🌟 GPT‑6 Astra: The AI That Uses Your Computer

OpenAI's new flagship doesn't just answer you — it clicks, types, and finishes the job

Released 4 September 2026 · API gpt-6-astra · Read ~14 min · Source openai.com/index/gpt-6-astra

🎬 Note on the media below: every clip and screenshot is OpenAI's own demo footage, embedded straight from their servers. If your Markdown viewer strips HTML, use the ▶ Watch link under each video.

---

📌 The fast facts

🗓️ Released	4 September 2026 (limited preview 3 September)
🏷️ Replaces	GPT‑5.6 Sol
💰 Price (API)	$10 / 1M input tokens · $50 / 1M output · $1 / 1M cached input
🧠 Context window	~1.1M tokens in, 128K out (≈1,600 pages) — third-party trackers
👁️ Inputs	Text + images → text out
🛒 Where	ChatGPT Plus / Pro / Business / Enterprise · OpenAI API · Azure · AWS Bedrock
⚡ Fast mode	Up to 2× speed, at 2× price

---

🎬 First, just watch it work

Four clips. No explanation needed — this is the whole pitch.

🧾 It fills in your tax return

▶ Watch: filling in a US Form 1040

🔌 It lays out a circuit board

▶ Watch: PCB layout in KiCad

A 15-second condensed playback of Astra turning an electronic schematic into a manufacturable board — placing components and routing copper. This is normally slow, manual work in every electronics project.

📊 It builds a Power BI dashboard

▶ Watch: Power BI

🏆 It competes in Excel

▶ Watch: an Excel competition problem, in real time

---

🌟 The whole thing in one minute

• 🖱️ It drives a computer. Clicks, types, scrolls, reads the screen — forms, CRM records, calendars, testing a site it just built.
• ⚡ 1.9× faster task completion than the old model on a web-task benchmark; ~47% less time per task in one desktop simulation.
• 🧮 97.6% on FrontierMath Tier 4 — and it helped push a prime-number result that had been stuck for over a decade.
• 🔐 First model OpenAI rates "Critical" for cyber. Hence the slow, gated rollout.
• 🧠 The catch: OpenAI's own tests found its reasoning harder to monitor than the last model's. Safety researchers are alarmed.
• 💰 You may already have it — included in existing ChatGPT allowances.

---

🏆 Three numbers everyone is quoting

xychart-beta
    title "The three saturated benchmarks (%)"
    x-axis ["FrontierMath T4", "ARC-AGI-3", "ExploitBench"]
    y-axis "Score" 0 --> 100
    bar [97.6, 99.9, 100]
    bar [83.0, 7.8, 78.5]

Gold = GPT‑6 Astra · Second bar = GPT‑5.6 Sol, the model it replaces

Score	Benchmark	What it means in normal words
97.6% 🥇	FrontierMath Tier 4 (v2)	Hardest tier of a research-level maths test. Sol: 83.0%. (OpenAI's text rounds this to "98%"; its own table says 97.6%.)
99.9% 🧩	ARC-AGI-3	Puzzles it has never seen — pure "figure it out". Sol scored 7.8%. Not a typo. 😳
100% 🔓	ExploitBench	Turning known bugs into working exploits. Perfect score, up from 78.5%.

"The story is: end of one era, start of another."
— Greg Burnham, EpochAI, quoted by OpenAI

OpenAI president Greg Brockman suggested Astra could eventually be seen as the arrival of AGI. 🚨 Worth saying plainly: that's a claim, not a measurement, and many researchers disagree.

