GPT‑6 Astra

gpt-6-astra

GPT-6 Astra is purpose-built for the most complex end-to-end work. It can be used for complex reasoning, coding, computer operation, research, and document creation.

Context window1.1M

ProviderOpenAI

Released2026/09/04

Playground

Pricing

The same model is available through multiple service channels — choose based on latency, reliability and cost.

Prices in $ / 1M tokens

To pick a channel, add a provider field to the request body, for example "provider": { "channel": "direct" }. Valid values are direct / stable / economical; omit it to use the default channel.

Direct

Direct upstream connection — best when you need native behavior and the full context window.

Input context	Input	Output	Cache read	Cache write
≤ 272K	10.00/M	50.00/M	1.00/M	12.50/M
> 272K	20.00/M	75.00/M	2.00/M	25.00/M

10% off

Preferred

A balanced route optimized for availability and speed — a good default for production traffic.

Input context	Input	Output	Cache read	Cache write
≤ 272K	10.009.00/M	50.0045.00/M	1.000.90/M	12.5011.25/M
> 272K	20.0018.00/M	75.0067.50/M	2.001.80/M	25.0022.50/M

Overview

Input

Text Image

Output

Text

GPT-6 Astra API: The Agent Flagship That Actually Puts Computers to Work

GPT-6 Astra is OpenAI's new flagship model released on September 3, 2026, and fully opened to API developers on September 5. It breaks out of the three-tier naming system of GPT-5.6 and represents OpenAI's largest training run to date—the VP of Research says it is the first time they have pretrained on the Stargate cluster in Texas with more than 100,000 GPUs.

The most tangible change this generation is in computer operation: ScreenSpot-Pro 92.7% (GPT-5.6 Sol 76.9%), OSWorld 2.0 72.6% (Sol 65.7%). Astra averages 40 minutes on the same batch of tasks, while Sol needs 75 minutes. It is the first model OpenAI has determined, under its Preparedness Framework, to cross the "critical" cybersecurity threshold, scoring a direct 100.0% on ExploitBench.

One thing to be clear about upfront: Astra does not dominate across the board. Its Artificial Analysis intelligence index is 61, tied with GPT-5.6 Sol and five points behind Claude Fable 5.1; yet its price tier has been raised from Sol's $4/$20 to $10/$50. It is a specialized machine built for agents and computer operation, not a cost-effective choice for general Q&A.

SeaWhale AI provides GPT-6 Astra through an OpenAI-compatible interface, with support for tool calling, reasoning effort control, streaming output, structured output, and image-text input.

Get API Key · Model ID: gpt-6-astra

Why Choose GPT-6 Astra

A generational lead in computer operation — ScreenSpot-Pro 92.7%, 15.8 percentage points higher than Sol, with the same batch of tasks cut from 75 minutes to 40 minutes
Perfect score in cybersecurity benchmarks — ExploitBench 100.0% (Sol 78.5%), SRE-Bench 88.0% (Opus 5 only 12.5%)
Frontier math near saturation — FrontierMath Tier 4 (v2) 97.6%, pulling nearly 10 points ahead of Fable 5.1's 87.8%
Hallucination rate sharply reduced — down from 92% to 51%, with accuracy up 4 points at the same time
1.05M-token context window — max input 922K, max output 128K tokens
Coding token efficiency improved — coding tasks use roughly one-third of the previous generation's output tokens

Core Capabilities

01 Computer Operation and Agent Execution

This is Astra's most uncontested strength. It can operate software interfaces directly: filling in forms, updating records, conducting online research, managing calendars, analyzing scientific data, building websites, and testing applications. ScreenSpot-Pro 92.7%, OSWorld 2.0 72.6%, and Agents' Last Exam 59.3%—all three are clearly higher than Sol and the Claude family.

