GPT-6 Sol

gpt-6-sol

GPT-6 Sol is designed for the complex tasks developers and knowledge workers repeatedly perform every day—writing features, reviewing code, troubleshooting defects, and analyzing data.

Context window1.0M

ProviderOpenAI

Released2026/09/23

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	2.00/M	10.00/M	0.20/M	2.50/M
> 272K	4.00/M	15.00/M	0.40/M	5.00/M

Overview

Input

Text Image

Output

Text

GPT-6 Sol API: The Half-Price Frontier Workhorse Model

GPT-6 Sol is the main tier of the GPT-6 series released by OpenAI on September 22, 2026. Together with GPT-6 Luna released the same day, it fills out the two tiers below GPT-6 Astra from early September. It targets the complex tasks that developers and knowledge workers repeatedly do every day—writing features, reviewing code, troubleshooting defects, and analyzing data.

The most direct change in this generation is price: official pricing is exactly half of GPT-5.6 Sol, for both input and output, and OpenAI confirms this is long-term pricing rather than a limited-time discount. Artificial Analysis measured the per-task cost of running the full Intelligence Index dropping from $1.99 to $1.06. In terms of capability, its Intelligence Index score is 48, essentially on par with GPT-5.6 Sol's 47; OpenAI's internal factual evaluation shows the number of errors roughly halved, reaching Astra-level reliability.

It should be made clear: Sol is not a comprehensive upgrade. DeepSWE, OSWorld 2.0, and GDPval-AA are all lower than GPT-5.6 Sol. Its selling point is "same intelligence, half the cost, fewer errors", not a leap in capability.

SeaWhale AI provides GPT-6 Sol through an OpenAI-compatible API, supporting tool calling, reasoning effort control, streaming output, structured output, and text-and-image input.

Get API Key · Model ID: gpt-6-sol

Why Choose GPT-6 Sol

Price Halved — Official pricing is 1/2 of GPT-5.6 Sol; AA Intelligence Index per-task cost is $1.06, versus $1.99 for the previous generation
Fewer Errors — OpenAI internal factual evaluation error count is about half that of the previous generation; AA-Omniscience hallucination rate drops from 92% to 60%
Coding Agent Index +2 — AA Coding Agent Index is 57, with Terminal-Bench 4.0 improving from 37% to 43%
High Cost-Effectiveness for Business Automation — AutomationBench 1.0.6 score is 33.2%, at $0.27 per task, 6.3 percentage points higher than Claude Opus 5
Coding Deception Rate Drops Sharply — 1.3% in internal testing, versus 10.4% for GPT-5.6 Sol
1.05M token Context — Maximum input 922K tokens, maximum output 128K tokens

Core Capabilities

01 Business Process Automation

On AutomationBench 1.0.6 (by Zapier, covering 47 cross-functional business tools), Sol scores 33.2% at xhigh, 2.9 percentage points higher than GPT-6 Astra (low), and 6.3 percentage points higher than Claude Opus 5 (max) at 26.9%, while costing less than 1/11 per task of Opus 5. Repetitive business processes that chain multiple tools are where Sol's cost-effectiveness stands out most.

Multi-step workflow orchestration across SaaS tools
Batch maintenance of records such as CRM, tickets, and spreadsheets
Office automation requiring stable, repeated execution

02 Coding and Code Review

The AA Coding Agent Index is 57, 2 points higher than GPT-5.6 Sol, with the main progress in terminal tasks (Terminal-Bench 4.0 43% vs 37%) and codebase Q&A (SWE-Atlas-QnA 58% vs 54%). On the other hand, DeepSWE v1.1 scores 68.8% at max, lower than GPT-5.6 Sol's 72.7%. It is better suited for everyday feature development, code review, and issue diagnosis, rather than pushing the hardest software engineering benchmarks.

Multi-step development tasks in terminal environments
Codebase understanding and Q&A
Code review and defect investigation

03 Reliability and Factuality

This is what OpenAI emphasized most this time: in internal factual evaluations (based on real conversations users flagged as errors), Sol's error count is about half of GPT-5.6 Sol's. Third-party data points in the same direction—AA-Omniscience score rises from 22 to 27, and hallucination rate drops from 92% to 60%. The deception rate in coding scenarios falls from 10.4% to 1.3%.

Analysis reports that need cited facts and reduced fabrication
Data processing with low tolerance for errors
More trustworthy result self-reporting in long-chain tasks

04 Professional Tasks and Computer Use

Agents' Last Exam (covering long-horizon economic tasks across 55 sub-industries) scores 56.4% at max, which OpenAI says exceeds Claude Opus 5's best result, at 60% lower cost per task. OSWorld 2.0 scores 60.5% at xhigh, on par with Opus 5 (medium) at 60.3%, and about 80% lower cost—but lower than GPT-5.6 Sol's 65.7%.

