The Cheaper AI Model Can Cost You More: Claude Opus 5.5 vs GPT-6 Astra, Sol and Luna

Quick answer This week Anthropic released a new AI model, Claude Opus 5.5, priced at less than half of OpenAI's top model, GPT-6 Astra. An independent tester also ranks it the top model overall. But in its most thorough mode it writes about four times as much to finish each job, so a single job can end up costing more, not less. The cheaper price list does not guarantee a cheaper bill. Before switching, test both on your own real tasks and compare what each finished job costs.

What Was Released

Two of the biggest AI companies released new models within weeks of each other: OpenAI, the company behind ChatGPT, and Anthropic, the company behind Claude.

• OpenAI, GPT-6 Astra (September 3): its most capable model and its most expensive.
• Anthropic, Claude Opus 5.5 (September 22): a new model that Anthropic says matches Claude Fable 5.1, its top publicly available model, on most work, while costing 40% less to run than Anthropic's previous version (Anthropic).
• OpenAI, GPT-6 Sol and GPT-6 Luna (also September 22): two cheaper versions of GPT-6. Sol is the mid-range option and Luna the budget one (OpenAI).

AI companies charge by the amount of text the model reads and writes, measured in "tokens" (a token is roughly three-quarters of a word). Prices are usually quoted per million tokens, with writing costing more than reading. Here is what each model charges for the text it writes:

Price for text written · list prices, September 2026

Model	Price per million tokens written	Best described as
GPT-6 Astra	$50	OpenAI's premium model
Claude Opus 5.5	$20	Anthropic's new all-rounder
GPT-6 Sol	$10	OpenAI's mid-range model
GPT-6 Luna	$0.50	OpenAI's budget model

Looking only at this table, Claude Opus 5.5 looks 60% cheaper than Astra. That is where most comparisons stop, and it is where they go wrong.

The Hourly-Rate Trap

Think about hiring two contractors. One charges $50 an hour, the other $20. The second looks like the obvious choice. But if the $50 contractor finishes the job in one hour and the $20 contractor takes four hours, the "cheap" one costs you $80 and the "expensive" one costs you $50.

AI models work the same way. Modern models "think" before they answer: they write out their reasoning step by step, and you pay for every word of that thinking. Some models think far longer than others to reach an answer.

Artificial Analysis, an independent company that tests AI models, measured how much each model writes to finish a task in its test suite, with Opus 5.5 at its maximum setting:

Writing per task · Opus 5.5 at maximum setting · source: Artificial Analysis

Model	Tokens written per task	Price per million	Cost of that writing per task
Claude Opus 5.5	about 119,000	$20	about $2.38
GPT-6 Astra	about 27,000	$50	about $1.35

The last column is my own calculation: the amount written multiplied by the price. It leaves out smaller costs, so treat it as a rough guide rather than an exact bill. But the picture is clear. At this setting, the model with the lower price costs about 75% more per task, because it writes about four times as much.

Thinking Longer Also Means Waiting Longer

There is a second cost that never shows up on the invoice: time. A model that writes four times as much usually takes longer to reply. For a report generated overnight, that doesn't matter. For a chatbot answering a customer, or a feature where a user is watching a loading spinner, it matters a lot.

So before comparing prices, ask a simpler question: can this task wait? If the answer is no, the most thorough modes of the most powerful models are often the wrong tool, however good they are.

"Effort" Is a Dial You Control

This doesn't make Claude Opus 5.5 a bad deal. Most of these models let you choose how hard they think, often called the "effort" level. Opus 5.5 has five settings, from low to maximum, and the numbers above come from the maximum setting.

At lower settings the picture changes. Rogo, one of the companies that tested Opus 5.5 before launch, says that even on its lowest setting, Opus 5.5 did better than Anthropic's previous version running on a high setting, while writing 60% less. Another tester, Walleye Capital, says the lowest setting "largely solved" its test (Anthropic). Anthropic chose which customer quotes to publish, so treat these as promising rather than proven. They do show where to start testing.