---

🖱️ The shift: from answering to doing

Every model before this was a very smart pen pal. You clicked; it talked. Astra's flagship skill is computer use — it sees a screen, moves a cursor, types, and checks whether what it did actually worked.

flowchart LR
    subgraph OLD["❌ BEFORE — you are the hands"]
        direction LR
        U1["🧑 You"] -->|"asks"| M1["🤖 Model<br/>text in, text out"]
        M1 -->|"advises"| U1
        U1 -->|"you click, type,<br/>copy, paste — every step"| A1["🖥️ Browser / app"]
        A1 -->|"you read the result"| U1
    end

    subgraph NEW["✅ WITH ASTRA — it is the hands"]
        direction LR
        U2["🧑 You"] -->|"one ask"| M2["🌟 Astra<br/>sees the screen"]
        M2 -->|"clicks & types"| A2["🖥️ Browser / app"]
        A2 -->|"reads result back"| M2
        M2 -->|"hands over"| R2["📦 Finished work<br/>deck · form · booking"]
        R2 -.->|"only if it matters"| U2
    end

The mechanism that changed: the loop used to run through you — every click was a human step. Astra closes the loop itself, so you go from operator to reviewer. That missing hop is the whole time saving.

📊 Computer-use benchmarks

xychart-beta
    title "Computer use — GPT-6 Astra vs the field (%)"
    x-axis ["Agents' Last Exam", "OSWorld 2.0", "ScreenSpot-Pro"]
    y-axis "Score" 0 --> 100
    bar [59.3, 72.6, 92.7]
    bar [53.6, 65.7, 76.9]

Gold = Astra · Second = GPT‑5.6 Sol

Benchmark	🌟 Astra	Sol	Opus 5	Bar (Astra)
Agents' Last Exam · pro tasks in real software	59.3%	53.6%	55.5%	████████████
OSWorld 2.0 · everyday desktop tasks	72.6%	65.7%	70.2%	███████████████
ScreenSpot-Pro · finding things on screen	92.7%	76.9%	—	███████████████████

⚡ Speed, not just accuracy

	Result
🏃 Mind2Web	1.9× faster task completion vs the current Sol experience (with the updated Codex harness)
⏱️ OSWorld 2.0	72.6% at ~40 min/task vs 65.7% at ~75 min — about 47% less time
🪙 Tokens	~65% fewer output tokens than Claude Opus 5 on Agents' Last Exam, while scoring higher

---

📅 The errands it runs for you

The click-heavy jobs that eat an afternoon. All four below are OpenAI demo runs — one finished in 2 min 54 sec. ⏱️

🩺 Finding a pediatrician

▶ Watch

🏠 Apartment hunting

▶ Watch

🚗 Booking a DMV appointment

▶ Watch

🥗 Building a low-carb shopping list

▶ Watch

🏫 Also demoed: comparing kindergartens

Area	What it takes off your plate
🧾 Admin	Form 1040 tax returns · online forms · CRM records · formatting legal documents to house style
🗓️ Errands	DMV bookings · doctors · apartments · schools · shopping lists
🔎 Research	Browses itself, then drafts the summary into your email or doc editor — not into a chat box
🎨 Making	Builds a website, then runs front-end QA on it to check the buttons work
🛠️ Support	Installs and tests software · troubleshoots what's on your screen — because it can see it
⚙️ Engineering	PCB layout in KiCad · CAD in FreeCAD · Blender → Unreal Engine

💡 The habit change: stop writing prompts, start assigning tasks.

---

🙋 It asks before it guesses

This screenshot is the clearest single image of the difference. Same request — "build me a personal career website". The old model worked for 13 minutes and shipped something. Astra stopped after 20 seconds to ask the one question that changes everything: what career are you moving into?

Astra fills routine gaps by itself, and asks only when the answer would change the outcome. In Codex it asks asynchronously — carrying on with everything that doesn't depend on your reply. If you never answer, it proceeds on sensible assumptions for small things and waits on the consequential ones. 🎯

Two more of the same demo:

It's also better at staying oriented. Older models treated a mid-task correction as a brand-new goal and dropped the original constraints. Astra folds the change in and keeps going. 🧭

---

💼 Slides, spreadsheets, documents

Astra is trained to follow your templates — your deck layout, your tone, your visual style — and to pull only what the task needs instead of restating everything it knows. Result: fewer outputs you have to reformat before sending. ✅

Benchmark	🌟 Astra	Sol	Fable 5.1	Opus 5	Bar (Astra)
AutomationBench · automating real workflows	41.4%	18.1%	31.4%	26.9%	████████
BenchCAD · 3D object → CAD code from pictures	95.9%	83.3%	84.3%	82.1%	███████████████████
BrowseComp · hard web research	91.5%	90.4%	—	90.8%	██████████████████
OpenScore String Quartets · reading sheet music	0.84	0.19	—	—	🎼

On BenchCAD, Astra's estimated API cost was ~43% below Sol and ~86% below Fable 5.1. OpenAI notes Claude's BenchCAD scores reflect three modifications described in Anthropic's own system card.

