You pay for one AI subscription, and two new models at the same price ship the same week. You have neither the time nor the budget to test both on all your work.
GPT-6 Sol from OpenAI and Claude Sonnet 5.5 from Anthropic cost exactly the same at launch, $2 per million input tokens and $10 per million output tokens. Sonnet 5.5 leads on the office-work tests Anthropic published. GPT-6 Sol puts its cross-app automations front and center, at a very low cost per task.
The right answer mostly depends on the tool you already work in and how much you send it at once. Below are the prices line by line, the tests to read with care, and a pick for each profile, including if you don't code.
Two mid-tier models, released days apart
Both companies stack their models in three tiers. At OpenAI, GPT-6 Astra sits at the top, Sol in the middle, Luna at the bottom. At Anthropic, Opus 5.5 sits at the top, Sonnet in the middle, Haiku at the bottom.
Sol and Sonnet 5.5 hold the same spot, the model you run all day long. Neither targets the hardest tasks, which each company keeps for its top tier.
Our article on Claude Sonnet 5.5 covers its scores against Opus 5.5. The one on ChatGPT Astra breaks down OpenAI's top tier.
The price is identical, up to 272,000 tokens
The price sheets published at launch match to the cent. The sources are OpenAI's model page and Anthropic's announcement.
| Price per million tokens, at launch | GPT-6 Sol | Claude Sonnet 5.5 |
|---|---|---|
| Input | $2 | $2 |
| Output | $10 | $10 |
| Cache reads | $0.20 | $0.20 |
| Cache writes | $2.50 | $2.50 |
| Batch processing | 50% off | 50% off |
A token is a chunk of a word, the unit the API bills you for. The cache saves you from paying full price again for text you resend with every request, like an agent's instructions.
The difference hides above 272,000 tokens sent at once. OpenAI then doubles Sol's input price and multiplies its output price by 1.5, for the whole request. Sending an entire code repository or a big folder of PDFs lands there quickly.
A request with 300,000 input tokens and 10,000 output tokens costs $0.70 on Sol at the normal rate. With the surcharge, it goes up to $1.35 ($1.20 of input and $0.15 of output). For Sonnet 5.5, check the price of very long requests on Anthropic's pricing page before comparing.
The token price doesn't tell the whole story. Each company says its model burns fewer tokens to finish a task. Anthropic claims up to 30% lower cost per task than Sonnet 5, and OpenAI halves Sol's price compared with GPT-5.6 Sol.
The benchmarks, where each side picks its ground
Each company publishes the tests where it shines, and rarely compares against its rival's latest model. So read every number together with who published it.
What Anthropic measures
Anthropic put GPT-6 Sol in its own table. On office work and chart reading, Sonnet 5.5 comes out clearly ahead, according to its measurements.
| Test (Anthropic's measurements) | What it measures | Sonnet 5.5 | GPT-6 Sol |
|---|---|---|---|
| GDPval-AA v2.1 | Real tasks from 44 occupations, Elo-style score | 1,844 | 1,487 |
| AA-Briefcase v1.1 | Long office work, Elo-style score | 1,811 | 1,483 |
| Chartography | Reading charts, without tools | 61.6% | 53.6% |
On an Elo-style score, only the gap means something. About 360 points separate the two models on GDPval-AA, almost as much as between Sonnet 5 and Sonnet 5.5.
What OpenAI measures
OpenAI doesn't mention Sonnet 5.5 in its announcement. Its comparisons target Opus 5 and Fable 5, two older and pricier Anthropic models.
- AutomationBench, chains of tasks across 47 sales, support and finance tools. Sol scores 33.2% for $0.27 per task, against 26.9% for Opus 5 at 11 times the cost.
- DeepSWE v1.1, software engineering in real projects. Sol reaches 68.8%, 1.1 points behind Fable 5, for about 80% less per task.
- Factual reliability. On an internal test, Sol makes about half as many mistakes as GPT-5.6 Sol.
Both companies cite OSWorld, the test where a model operates a computer, but not the same version. Anthropic reports 80.1% for Sonnet 5.5 on OSWorld 2.1, OpenAI reports 60.5% for Sol on OSWorld 2.0 offline. These two numbers can't be compared, the tasks differ.