GUI element localization and multi-step operations
Repetitive office workflows such as form filling and record maintenance
Browser research and cross-application data transfer

02 Long-Horizon Coding

Astra ships with an updated Codex harness, which OpenAI says delivers task completion on Mind2Web at 1.9x the speed of the GPT-5.6 Sol experience. On benchmarks, Terminal-Bench 4.0 scores 57.7% (Fable 5.1 55.8%, Opus 5 52.3%) and DeepSWE v1.1 scores 74.1% (Opus 5 73.7%, Gemini 3.8 Flash 73.8%)—the lead is not huge, but it is consistently ahead. More noteworthy is efficiency: output tokens in coding scenarios are about one-third of the previous generation's.

Codebase-level changes in long sessions
Searchable notes across context windows (experimental feature)
Defect localization, patch generation, and regression validation

03 Defensive Cybersecurity

ExploitBench 100.0% and SRE-Bench 88.0% represent the biggest gap this generation has opened, which is exactly why OpenAI has placed it in the "critical" capability tier with additional controls. Its official positioning is defensive: secure code review, vulnerability patching, and incident response.

Secure code review and vulnerability identification
Patch authoring and validation
Fault localization and handling in SRE scenarios

04 Professional Documents and Scientific Research

AA-Briefcase (long-horizon knowledge work) improved by about 80 points over the previous generation—Astra's largest gain on the knowledge-work front. Combined with a 1.05M-token context window, it is well suited to reading full document collections end to end, producing professional documents and slides, and analyzing scientific data.

Comprehensive analysis of long document collections
Generation of professional reports and presentation materials
Scientific data processing and conclusion synthesis

Best Use Cases

Scenario	Description
Computer operation agents	ScreenSpot-Pro 92.7%, OSWorld 2.0 72.6%—currently the strongest tier in this direction
Office workflow automation	Cross-application repetitive work on forms, records, calendars, etc., taking about half the time of Sol
Long-session coding	Paired with the Codex harness; coding output tokens about one-third of the previous generation
Defensive security engineering	Perfect ExploitBench score, used for code review, patching, and incident response
Frontier math and scientific research	FrontierMath Tier 4 (v2) 97.6%, near saturation
Ultra-long document research	1.05M-token context window, max input 922K tokens

Differences Between GPT-6 Astra, GPT-5.6 Sol, and Claude Fable 5.1

Capability	GPT-6 Astra	GPT-5.6 Sol	Claude Fable 5.1
Model ID	`gpt-6-astra`	`gpt-5.6-sol`	`claude-fable-5-1`
Vendor	OpenAI	OpenAI	Anthropic
AA Intelligence Index	61	61	66
ScreenSpot-Pro	92.7%	76.9%	87.3% (Fable 5)
OSWorld 2.0	72.6%	65.7%	—
Terminal-Bench 4.0	57.7%	—	55.8%
DeepSWE v1.1	74.1%	—	67.4%
FrontierMath Tier 4 (v2)	97.6%	—	87.8%
HLE (with tools)	57.2%	—	65.0%
ExploitBench	100.0%	78.5%	—
Context window	1.05M tokens	—	—
Max output	128K tokens	—	—
Reasoning effort levels	`low`/`medium`/`high`/`xhigh`/`max`	Adjustable	Adjustable
Positioning	Computer operation and security engineering flagship	Frontier coding and agent flagship	Strongest overall intelligence tier

Specific pricing is subject to the real-time price card at the top of the page.

FAQ

When was GPT-6 Astra released? It launched in limited release on September 3, 2026, first opening to organizations in the Daybreak Access program; on September 5 it fully opened to ChatGPT Plus/Pro/Business/Enterprise and API developers.

What has been upgraded compared with GPT-5.6 Sol? The biggest change is computer operation: ScreenSpot-Pro from 76.9% to 92.7%, OSWorld 2.0 from 65.7% to 72.6%, and the same batch of tasks dropping from 75 minutes to 40 minutes. Next is cybersecurity (ExploitBench 78.5% → 100.0%) and hallucination rate (92% → 51%, with accuracy up 4 points at the same time). Coding improvements are modest, but output tokens drop to roughly one-third.