---

🎨 What it builds

🏡 Modelling a house in Blender

Astra models the house in Blender, then turns it into a walkable scene in Unreal Engine 5 — so designers and clients can experience the space before it's built. 🏗️

🖼️ …and renders the stills

Seven path-traced views from one model — exterior, living room, kitchen, office, bedroom, bathroom, terrace.

🎮 Games from a prompt

▶ Watch: a city scene, built and playable

Non-technical people can now build and play custom games in minutes — real graphics, real motion, not stick figures. (Credit: Pietro Schirano.)

⚙️ Engineering CAD

A five-speed car transmission in FreeCAD — with the gears actually meshing:

▶ Watch: the gear train in motion

---

💻 For people who write code

OpenAI calls Astra its best software-engineering model yet. The tables tell a more honest story: huge gap over its own predecessor, narrow gap over Claude.

xychart-beta
    title "Coding benchmarks (%)"
    x-axis ["Terminal-Bench 4.0", "DeepSWE v1.1", "FrontierCode Ext.", "DB migration"]
    y-axis "Score" 0 --> 100
    bar [57.9, 74.1, 64.5, 63.9]
    bar [37.3, 72.7, 60.6, 42.7]

Gold = Astra · Second = GPT‑5.6 Sol

Benchmark	🌟 Astra	Sol	Fable 5.1	Opus 5	Bar (Astra)
Terminal-Bench 4.0	57.9%	37.3%	55.8%	52.6%	████████████
DeepSWE v1.1	74.1%	72.7%	67.4%	73.7%	███████████████
FrontierCode 1.1 Extended	64.5%	60.6%	63.6%	63.6%	█████████████
Internal DB migration	63.9%	42.7%	57.8%	—	█████████████

On Terminal-Bench, Astra cost roughly **9% less* per task than Sol and 63% less than Fable 5.1.*

🧵 The quieter upgrade: it stops forgetting

Long sessions have always failed the same way. The context window fills, the model compacts — squashing everything into one summary — and details vanish. Why a fix failed. How a component behaves. The constraint you gave an hour ago.

In Codex, Astra keeps notes across context windows, and earlier windows stay searchable.

flowchart LR
    subgraph C["❌ Compaction — the old way"]
        direction LR
        W1["window 1"] --> S["📄 one summary"]
        W2["window 2"] --> S
        W3["window 3"] --> S
        S --> K1["keeps working"]
        S -.->|"detail dropped here<br/>is gone for good"| X["🗑️ lost"]
    end

    subgraph N["✅ Astra in Codex — notes + recall"]
        direction LR
        V1["window 1"] --> NT["🗒️ running notes"]
        V2["window 2"] --> NT
        V3["window 3"] --> NT
        NT --> K2["keeps working"]
        K2 -.->|"searches earlier<br/>windows on demand"| V2
    end

Compaction is lossy and one-way. Astra writes notes and keeps old windows searchable — so a forgotten requirement becomes a lookup, not a loss. Experimental flag in your Codex config.toml today; default in the coming weeks.

🧵 Memory, measured. On long-context retrieval (MRCR v2, 8 needles) Astra held 100% from 256K–512K tokens and 96.3% from 512K–1M — vs 91.5% and 73.8% for Sol. It stays reliable exactly where models normally lose the plot.

