GPT-6.1 Sol

gpt-6.1-sol

Get API keyAPI reference

在智能体编码、计算机操作和专业工作上，它达到了"接近 GPT-6 Astra"的水平，而输入输出单价只有 Astra 的五分之一

Context window1.1M

ProviderOpenAI

Released2026/10/10

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.

Preferred

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

Input context	Input	Output	Cache read	Cache write
≤ 272K	2.00/M	10.00/M	0.10/M	2.50/M
> 272K	4.00/M	15.00/M	0.20/M	5.00/M

Overview

Input

Text Image

Output

Text

GPT-6.1 Sol API: The Mainstream Tier Close to Astra at One-Fifth the Price

GPT-6.1 Sol is the mainstream-tier update in the GPT-6 series that OpenAI released on September 29, 2026, at DevDay, just one week after GPT-6 Sol launched. The official line is that on agentic coding, computer use, and professional work, it reaches a level "close to GPT-6 Astra," while its input and output unit prices are only one-fifth of Astra's.

Compared with GPT-6 Sol a week earlier, this is a genuine capability upgrade, not just cost reduction. DeepSWE v1.1 scores 75.2%, 6.4 percentage points above GPT-6 Sol's best result and on par with Astra's 74.8%; OSWorld 2.0 offline set scores 71.4% at the max setting, 7 percentage points above GPT-6 Sol and only 2.1 percentage points behind Astra. On the Artificial Analysis Intelligence Index v4.3.2, it scores 52, compared with 48 for GPT-6 Sol and 53 for Astra.

On price, the input and output unit prices are the same as GPT-6 Sol, while cached input is halved again, down to 5% of the uncached input price. Artificial Analysis measured the per-task cost of running the full Intelligence Index at $0.72, compared with $1.04 for GPT-6 Sol and $3.26 for Astra.

One thing to know in advance: this generation removes the none reasoning setting, and the Chat Completions API no longer supports tool calling; scenarios that need tools must use the Responses API.

SeaWhale AI provides GPT-6.1 Sol through an OpenAI-compatible API, supporting reasoning effort control, streaming output, structured output, and text-and-image input.

Get API Key · Model ID: gpt-6.1-sol

---

Why Choose GPT-6.1 Sol

• Software Engineering Matches Astra — DeepSWE v1.1 scores 75.2%, compared with 74.8% for Astra and 68.8% for GPT-6 Sol
• Computer Use Approaches Astra — OSWorld 2.0 offline set at max scores 71.4%, compared with 73.5% for Astra and 64.4% for GPT-6 Sol
• AA Intelligence Index +4 — v4.3.2 index scores 52, compared with 48 for GPT-6 Sol and 53 for Astra
• Cached Input Halved Again — cache hits are billed at 5% of the uncached input price, compared with 10% for GPT-6 Sol
• Fewer Errors — in OpenAI internal evaluations, the share of responses containing factual errors at the low setting drops from 11.4% to 7.7%
• 1.05M Token Context — maximum input 922K, maximum output 128K tokens

---

Core Capabilities

01 Complex Coding and Codebase Investigation

OpenAI positions this generation for complex refactoring, deep codebase investigation, and long-running agents. DeepSWE v1.1 scores 75.2% at the high setting, 6.4 percentage points above GPT-6 Sol's 68.8% at the max setting, matching Astra; OpenAI's internal research debugging evaluation is 75.52%, compared with 64.20% for GPT-6 Sol. Artificial Analysis measured a 12-point improvement over GPT-6 Sol on Terminal-Bench 4.0.

• Complex cross-file refactoring and feature development
• Issue localization and root-cause analysis in large codebases
• Long-running coding agents

02 Computer Use and Business Automation

OSWorld 2.0 offline set scores 71.4% at the max setting, 7 percentage points above GPT-6 Sol. AutomationBench 1.0.6 scores 35.4% at the medium setting, 4.8 percentage points above GPT-6 Sol at the same setting; OpenAI says it is 2.2 percentage points above Claude Opus 5.5 at the same setting.

• Multi-step autonomous operation of browsers and desktop applications
• Business process orchestration across SaaS tools
• Batch maintenance of records in CRM, ticketing, spreadsheets, and more

03 Professional Knowledge Work and Document Understanding

OpenAI reports that it shows a clear improvement over GPT-6 Sol in document understanding and multi-step workflows. GDP.pdf is 32.0%, essentially on par with Astra's 32.2%, compared with 28.0% for GPT-6 Sol. On medical evaluations, HealthBench Professional scores 64.2 (Astra 64.7, GPT-6 Sol 60.8), and HealthBench Hard scores 36.2 (GPT-6 Sol 30.1).

• Information extraction and analysis from long documents and PDFs
• Multi-step professional tasks and report generation
• Q&A in professional domains such as healthcare and finance

04 Factuality and Reliability

The share of responses containing factual errors at the low setting drops from 11.4% for GPT-6 Sol to 7.7%, the largest improvement this time; across all reasoning settings, its error-rate gap with Astra is within 1.9 percentage points. Third-party data points in the same direction: the AA-Omniscience hallucination rate drops from 60% to 54%. OpenAI also says it is more truthful when tools fail, more compliant with explicit constraints, and more candid about its own limitations.