Cross-industry professional knowledge tasks
Computer use in graphical interfaces
Data analysis and conclusion synthesis

Best Use Cases

Scenario	Description
Business process automation	AutomationBench 33.2%, low per-task price for multi-tool chained tasks
Daily coding workhorse	Coding Agent Index 57, clear improvement on terminal tasks
Code review and Q&A	SWE-Atlas-QnA 58%, suited for understanding existing codebases
Accuracy-sensitive analysis	Error count about half of the previous generation, hallucination rate down to 60%
Migrating from GPT-5.6 Sol to reduce cost	Intelligence Index on par, cost halved
Long document processing	1.05M token context, maximum input 922K tokens

Differences Between GPT-6 Sol, GPT-5.6 Sol, and GPT-6 Astra

Capability	GPT-6 Sol	GPT-5.6 Sol	GPT-6 Astra
Model ID	`gpt-6-sol`	`gpt-5.6-sol`	`gpt-6-astra`
Release date	September 22, 2026	—	September 3, 2026
AA Intelligence Index (v4.3)	48	47	53
AA Coding Agent Index	57	55	—
DeepSWE v1.1	68.8%	72.7%	74.1%
OSWorld 2.0	60.5%	65.7%	72.6%
AA-Omniscience hallucination rate	60%	92%	—
Context window	1.05M tokens	—	1.05M tokens
Maximum output	128K tokens	—	128K tokens
Knowledge cutoff	April 20, 2026	—	April 30, 2026
Price tier	1/2 of GPT-5.6 Sol	Baseline	Highest tier
Positioning	Cost-effective main tier	Previous-generation main tier	Flagship for computer use and safety

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

FAQ

When was GPT-6 Sol released? September 22, 2026, released the same day as GPT-6 Luna and made available that day in the API, Codex, and paid ChatGPT accounts.

What upgrades does it have compared with GPT-5.6 Sol? Three things: price halved; factual errors roughly halved, with hallucination rate dropping from 92% to 60%; Coding Agent Index +2, with terminal tasks showing the clearest progress. The overall Intelligence Index only goes from 47 to 48, essentially flat—the main theme of this generation is cost reduction and reliability, not a leap in capability.

Are there any areas where it falls behind? Yes, and quite a few. DeepSWE v1.1 is 68.8%, lower than GPT-5.6 Sol's 72.7%; OSWorld 2.0 is 60.5%, lower than the previous generation's 65.7%; GDPval-AA v2.1 drops from 1588 Elo to 1487. In addition, when a single prompt exceeds 272K tokens, input is billed at 2x and output at 1.5x, essentially offsetting the price reduction in ultra-long-context scenarios. If you need the strongest computer use or software engineering capability, you should choose GPT-6 Astra.

What are the context and output limits? The context window is 1.05M tokens, with maximum input 922K and maximum output 128K tokens. Input supports text and images, and output is plain text. Knowledge cutoff is April 20, 2026.

How do I set reasoning effort? reasoning_effort supports six levels: none, low, medium (default), high, xhigh, and max, and can be set to none to fully disable thinking. Official benchmarks are mostly measured at xhigh or max, while everyday tasks can use the default level. Note: when using function calling through the Chat Completions API, the official requirement is to set reasoning_effort to none; if you need reasoning and tool calling enabled at the same time, it is recommended to use the Responses API.

Which tools and features are supported? Function calling, streaming output, structured output, Prompt caching, image input, and Batch are all supported. Official hosted tools cover web search, file retrieval, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search. Cached input is billed at 10% of the uncached price.

Should I choose Sol or Luna? Choose by task complexity. For high-frequency tasks with clear goals, such as summarization, extraction, and simple Q&A, use Luna, whose price tier is about 1/20 of Sol; for tasks requiring sustained reasoning, such as writing features, reviewing code, and multi-tool business workflows, use Sol.

Why Choose SeaWhale AI to Use the GPT-6 Sol API

No Overseas Account Required — Direct connection within China, no need to self-host a proxy
OpenAI-Compatible API — Existing code can be integrated by changing two lines
Cross-Vendor Side-by-Side Comparison — Compare price and performance directly with Claude, Gemini, and Grok under the same account
Unified Billing — Centralized management of multi-model usage and spending

API

API integration

Model IDUse this value as the model in inference requests

gpt-6-sol

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-sol 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-sol",
    "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-sol",
    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-sol',
  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 ?? '')
}