---

🔬 Science, maths and health

Astra did something models haven't done before: it contributed new mathematics. 🧮

🔢 Result	What changed
Small prime gaps	For a decade the best result said infinitely many primes sit ≤ 246 apart. Julia Stadlmann recently got it to 240. Astra helped establish 186.
Large prime gaps	Astra improved a term in a bound that had stood unchanged for 80+ years.

OpenAI published the proofs, an abridged chain of thought, and verification materials for both.

xychart-beta
    title "Science & maths benchmarks (%)"
    x-axis ["FrontierMath T4", "GPQA Diamond", "TB Science 0.1", "HealthBench Pro"]
    y-axis "Score" 0 --> 100
    bar [97.6, 96.0, 64.6, 63.4]
    bar [83.0, 94.6, 22.4, 60.5]

Gold = Astra · Second = GPT‑5.6 Sol

Benchmark	🌟 Astra	Sol	Fable 5.1	Opus 5
FrontierMath Tier 4 (v2)	97.6%	83.0%	87.8%	73.2%
GPQA Diamond	96.0%	94.6%	93.7%	93.7%
Terminal-Bench Science 0.1	64.6%	22.4%	52.6%	30.0%
HealthBench Professional	63.4%	60.5%	58.1%	56.4%
LifeSciBench	60.3%	59.9%	—	—
GeneBench Pro	37.1%	32.3%	—	—

🧬 And it can drive the lab software too

▶ Watch: a cell-tracking workflow, in real time

Because it can operate specialist software, Astra can open a genomics tool, inspect sequencing quality, visualise genetic variation and tell a researcher where to look next — instead of describing how one might.

---

🧠 How Astra thinks — and why experts are worried

This is the most important part of the launch, and most coverage skips it.

Older reasoning models wrote their thinking down: step 1, step 2, step 3. That trail was the safety community's best window into why a model did something — you could put a monitor on it.

Astra uses a technique widely reported as "recurrent depth" (also called opaque recurrence): it loops over a problem internally, in its own numeric representations, instead of writing each step out as text.

flowchart TB
    subgraph OLD2["🔍 Written reasoning — a monitor can read it"]
        direction TB
        P1["Prompt"] --> S1["step 1 (text)"]
        S1 --> S2["step 2 (text)"]
        S2 --> S3["step 3 (text)"]
        S3 --> A1["Answer"]
        S2 -.->|"👁️ readable"| MON1["Safety monitor"]
    end

    subgraph NEW2["🌀 Recurrent depth — fewer written steps"]
        direction TB
        P2["Prompt"] --> L["🌀 internal loop<br/>×N passes<br/><i>not written down</i>"]
        L --> L
        L --> A2["Answer"]
        L -.->|"👁️ much less to read"| MON2["Safety monitor"]
    end

The trade-off in one line: the same property that lets Astra solve a problem in fewer written steps means there are fewer written steps to inspect.

OpenAI says so itself in the launch post — its evaluations found Astra's written reasoning harder to monitor than Sol's on tests that explicitly asked it to evade monitoring.

😟 What the critics said

Who	What they said
Ryan Greenblatt · chief scientist, Redwood Research	Called it the single worst development for AI safety and security to date
Buck Shlegeris · CEO, Redwood Research	Pushing the technique further moves reasoning out of visible channels entirely
Micah Carroll · preparedness lead, OpenAI	A major capability jump and an important monitorability regression; labs need shared minimum standards
Jakub Pachocki · chief scientist, OpenAI	OpenAI would hold back scaling rather than accept further loss of monitoring confidence

⚖️ The fair summary: OpenAI disclosed this itself, in the same post where it called Astra its most aligned model. Both are true at once — better behaviour, worse visibility.

---

🔐 Cybersecurity: the double-edged sword

Astra is the first OpenAI model to meet the "Critical" threshold for cybersecurity under the company's Preparedness Framework. It can find and build zero-days.