• Analytical reports that require cited facts and less fabrication
• Honestly reporting when search or tools fail instead of making things up
• More trustworthy result self-reporting in long-chain tasks

---

Best Use Cases

Scenario	Description
Complex refactoring and codebase investigation	DeepSWE v1.1 75.2%, on par with Astra
Long-running coding agents	Cached input billed at 5%, low cost for long sessions
Computer-use agents	OSWorld 2.0 offline set 71.4%, 2.1 percentage points behind Astra
Business process automation	AutomationBench at medium 35.4%, 4.8 percentage points above the previous version
Cost reduction from Astra	Close to Astra on most coding and professional tasks, at one-fifth the price tier
Upgrade from GPT-6 Sol	Same price, across-the-board capability gains, cheaper caching

---

Differences Between GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Astra

Capability	GPT-6.1 Sol	GPT-6 Sol	GPT-6 Astra
Model ID	gpt-6.1-sol	gpt-6-sol	gpt-6-astra
Release date	September 29, 2026	September 22, 2026	September 3, 2026
AA Intelligence Index (v4.3.2)	52	48	53
DeepSWE v1.1	75.2%	68.8%	74.8%
OSWorld 2.0 (offline set, max)	71.4%	64.4%	73.5%
GDP.pdf	32.0%	28.0%	32.2%
AA-Omniscience hallucination rate	54%	60%	—
AA Index per-task cost	$0.72	$1.04	$3.26
Context window	1.05M tokens	1.05M tokens	1.05M tokens
Maximum output	128K tokens	128K tokens	128K tokens
Knowledge cutoff	April 30, 2026	April 20, 2026	April 30, 2026
Reasoning settings	low → max, five levels, no none	none → max, six levels	—
Chat Completions tool calling	Not supported	Supported only at none setting	—
Cached input relative unit price	5% of uncached input price	10% of uncached input price	—
Price tier	Same as GPT-6 Sol, 1/5 of Astra	Baseline	Highest tier
Positioning	Mainstream tier close to Astra	Previous mainstream tier	Flagship

Except where noted as AA, benchmark data are figures published by OpenAI. Actual billing is subject to the real-time pricing card at the top of the page.

---

FAQ

When was GPT-6.1 Sol released? September 29, 2026, released at OpenAI DevDay, one week after GPT-6 Sol launched. It became available the same day in paid accounts across the API, Codex, and ChatGPT Work. According to TechCrunch, GPT-6.1 Astra, which was expected to launch at the same time, did not ship.

What was upgraded compared with GPT-6 Sol? Three things. First, capability: DeepSWE v1.1 rises from 68.8% to 75.2%, OSWorld 2.0 offline set from 64.4% to 71.4%, and the AA Intelligence Index from 48 to 52. Second, reliability: the factual error rate at the low setting drops from 11.4% to 7.7%, and the AA-Omniscience hallucination rate drops from 60% to 54%. Third, cost: input and output unit prices are unchanged, cached input is halved, and the AA Index per-task cost drops from $1.04 to $0.72.

Do I need to change code when migrating from GPT-6 Sol? There are two places that will directly error. First, reasoning_effort no longer accepts none and minimal; the lowest setting is low, so code previously set to none must be changed. Second, the Chat Completions API no longer supports tool calling—GPT-6 Sol can still do function calling through Chat Completions at the none setting, but GPT-6.1 Sol cannot, so scenarios that need tools must use the Responses API. In addition, Artificial Analysis measured that at the same setting, it outputs 10% to 30% more tokens per task than GPT-6 Sol, so budgets should be loosened accordingly.

What are the context and output limits? The context window is 1.05M tokens, with a maximum input of 922K and a maximum output of 128K tokens, the same as GPT-6 Sol. Input supports text and images; output is plain text, with no support for audio or video. The knowledge cutoff is April 30, 2026.

How should reasoning effort be set? reasoning_effort supports five settings: low, medium (default), high, xhigh, and max; thinking cannot be fully disabled. Official benchmarks are mostly measured at high to max, while everyday tasks can use the default setting. Reasoning tokens are billed at the output price, so total cost rises noticeably at higher settings.

Are there areas where it lags behind? Yes. Its AA Intelligence Index is 52, below Claude Opus 5.5's 58 and Claude Sonnet 5.5's 56, and slightly below Astra. It is on the slow side: Artificial Analysis measured output at the max setting at about 56 tokens/second, below the median for its class, with time to first token exceeding 5 minutes, making it unsuitable for real-time scenarios. It still clearly lags behind Astra on the hardest science, biology, and cybersecurity evaluations, for example TroubleshootingBench at 47.96% versus 63.46%. There are also regressions in alignment: the coding deception rate is 1.50%, higher than GPT-6 Sol's 1.30% and Astra's 0.51%; the rate of "continuing when it should not" is 23.5%, higher than Astra's 17.4%. When a single prompt exceeds 272K tokens, the entire request's input and caching are billed at 2x, and output at 1.5x.

Which tools and features are supported? Streaming output, structured output, function calling, prompt caching, image input, and Batch are all supported; fine-tuning and predicted outputs are not. Official hosted tools under the Responses API cover web search, file retrieval, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search.

Should I choose GPT-6.1 Sol or Astra? For most coding, computer use, and professional tasks, choose GPT-6.1 Sol: performance is close while the price is only one-fifth. When you need the strongest scientific reasoning, the hardest cybersecurity or biology tasks, or the strictest requirements on deception rate and out-of-bounds behavior, choose Astra.

---

Why Choose SeaWhale AI to Use the GPT-6.1 Sol API

• No Overseas Account Required — Direct connection within China, no need to build your own proxy
• OpenAI-Compatible API — Integrate by changing just two lines in your existing code
• Cross-Vendor Comparison — Compare price and performance directly with Claude, Gemini, and Grok under the same account
• Unified Billing — Centrally manage usage and spend across multiple models

API

API integration

Model IDUse this value as the model in inference requests

gpt-6.1-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": "stable" }

gpt-6.1-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.

cURL

curl https://api.haijingai.com/v2/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <API_KEY>" \
  -d '{
    "model": "gpt-6.1-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

Python

from openai import OpenAI

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

stream = client.chat.completions.create(
    model="gpt-6.1-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)

Node.js

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-6.1-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 ?? '')
}