Are there areas where it lags? Yes, and more than one. Its Artificial Analysis intelligence index of 61 merely ties GPT-5.6 Sol, trailing Claude Fable 5.1 by five points and also behind Meta's Muse Spark 1.3 (max). Humanity's Last Exam (with tools) is 57.2%, notably lower than Fable 5.1's 65.0%. GDPval-AA v2 regressed about 80 Elo from the previous generation, with τ³-Banking, SciCode, and AA-LCR each down 2-3 points. Pricing has gone from $4/$20 to $10/$50, and AA estimates per-task cost is about 75% higher than Sol—it is not cost-effective for general intelligence scenarios; its strengths are concentrated in computer operation, security, and math.

What are the context and output limits? The context window is 1.05M tokens, with max input of 922K and max output of 128K tokens. Input supports text and images; output is text only. Knowledge cutoff: April 30, 2026.

How do I set reasoning effort? reasoning.effort supports five levels: low, medium, high, xhigh, and max. The default level is sufficient for everyday use; for its strongest scenarios, such as computer operation and math, xhigh or max is recommended; for simple tasks, use lower levels to save tokens.

What tools and features are supported? Function calling, streaming output, structured output, prompt caching, and image input are all supported. Officially hosted tools cover web search, file retrieval, image generation, code interpreter, hosted shell, apply patch, computer use, MCP, and tool search.

Why are its cybersecurity capabilities specially controlled? Astra is the first OpenAI model determined under the Preparedness Framework to reach the "critical" cybersecurity threshold—meaning it could potentially discover and exploit unknown vulnerabilities in well-defended systems without step-by-step human guidance. Advanced cybersecurity capabilities are therefore initially open only to testers, then gradually expanded through Daybreak Blue. The official positioning is defensive use.

Why Use SeaWhale AI for the GPT-6 Astra API

No overseas account needed — direct connection in mainland China, no need to build your own proxy
OpenAI-compatible API — integrate by changing two lines in your existing code
Cross-vendor comparison — compare price and performance with Claude, Gemini, and Grok directly under the same account
Unified billing — usage and spend across multiple models managed centrally

API

API integration

Model IDUse this value as the model in inference requests

gpt-6-astra

API KeyBearer token used to authenticate inference requests

Manage API keys

Base URLOpenAI compatible · /chat/completions

https://api.haijingai.com/v2/

provider OptionalSelects a service channel; omit it and the system picks the default

"provider": { "channel": "direct" }

gpt-6-astra usage examples

SeaWhale AI is compatible with the OpenAI API protocol, so you can call it with the OpenAI SDK or plain HTTP requests. Streaming is enabled by default.

cURLPythonNode.js

curl https://api.haijingai.com/v2/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <API_KEY>" \
  -d '{
    "model": "gpt-6-astra",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Hello!"}
    ],
    "provider": { "channel": "direct" },
    "stream": true
  }'
# provider is optional — remove this line to use the default channel

from openai import OpenAI

client = OpenAI(
    base_url="https://api.haijingai.com/v2",
    api_key="<API_KEY>",
)

stream = client.chat.completions.create(
    model="gpt-6-astra",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello!"},
    ],
    stream=True,
    # Optional: pick a service channel; omit to use the default
    extra_body={"provider": {"channel": "direct"}},
)

for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

import OpenAI from 'openai'

const client = new OpenAI({
  baseURL: 'https://api.haijingai.com/v2',
  apiKey: '<API_KEY>',
})

const stream = await client.chat.completions.create({
  model: 'gpt-6-astra',
  messages: [
    { role: 'system', content: 'You are a helpful assistant.' },
    { role: 'user', content: 'Hello!' },
  ],
  stream: true,
  // Optional: pick a service channel; omit to use the default
  // @ts-expect-error provider is a SeaWhale AI extension, not in the OpenAI SDK types
  provider: { channel: 'direct' },
})

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? '')
}