Tested without production safeguards:

xychart-beta
    title "Cyber capability — Astra vs GPT-5.6 Sol (%)"
    x-axis ["ExploitBench", "ExploitGym", "New bugs Jun-Aug", "SRE-Bench", "SEC-Bench Pro"]
    y-axis "Score" 0 --> 100
    bar [100, 42.4, 39.0, 88.0, 85.4]
    bar [78.5, 30.3, 5.5, 55.9, 79.1]

Benchmark	🌟 Astra	Sol	What it tests
ExploitBench	100.0%	78.5%	Known bugs → working exploits
ExploitGym	42.4%	30.3%	Harder, open-ended exploitation
ExploitBench (Jun–Aug 2026)	39.0%	5.5%*	Brand-new bugs, contamination-free
SRE-Bench	88.0%	55.9%	Reverse-engineering binaries with no source
SEC-Bench Pro	85.4%	79.1%	Broad security tasks

* OpenAI notes Sol's 5.5% is an artefact of a 300-turn limit; with fewer limits it reached 11.5%.

On SRE-Bench, Astra solved 88.0% first try and 99.2% within four attempts. During the fresh-vulnerability test it found and used two previously unknown zero-days — which OpenAI says it is disclosing to the maintainers. 🐛

✅ Allowed today	⏳ Coming via OpenAI Daybreak	🚫 Refused
Secure code review	Vulnerability & PoC validation	Writing proof-of-concept exploits
Patching	Malware analysis	Advanced offensive security work
	Detection engineering

---

🛡️ Safety: the genuinely good news

On behaviour, Astra is a big improvement. OpenAI built a new evaluation informed by the "Hugging Face incident" of July 2026, testing whether a model facing an impossible task goes beyond its authorised scope.

⬇️ Lower is better on every row.

Test	🌟 Astra	Sol	Fable 5.1	Opus 5
Went beyond authorised target (impossible task)	0%	48%	—	—
Internal computer-use safety benchmark	2.4%	22.0%	9.5%	11.5%
…with Auto-Review on	1.8%	4.3%	—	—
ExploitGym honeypot	0.0%	48.2%	—	—
Tried to circumvent an Auto-Review denial	0.00%	0.29%	—	—
Internal hallucination benchmark	4.2%	12.2%	—	—

• 🎯 0% out-of-scope. Where Sol overstepped nearly half the time, Astra did so in none of the tested cases.
• 🗣️ 3× less overclaiming about its own capabilities — fewer confident promises it can't keep.
• 🛑 Live monitoring in production. Classifiers watch reasoning and actions and can auto-stop unauthorised activity. Side effect: legitimate work occasionally gets paused — you confirm in ChatGPT/Codex; in the API the task simply stops.

---

⚖️ The honest bit: where Astra is not best

Every launch post is a highlight reel. Read OpenAI's own tables carefully and you find:

Benchmark	🌟 Astra	Best rival	Verdict
Humanity's Last Exam (w/ tools)	57.2%	65.0% — Claude Fable 5.1	❌ Clear loss
AA Intelligence Index v4.1.1	61.2	65.7 — Claude Fable 5.1	❌ Loss
AA Coding Agent Index v1.4	67.0	68.1 — Claude Opus 5	❌ Loss
FrontierCode 1.1 Main	53.3%	53.5% — Claude Fable 5	🤏 Tie
DeepSWE v1.1	74.1%	73.8% — Gemini 3.8 Flash	🤏 Within noise
Chain-of-thought monitorability	worse	GPT‑5.6 Sol	⚠️ Regression, disclosed by OpenAI

Three caveats that apply to every number above:

1. 📐 Scores are the maximum at any effort setting — best-case runs, not casual chat results.
2. 🏠 OpenAI ran the comparisons. Rival scores carry footnotes about modified evals, fallbacks and different harnesses.
3. 🔬 Benchmarks aren't your job. 97.6% on research maths says nothing about whether it'll format your quarterly report right.

---

💰 Price and access

	Standard	Fast mode
📥 Input / 1M tokens	$10	$20
📤 Output / 1M tokens	$50	$100
♻️ Cached input / 1M	$1	—
⚡ Speed	baseline	up to 2×

ChatGPT — rolling out to Plus, Pro, Business, Enterprise; included in your existing allowance, extra credits purchasable. Pro/Business/Enterprise also get GPT‑6 Astra Pro. Enterprise admins must switch it on — it's off by default.

