Top models from OpenAI, Anthropic and Google, for less money
Within three days Anthropic, OpenAI and Google released new models that come close to their best, at a lower price. What a task costs with AI is changing fast.
What happened
Within three days Anthropic, OpenAI and Google each released a model that comes close to their best, all three at exactly the same price: $2 per million tokens in and $10 per million tokens out. Tokens are the small pieces of text an AI model reads and writes, and you pay for each one.
Anthropic went first on 28 September with Claude Sonnet 5.5. Its price per token is the same as its predecessor Sonnet 5, but Anthropic says a task costs up to 30% less, because the model needs far fewer tokens for the same work. At the investment firm Balyasny it used about 121,000 tokens per answer, where Sonnet 5 needed 497,000.
A day later, at its DevDay event, OpenAI followed with GPT-6.1 Sol, an improved version of GPT-6 Sol, which had come out a week earlier at the same price. OpenAI says Sol comes close to its top model, GPT-6 Astra, at coding, operating a computer and office work, for a fifth of the price: Astra costs $10 and $50. In OpenAI’s own tests Sol tried to get around an explicit restriction, such as an access-denied message, in 23.5% of cases; for GPT-6 Sol that was 64.4%. The day before DevDay, The Wall Street Journal reported that OpenAI had cancelled the launch of GPT-6.1 Astra, because in tests that model was more deceptive and overstepped its brief.
Google closed the week on 30 September with Gemini 4 Argon, its new top model. The $2 and $10 are an introductory price: according to Google’s price list, it goes up to $4 and $20 from 1 December. Argon went first only to a select group of security specialists in its Fairwind programme. They get the model without the brakes on security work, so they can use it to find and fix vulnerabilities.
Near-Astra intelligence for a fifth of the price.
How OpenAI announced GPT-6.1 Sol, as quoted by The Next Web
The background
What you actually pay for
An AI model charges in tokens: what you put in, such as your question, a document and earlier messages, and what comes out. With all three, output costs five times as much as input. If a model thinks before it answers, that thinking counts as output too, even when you never see it, as Google’s price list spells out. That is why the setting for how thoroughly a model thinks decides much of the bill.
Text that keeps coming back, such as a standing instruction or a manual sent along with every question, can be held by the model for a while. Reusing it costs 95% less with all three.
How fast AI got cheaper
In March 2023 GPT-4, then the best model, cost around $37.50 per million tokens on average, according to the research group Epoch AI. In 2025 Epoch worked out that the price of reaching a fixed level falls by a factor of 9 to 900 a year, depending on the task. The steepest falls came in the most recent year, and Epoch itself calls it uncertain whether they will continue.
At the end of July OpenAI cut the price of its smallest model by another 80%. According to Reuters, open models from China, such as Z.ai’s GLM-5.2, come close to the same level for less money. Reuters also noted that while the price per token is falling, completing a task is getting more expensive as companies move from flat subscriptions to paying per use.
The gap with the top is narrowing
The real top models stay more expensive: GPT-6 Astra costs $10 and $50, Claude Opus 5.5 $4 and $20, and Argon will soon cost $4 and $20 too. But the gap in ability is shrinking. On the index of the independent Artificial Analysis, an average of ten tests, Sol scores 52 points and Astra 53. A task from those tests cost $0.72 with Sol and $3.26 with Astra. Sonnet 5.5 reaches 56 points, just below Opus 5.5 at 58.
Our analysis
Same price per token, a very different price per task
Three rivals landing on exactly $2 and $10 within three days is no coincidence. It is a price point buyers compare on, like a slot on a shelf. But the price tag says little about your bill. On Artificial Analysis’s tests a task cost $0.72 with Sol, $1.99 with Argon and $5.46 with Sonnet 5.5. Sonnet scored highest, but at its most thorough setting it wrote more than six times as many tokens as Sol.
What the coverage misses is that much of the “cheaper” is relative. Sonnet 5.5 costs the same per token as its predecessor, Sol the same as GPT-6 Sol a week earlier, and Argon will double in price after its introduction. What does change fast is what that price buys you. Our expectation: your bill won’t come down by itself. It only comes down when someone picks the model and the setting for each task, and work an agent does in many steps can even turn out more expensive.
Test with twenty to fifty real examples from your own work and measure what a finished task costs, retries included. Do the sums at the full price, not an introductory one. And set things up so you can switch models: at this price point the three are interchangeable for many tasks, and Anthropic has already announced Haiku 5.5.
What stands out
- Same price tag, different bill On Artificial Analysis’s tests a task cost more than seven times as much with Sonnet 5.5 as with Sol, at exactly the same price per token.
- A lower setting saves a lot At a low or medium setting Sonnet 5.5 beats Sonnet 5’s best score on several tests for about a tenth of the cost per task, Anthropic says.
- All three got brakes on security work Anthropic hands risky requests to an older model, OpenAI sometimes to an unnamed fallback model. Google gives Argon without those brakes only to selected defenders.
- The top model stayed on the shelf OpenAI cancelled GPT-6.1 Astra because in tests it was more deceptive and went further than it was allowed to. The cheaper Sol did ship.
What it means for you
Did you once work out that AI was too expensive for a process? Then it is time to do the sums again.
- Count per task, not per token One model uses far more tokens than another for the same work. Let a test with your own work show what a task really costs.
- Watch out for introductory prices For now Google charges half of what Argon will cost later. Run the numbers at the full price too, so nothing catches you out.
- Pick a model per task The most expensive model isn’t needed for every task. Test with your own work which model is good enough.
Where we come in · Advisory
We work out with you where AI now pays off
In a Discovery or Roadmap we look at each process: what AI can take over, what it costs and which model suits it. You get independent advice you can act on straight away.
Which process has been too expensive to automate until now?
Tell us what work you have in mind. We’ll gladly run the numbers with you, and we always reply within one working day.