The same logic works in OpenAI's favour too. Its new mid-range model, Sol, scores slightly lower on a coding test than the version it replaces. On a results chart, that looks like a step backwards. But according to OpenAI's own figures, it costs less than half as much per finished task: about $2.74 against $6.46 (Digital Applied). Giving up a little quality for a much lower bill is often the right trade.

So Which One Is Actually Smarter?

On overall ability, Claude Opus 5.5 comes out on top, and it's an independent tester saying so, not just Anthropic. Artificial Analysis ranks it first overall, ahead of GPT-6 Astra.

Three caveats:

• Astra is still the best at advanced maths and science. Its scores on expert-level maths and science tests are among the highest published, and Anthropic did not publish comparable results (OpenAI).
• The gap shrinks when someone neutral checks. On one test where Anthropic reported a clear lead, the independent re-test came out as an exact tie.
• Neither chart includes the other company's new model. Because both launched on the same day, Anthropic's chart compares against OpenAI's older Sol, not the new one (OrcaRouter), and OpenAI's charts leave out Opus 5.5 (Digital Applied). The comparison you actually need appears in neither.

Which One Should Your Business Use?

Based on the published results, this is where I would start. Treat it as a starting point to test, not a final answer.

Where to start testing · based on published results

If the job is...	Start with	Why
Sorting, tagging, pulling details out of documents	GPT-6 Luna	Very cheap and fast; good enough for simple, repetitive work
Large volumes of everyday work, including code	GPT-6 Sol	A fifth of Astra's price, built for everyday coding and professional work
Office work, writing software, AI assistants that act on your behalf	Claude Opus 5.5, medium setting	Strongest all-round results, and, by Anthropic's own testing, among the hardest to trick with malicious instructions hidden in emails or web pages
Advanced maths, science, research	GPT-6 Astra	Clearly the best in these areas, and it writes far less per task

The smartest setup usually isn't choosing one model. A cheap model handles the simple majority of requests and passes the hard ones to an expensive model. With the budget option now about 100 times cheaper than the premium one, that approach can save a lot. And whichever model you use for an AI assistant that takes actions, limit what it is allowed to do, and have a person approve anything that can't be undone.

Before You Switch: A Simple Test

You don't need to be technical to run this. Your developers can do it in a day or two:

1. Pick 30 to 50 real tasks your team actually needs done. Use real work, not demo examples.
2. Run them through each model you are considering, at the settings you would actually use.
3. Check each result and mark it as done correctly or not.
4. Divide the total bill by the number of tasks done correctly. That is your real cost per job.
5. Note how long each one took, if anyone will be waiting for the answer.

Dividing by successful tasks matters. A cheap model that gets a third of the work wrong isn't cheap once someone has to redo it.

Key Takeaways

• Claude Opus 5.5 is the strongest all-round AI model this week, according to an independent tester.
• It isn't automatically the cheapest to use. At its maximum setting it writes about four times as much as GPT-6 Astra per task, which cancels out its lower price.
• The effort setting matters as much as the model. Try the medium setting before the maximum.
• Match the model to the job: Luna for simple work, Sol for volume, Opus 5.5 for office work and assistants, Astra for advanced maths and science.
• Judge by cost per finished job, measured on your own work, not by the price list.

Has your team measured what a finished AI task actually costs you, or are you still comparing price lists?

Sources

• Anthropic, Introducing Claude Opus 5.5
• OpenAI, GPT-6 Astra: A new generation of intelligence
• OpenAI, Introducing GPT-6 Sol and Luna
• Artificial Analysis, Claude Opus 5.5 takes the top spot on the Intelligence Index
• Digital Applied, GPT-6 Sol and Luna: API prices, benchmarks and trade-offs
• OrcaRouter, Claude Opus 5.5 vs GPT-5.6 Sol: two launch tables diverge

Update, October 4, 2026: Choosing the model is the smaller decision. How you configure the agent around it matters more, and I covered that in How to Configure AI Coding Agents to Work Like an Engineering Team.

You might also like GPT-6 Astra and the AGI Era: What Actually Changed and Shipping AI Features in Production: GPT-4o Inside a Live Platform.