API — gpt-6-astra, also on Microsoft Azure and AWS Bedrock. Zero Data Retention for eligible customers; Private Safety Processing in testing.

💡 Cost tip: Astra costs more per token but repeatedly used fewer tokens to reach a better score — up to 65% fewer on one benchmark. Judge it on cost per finished task, not cost per token.

---

🗓️ The rollout

timeline
    title GPT-6 Astra rollout
    3 Sep 2026 : Limited preview to trusted partners : Cybersecurity programme partners first
    4 Sep 2026 : Public release begins : Limited set of organisations
    Following days : ChatGPT Plus, Pro, Business, Enterprise : OpenAI API, Azure, AWS Bedrock
    Coming weeks : Codex notes become the default : OpenAI Daybreak expands cyber access

---

🙋 Should you care?

If you are…	What changes
👔 An office worker	Delegate the click-heavy stuff — forms, CRM, calendars, decks on your template. Review instead of produce.
💻 A developer	Better terminal and migration work; long sessions stop losing context. Turn on the Codex notes flag.
🔬 A researcher	It operates your specialist software, not just describes it. On hard maths it's a genuine collaborator.
🔐 A security team	Secure code review and patching today, more as Daybreak expands. Also: attackers get better tools too.
🏢 An IT admin	Off by default on Enterprise. Plan the rollout; expect occasional safety pauses on legitimate work.
🙂 Just curious	Hand it a whole errand — "find three pediatricians near me who take my insurance and book the earliest" — not a question.

---

📊 The full benchmark table

All figures as published by OpenAI. Scores are the maximum at any effort setting. "—" = not reported.

Benchmark	🌟 Astra	GPT‑5.6 Sol	Fable 5.1	Fable 5	Opus 5	Gemini 3.8 Flash
🖱️ Computer use
Agents' Last Exam	59.3%	53.6%	—	48.7%	55.5%	—
OSWorld 2.0 (offline, partial)	72.6%	65.7%	—	—	70.2%	—
ScreenSpot-Pro (no tools)	92.7%	76.9%	—	87.3%	—	—
💼 Professional
AutomationBench	41.4%	18.1%	31.4%	17.4%	26.9%	—
BenchCAD	95.9%	83.3%	84.3%	67.5%	82.1%	—
BrowseComp	91.5%	90.4%	—	87.4%	90.8%	—
OpenScore String Quartets	0.84	0.19	—	—	—	—
Internal design tasks	50.0%	47.4%	—	35.8%	—	—
Internal data-science tasks	40.9%	30.5%	—	34.7%	—	—
AA Intelligence Index v4.1.1	61.2	60.9	65.7	62.1	63.1	58.7
💻 Coding
Terminal-Bench 4.0	57.9%	37.3%	55.8%	44.5%	52.6%	19.1%
DeepSWE v1.1	74.1%	72.7%	67.4%	69.9%	73.7%	73.8%
FrontierCode 1.1 Extended	64.5%	60.6%	63.6%	64.9%	63.6%	56.3%
FrontierCode 1.1 Main	53.3%	47.5%	50.9%	53.5%	53.4%	43.6%
Internal DB migration tasks	63.9%	42.7%	57.8%	50.3%	—	—
AA Coding Agent Index v1.4	67.0	65.1	—	67.2	68.1	61.2
🔬 Academic
Terminal-Bench Science 0.1	64.6%	22.4%	52.6%	21.4%	30.0%	—
FrontierMath Tier 4 (v2)	97.6%	83.0%	87.8%	87.8%	73.2%	—
GPQA Diamond	96.0%	94.6%	93.7%	92.6%	93.7%	95.3%
Humanity's Last Exam (w/ tools)	57.2%	—	65.0%	63.8%	63.6%	—
🧬 Science & health
GeneBench Pro	37.1%	32.3%	—	—	—	—
MedChemBench (internal)	49.3%	47.4%	—	—	—	—
LifeSciBench	60.3%	59.9%	—	—	—	—
HealthBench Professional	63.4%	60.5%	58.1%	60.9%	56.4%	52.1%
🔐 Cybersecurity
ExploitBench	100.0%	78.5%	—	—	70%	—
ExploitGym	42.4%	30.3%	30.4%	28.4%	22.0%	—
ExploitBench (Jun–Aug 2026)	39.0%	5.5%	—	—	—	—
SRE-Bench	88.0%	55.9%	—	—	12.5%	—
SEC-Bench Pro	85.4%	79.1%	—	—	—	—
🛡️ Alignment (lower is better)
Internal computer-use safety	2.4%	22.0%	9.5%	18.3%	11.5%	—
…with Auto-Review	1.8%	4.3%	—	—	—	—
Internal circumvention	0.00%	0.29%	—	—	—	—
ExploitGym honeypot	0.0%	48.2%	—	—	—	—
Internal hallucination	4.2%	12.2%	—	—	—	—
🧵 Long context
MRCR v2, 8-needle, 256K–512K	100.0%	91.5%	—	—	—	—
MRCR v2, 8-needle, 512K–1M	96.3%	73.8%	—	—	—	—
🧩 Abstract reasoning
ARC-AGI-3	99.9%	7.8%	—	—	30.2%	—
ARC-AGI-2	95.0%	92.5%	90.0%	89.2%	90.4%	—
ARC-AGI-1	98.5%	97.5%	97.5%	98.5%	97.5%	—

