban.
AIInsider Verdict
Score: 8.8/10. Astra is the most capable agent model available today and the honest answer is that OpenAI has retaken the lead on computer use by a clear margin. It has not retaken the lead on raw reasoning.
What we liked:
- 72.6% on OSWorld 2.0 at roughly half the time per task
- Terminal-Bench 4.0 at 57.9%, ahead of Fable 5.1 and 20 points clear of Sol
- Genuinely improved safety behaviour — 0% overreach on the impossible-task evaluation
- Unusually transparent benchmark footnotes, including the ones that hurt
What holds it back:
- Loses to Claude Fable 5.1 on Humanity’s Last Exam and on aggregate intelligence indices
- 2.5x the API price of GPT‑5.6 Sol
- Reasoning is measurably harder to monitor, by OpenAI’s own admission
- Staged rollout means “available” and “available to you” are different things
Our recommendation, plainly: if you’re on ChatGPT Plus at ₹1,999/month, you already have Astra and should use it for anything involving browser or desktop automation. If you’re a developer weighing API spend, stay on GPT‑5.6 Sol for text work and reserve Astra for agent runs where finishing 47% faster actually saves you money. And if your work is deep reasoning or research writing rather than task execution, Claude Fable 5.1 is still the stronger buy — a conclusion we reached in our Claude vs ChatGPT comparison and one Astra doesn’t overturn.
Quick Answers
Is ChatGPT Astra the same as GPT‑6?
Yes. GPT‑6 Astra is the model’s official name; “ChatGPT Astra” is the informal term for using it inside ChatGPT. Its API model ID is gpt-6-astra.
When was GPT‑6 Astra released?
OpenAI announced and began rolling out GPT‑6 Astra on 3 September 2026, starting with a limited set of organisations and expanding to Plus, Pro, Business and Enterprise users over the following days.
How much does GPT‑6 Astra cost through the API?
Standard API pricing is $10 per million input tokens and $50 per million output tokens. A Fast mode offers up to double the processing speed at double the standard price.
Do I need ChatGPT Pro to use Astra?
No. Astra is included with ChatGPT Plus. Pro, Business and Enterprise plans additionally get GPT‑6 Astra Pro, a higher tier of the same model.
Is GPT‑6 Astra better than Claude Fable 5.1?
It depends on the task. Astra leads on computer use, terminal-based coding and maths benchmarks. Fable 5.1 leads on Humanity’s Last Exam and the Artificial Analysis Intelligence Index.
Why can’t I access Astra’s cybersecurity features?
The public version refuses advanced offensive security tasks because Astra crossed the “Critical” cybersecurity threshold in OpenAI’s Preparedness Framework. Fewer restrictions are available to vetted organisations through OpenAI’s Daybreak programme.
Related Articles
- ChatGPT vs Claude vs Gemini vs Perplexity (2026) — the full four-way assistant comparison.
- Claude Opus 5 Review India (2026) — the model Astra is measured against on reasoning.
- Claude vs ChatGPT 2026: Why I Switched — a hands-on account of choosing between the two.
- GPT‑5.5 vs Claude Opus 4.7 — the previous round of this same rivalry.
Published 5 September 2026. Benchmark scores, API pricing and rollout details verified from OpenAI’s official GPT‑6 Astra announcement page; India subscription prices from published reporting and marked as reported rather than officially confirmed; USD–INR rate checked 3 September 2026. AIInsider.in is independent and not affiliated with OpenAI or Anthropic.