On September 3, OpenAI launched its most expensive model. Seven days later, DeepSeek launched one that costs 67 times less and whose weights anyone can download for free.
The lazy takeaway is "the cheap model caught up with the expensive one." It hasn't, and the numbers themselves show where it falls short. The useful takeaway is different: with this price gap, choosing a single model for the entire company has become wasteful.
The two prices, side by side
Figures per million tokens, the billing unit. One million tokens is roughly equivalent to 700,000 words.
- GPT-6 Astra, launched on September 3: $10 input and $50 output. Cached input at $1, cache write at $12.50.
- DeepSeek V4.1 Flash, launched on September 10: $0.15 input and $0.60 output off-peak, cached input at $0.003. During peak hours, prices double.
Off-peak, the gap is 67 times on input and 83 times on output. At peak, it drops to 33 and 42 times. It's still a different order of magnitude.
And there's a detail that changes the conversation: V4.1 Flash's weights are on Hugging Face under an MIT license, with commercial use allowed. The backbone has 552 billion parameters, of which only 8 billion are active on input and 16 billion on output — that's why it's so cheap.
Where the cheap model performs well, and who's saying so
DeepSeek reports wins over both American flagships: 74.2 against 74.0 for Claude Opus 5 and 73.0 for GPT-5.6 Sol on a software engineering benchmark; 88.1 on a security benchmark; 54.8 on an automation benchmark.
The caveat is significant and must be stated: these evaluations were run by DeepSeek itself, and no one has independently replicated them. A manufacturer's numbers about its own product are a starting point, not proof. They're worth taking as a sign that the gap has narrowed, not as a measure of by how much.
Where the expensive model still wins, by a wide margin
On the benchmarks where the American models lead, the gap isn't measured in decimals.
- Terminal-Bench 3.0: 43.3 for Opus 5 versus 30.0.
- Terminal-Bench 4.0: 51.8 for Opus 5 versus 31.2.
- GPQA Diamond and SEC-Bench Pro: GPT-5.6 Sol leads.
These benchmarks measure long tasks with many chained steps, where a mistake on the third step ruins the other ten. It's exactly the kind of work where paying a premium is justified.
How to sort your work by requirement
The criterion isn't the subject matter, it's the cost of an error and the length of the chain. Four buckets cover almost everything:
- High volume, cheap error, short response. Classifying messages, summarizing, extracting data from forms, suggesting a reply that a person reviews. This is the bulk of a small company's usage, and it's where the cheap model does the job.
- Long task, many steps, expensive error. Migrating a system, touching production code, reconciling financial data. Use the expensive model, and still add human review.
- Text that goes to the client under your name. Decide based on perceived quality, not cost per token: the token's price is irrelevant next to the cost of a wrong answer signed by your company.
- Sensitive data. This bucket isn't decided by price at all, and it's the subject of the last section.
In practice, most companies find that 80% of their volume falls into the first bucket, and that they'd been paying fourth-bucket prices for everything.
The math, with real numbers
A medium-sized automation consuming 50 million input tokens and 10 million output tokens per month:
- With the expensive model: $1,000 per month.
- With the cheap one, off-peak: $13.50 per month.
Even small-scale usage — 2 million input and 400,000 output tokens — comes out to $40 versus $0.54. The difference rarely decides anything at small scale; it decides when you want to turn automation on everywhere instead of choosing where to turn it on, and that's what the new pricing makes possible.
Before switching anything, it's worth revisiting what caching does to this math, covered in AI pricing: the discount you pay for with your data — the cache discount only applies if the beginning of the request is identical across calls.
What no price can solve
The $0.15 is the price of DeepSeek's interface. Using that interface means sending your text to the company, under its terms. Open weights are a different matter: they mean you can run the model on your own infrastructure, without sending anything out — and then the math stops being per token and becomes per machine, as we covered in open and free AI models: how much it costs to run them in-house.
The question of who actually processes your text has become more important, not less, and it's addressed in who really processes the text you send to AI. A low price doesn't change contract clauses or the legal basis for data processing.
Setting this up separated by requirement, with the right model in each bucket, is what we do at AI implementation for businesses.
Sources
Launch and pricing table for DeepSeek V4.1 Flash per VentureBeat's coverage from September 10, 2026, which is also the source for the peak and off-peak figures, the MIT license for the weights on Hugging Face, the parameter count, and the note that the cited evaluations were run by DeepSeek itself. GPT-6 Astra pricing per OpenAI's public pricing table, with a launch date of September 3, 2026.
Frequently asked questions
How does DeepSeek V4.1 Flash pricing compare to GPT-6 Astra?
Per million tokens off-peak, V4.1 Flash charges $0.15 for input and $0.60 for output, with cached input at $0.003. GPT-6 Astra charges $10 for input and $50 for output, with cached input at $1. The off-peak gap runs about 67x on input and 83x on output; during peak hours DeepSeek’s prices double, narrowing the gap to 33x and 42x.
Is the cheap model already as good as the expensive one?
It depends on the task, and the numbers we have don’t support that claim either way. DeepSeek has published narrow wins on software engineering, security, and automation benchmarks, but those evaluations were run by DeepSeek itself and haven’t been independently replicated. On long-task benchmarks like Terminal-Bench, the American models lead by a wide margin: 51.8 versus 31.2 in version 4.0.
How do I decide which model to use for each case?
Split by the cost of an error and the length of the step chain, not by subject matter. High-volume work with a cheap error and a short answer — classifying, summarizing, extracting data — goes to the cheap model. Long tasks with many chained steps and an expensive error go to the expensive model, with human review. Text that goes out under the company’s name is decided by perceived quality. Sensitive data is decided by where it’s processed, not by price.
Does open weights mean I can use it for free in my company?
The MIT license allows commercial use, modification, and redistribution, so yes, you can use the weights. But running the model requires your own infrastructure, and the cost shifts to hardware, power, and operations instead of per token. Using DeepSeek’s paid interface is cheaper and simpler; running it yourself solves the question of where your text ends up.
Is using a Chinese model a data-protection problem?
The issue isn’t nationality, it’s where the data goes and under what terms. Whenever you use any provider’s interface, your text is processed by that provider. If the data is sensitive, the decision comes down to the contract, the legal basis for processing, and whether you can run the model on your own infrastructure — something open weights make possible.
Is it worth switching models right now?
It’s worth measuring before switching. Pull a month of real usage, see how much of the volume is simple tasks, and run the numbers at both price points. In most small companies, the bulk of the volume is simple tasks, and the savings show up without touching the critical tasks, which stay on the expensive model.