---

🎯 The bottom line

GPT‑6 Astra is the first model where the interesting question isn't "how good is its answer?" but "how much of the job can I hand over?" 🌟

The wins are real and measured: it operates software, it's roughly twice as fast at it, it stays coherent across a million tokens, it contributed genuine new mathematics, and it behaves dramatically better than the model it replaces.

The costs are real too, and OpenAI published them: a Critical cyber rating that forces a slow, gated rollout — and reasoning its own tests say is harder to watch than last year's.

Both halves of that sentence are the story. 🌗

---

📚 Sources

1. OpenAI — "GPT‑6 Astra: A new generation of intelligence" — primary source for all benchmark tables, quotes, pricing, availability, and every image and video embedded above: https://openai.com/index/gpt-6-astra/
2. Wikipedia — GPT‑6 Astra — training run, "recurrent depth", rollout history: https://en.wikipedia.org/wiki/GPT-6_Astra
3. 9to5Mac — ChatGPT and Codex upgrade details: https://9to5mac.com/2026/09/04/openai-releasing-major-upgrade-to-chatgpt-and-codex-with-gpt-6-astra-details-here/
4. CNBC — rollout announcement: https://www.cnbc.com/2026/09/03/open-ai-astra-gpt-6-cyber.html
5. Axios — the AGI claim, Brockman: https://www.axios.com/2026/09/03/openai-astra-gpt-6-agi-brockman
6. LLM-Stats — context window, knowledge cutoff, modalities: https://llm-stats.com/models/gpt-6-astra
7. TechRadar — experts on recurrent depth: https://www.techradar.com/pro/security/why-is-there-so-much-worry-about-openai-astra-and-what-issues-could-recurrent-depth-reasoning-cause-the-experts-weigh-in
8. Gizmodo — the monitorability problem: https://gizmodo.com/openai-says-humans-need-to-be-able-to-monitor-how-ai-thinks-its-new-model-astra-makes-that-much-harder-2000807665
9. Implicator.ai — OpenAI's own monitorability finding: https://www.implicator.ai/openai-says-its-own-tests-found-gpt-6-astra-harder-to-monitor/
10. Artificial Analysis — independent index scores: https://artificialanalysis.ai/models/gpt-6-astra

---

Written 5 September 2026. All media is hosted by OpenAI and embedded from their public servers. Benchmark figures are OpenAI's own published numbers unless marked otherwise; context window, knowledge cutoff and modality details come from third-party trackers and may be revised.