What you get out of this is a trend, not a verdict. Sonnet 5.5 looks stronger on documents and office work. Sol sells itself on business automations, at a very low cost per task.
Where you'll use them day to day
This is often what settles it. A model only helps you inside the tool you already work in.
- Claude Sonnet 5.5 runs on Claude.ai, open to everyone at launch according to Anthropic, and in Claude Code, the agent that codes in your terminal.
- GPT-6 Sol runs in ChatGPT Work and in Codex, OpenAI's coding agent, for Plus subscribers and above. It wasn't offered in regular ChatGPT chat at launch.
- Through the API, the model IDs are
gpt-6-solat OpenAI andclaude-sonnet-5-5at Anthropic.
If you code with an agent, the model follows the agent you picked. Our Claude Code course teaches you to work with Sonnet 5.5 in the terminal. Our Codex course covers the OpenAI agent that runs Sol.
The simplest move is to give both the same task, on your own files. Here are two minimal calls in Python, one per company.
from openai import OpenAI
import anthropic
task = "Summarize this quote in three lines for my client: YOUR_TEXT_HERE"
# GPT-6 Sol (reads the key from OPENAI_API_KEY)
sol = OpenAI().responses.create(model="gpt-6-sol", input=task)
print("GPT-6 Sol:", sol.output_text)
# Claude Sonnet 5.5 (reads the key from ANTHROPIC_API_KEY)
sonnet = anthropic.Anthropic().messages.create(
model="claude-sonnet-5-5",
max_tokens=1024,
messages=[{"role": "user", "content": task}],
)
print("Sonnet 5.5:", sonnet.content[-1].text)Compare the two answers, then the cost shown in each console. Ten real tasks teach you more than every table out there.
Which one to pick for your work
Neither wins everywhere. Here's the pick we'd make, profile by profile.
- You already code in Claude Code. Stay on Sonnet 5.5, it clearly improves on Sonnet 5 without changing how you work.
- You already code in Codex. Switch to Sol, OpenAI reports a big jump on FrontierCode at half the price.
- You produce documents, slide decks, analyses. Sonnet 5.5 leads on the office-work tests Anthropic published.
- You automate tasks across several apps. Try Sol, that's the ground OpenAI highlights with AutomationBench.
- You send very large volumes at once. Watch out for Sol's surcharge above 272,000 tokens.
- You don't want to pay for a subscription. Sonnet 5.5 was available on Claude.ai without a paid plan at launch, while Sol required at least the Plus plan.
In an automation tool like n8n, you don't even have to choose once and for all. Our n8n course shows how to plug a different model into each step of a workflow.
Frequently asked questions about GPT-6 Sol and Claude Sonnet 5.5
Is GPT-6 Sol cheaper than Claude Sonnet 5.5?
No, both cost $2 per million input tokens and $10 per million output tokens at launch, with the same cache pricing.
The difference shows up above 272,000 tokens sent in one request. OpenAI then doubles GPT-6 Sol's input price and multiplies its output price by 1.5.
Which one is better for coding?
Neither company publishes a clean coding comparison between the two. Anthropic reports 70.6% for Sonnet 5.5 on Terminal-Bench 4.0, OpenAI reports 68.8% for Sol on DeepSWE v1.1, two different tests.
In practice, the model follows the agent. Sonnet 5.5 runs in Claude Code, GPT-6 Sol in Codex.
Can you use GPT-6 Sol or Claude Sonnet 5.5 for free?
At launch, Claude Sonnet 5.5 was available to everyone on Claude.ai, on the web, iOS and Android.
GPT-6 Sol was limited to ChatGPT Plus, Pro, Business, Enterprise and Edu subscribers, in ChatGPT Work and Codex. Free users had access to GPT-6 Luna in the desktop app.
What are the API model IDs for GPT-6 Sol and Claude Sonnet 5.5?
The ID is gpt-6-sol on the OpenAI API and claude-sonnet-5-5 on the Anthropic API.
Sonnet 5.5 is also offered on Amazon Bedrock as anthropic.claude-sonnet-5-5, as well as on Google Cloud and Microsoft Foundry.