OpenAI®FRONTIER INTELLIGENCE
RESEARCH / 2026
INTRODUCING / GPT–6 ASTRA A NEW STATE OF INTELLIGENCE
THE NEXT CHAPTER

GPT–6

ASTRA

A new generation
of intelligence.

ENTER ASTRA
+ FIG. 006
INTELLIGENCE, IN FORM.
SCROLL TO EXPLORE ↓COMPUTER USE / CODE / SCIENCE / CYBERSECURITY / PROFESSIONAL WORK01 — 13

01 / INTELLIGENCE

ASTRA / 006

INTELLIGENCE.
WITHOUT A
FIXED SHAPE.

Not a single task.
A new capacity to act.

OpenAI’s frontier model for its most demanding end-to-end work.

A difficult problem. An unfamiliar environment. Room to think.

02 / CAPABILITY FIELD

OPENAI-REPORTED EVALUATIONS / 2026
X 001
Y 006
Z ACTIVE

ABSTRACT REASONING

99.9%

ARC–AGI–3

Learning to navigate an unfamiliar world.

Maximum reported scores across reasoning efforts. Research and API conditions may differ from ChatGPT.

03 / COMPUTER USE

ASTRA / 006

THE COMPUTER
BECOMES AN
ENVIRONMENT.

BROWSER / WORKSPACE

A question becomes a path.

  1. 01Research a question
  2. 02Inspect sources
  3. 03Compare findings
RESEARCH / SYNTHESISCONCEPTUAL WORKFLOW

Forms. CRM updates. Research. Documents. Scientific data. Websites and front-end QA.
Work moves across tools, with the task held in view.

04 / COMPRESSION OF TIME

ASTRA / 006

LESS TIME.
MORE WORLD.

47%LESS TIME PER TASK*
GPT–5.6 SOL65.7% / ~75 MIN
GPT–6 ASTRA72.6% / ~40 MIN
001530456075 MIN

*OpenAI latency simulation on OSWorld 2.0. Approximate task times.
Higher task performance in a shorter trajectory. Read the evaluation ↗

05 / PROFESSIONAL WORK

ASTRA / 006

FROM THOUGHT
TO ARTIFACT.

ASTRA / DOCUMENTS006
An idea.
Made useful.
Form follows
understanding.
DESIGN → STRUCTURE → ARTIFACTLAYOUT STUDY

Clarity is
a working material.

Stronger template following. Structured narratives. More usable business outputs.

06 / CODING

ASTRA / 006

CODE THAT STAYS
IN CONTEXT.

WORKING CONTEXT ASTRA / CODEX
01 / REPOSITORY

Read the structure.

Trace requirements through the repository before making a change.

REQUIREMENTS → TOOLS → EVIDENCE → RESULT
Terminal-Bench 4.057.9%
DeepSWE v1.174.1%
FrontierCode 1.1 Extended64.5%
Database Migration Tasks63.9%

Context retrieval across windows is an experimental Codex feature.
It is not permanent memory for every Astra API request.

07 / LONG CONTEXT

ASTRA / 006

A larger space
to hold a thought.

1,050,000

TOKENS OF CONTEXT

128,000MAXIMUM OUTPUT TOKENS
REASONING EFFORT / EXPLORE THE SCALE

KNOWLEDGE CUTOFF
APRIL 30, 2026

MODEL / gpt-6-astra

08 / SCIENCE

ASTRA / 006

INTELLIGENCE
AS AN INSTRUMENT.

OBSERVATION / PARAMETRIC FIELD
ILLUSTRATIVE SCIENTIFIC VISUALIZATION
FrontierMath Tier 4 (v2)97.6%
GPQA Diamond96.0%
Terminal-Bench Science 0.164.6%
GeneBench Pro37.1%
LifeSciBench60.3%

Scientific reasoning meets specialized software.
Inspect data. Explore results. Let the evidence guide the next question.

09 / CYBERSECURITY

ASTRA / 006

CAPABILITY
REQUIRES CONTROL.

01 / BOUNDARY+
02 / AUTHORIZATION+
03 / MONITORING+
04 / AUTO REVIEW+
05 / CONTAINMENT+
AUTHORIZED SCOPE
CONTROLLED CAPABILITY

Greater capability.
Stronger safeguards.

The first OpenAI model to meet the Critical cybersecurity capability threshold under the Preparedness Framework.

ExploitBench100.0%
ExploitGym42.4%
ExploitBench · Jun–Aug 202639.0%
SRE-Bench88.0%
SEC-Bench Pro85.4%

Cyber evaluations include testing without production safeguards. Deployed access includes stronger controls.
Advanced exploit creation is restricted at launch. Safety overview ↗

10 / ALIGNMENT

ASTRA / 006

POWER.
WITH BOUNDARIES.

INTERNAL COMPUTER-USE SAFETY BENCHMARK

2.4%GPT–6 ASTRA
22.0%GPT–5.6 SOL

↓ LOWER IS BETTER

INTERNAL HALLUCINATION BENCHMARK

4.2%GPT–6 ASTRA
12.2%GPT–5.6 SOL

↓ LOWER IS BETTER

Knowing what to do.
And where to stop.

OpenAI reports improved respect for user authorization and fewer misaligned outcomes. These evaluations do not imply perfect safety. Monitorability remains an active research challenge.

11 / MODEL ENGINE

ASTRA / 006

ASTRA

ENGINE SPECIFICATION
OPENAI / 2026
REV. 09.05

MODEL IDENTIFIER
CONTEXT WINDOW
1,050,000 tokens
MAXIMUM OUTPUT
128,000 tokens
KNOWLEDGE CUTOFF
April 30, 2026
INPUT
Text / Image
OUTPUT
Text
AUDIO / VIDEO
Not supported

STANDARD API / PER 1M TOKENS

INPUT$10.00
CACHED INPUT$1.00
CACHE WRITE$12.50
OUTPUT$50.00
SUPPORTED TOOLS

Web Search / File Search / Image Generation / Code Interpreter / Hosted Shell / Apply Patch / Skills / Computer Use / MCP / Tool Search

STREAMING / FUNCTION CALLING / STRUCTURED OUTPUTSOPEN MODEL DOCUMENTATION ↗

Requests above 272K input tokens: 2× input/cache rates and 1.5× output rates for the entire request.
Prices and specifications checked against OpenAI documentation on September 5, 2026.

12 / AVAILABILITY

ASTRA / 006

A new chapter.
Opening gradually.

Announced September 3, 2026. Initial access began with limited organizations, with broader rollout planned over the following days.

ChatGPT Plus / Pro / Business / Enterprise
OpenAI API / Microsoft Azure / AWS Bedrock

Astra Pro is planned for Pro, Business and Enterprise. Enterprise access is off by default at launch. Availability depends on rollout and account access.

OFFICIAL AVAILABILITY

13 / A NEW STATE OF INTELLIGENCE

ASTRA

THE MODEL
IS NO LONGER
JUST THE INTERFACE.

IT ACTS.
GPT–6 ASTRA2026