DeepSeek Flash Pricing and Google's Willow Quantum Milestone

Two things happened today. DeepSeek released a new Flash checkpoint at 15 cents per million input tokens, and Google cut a specific class of quantum errors by 40 times on its Willow chip.

One of those affects your invoice this month. The other affects your encryption in a year nobody is willing to name out loud.

DeepSeek Flash at 15 cents per million tokens

DeepSeek V4.1 Flash got a new checkpoint, and the headline is the price. 15 cents per million input tokens. The weights are under an MIT license, which in plain words means anyone can download them, change them and use them commercially.

It also takes over a million tokens of context. So you can hand it a pile of documents and it will not panic.

The benchmark numbers are self-reported, so adjust your eyebrows accordingly. The previous Flash checkpoint scored 74.2 on their internal agent eval. The new one is reported at 77.

On Humanity's Last Exam it lands at 42.6 without tools and 44 with tools. On Terminal Bench it scores 42.3, ahead of GPT 5.6 at 39.9 and well behind Opus 5 at 51.8.

I have reached the point where every new benchmark chart reads like a restaurant reviewing its own food. The pattern still tells you something, though. A model this cheap is now close enough to the expensive ones that the gap stops mattering for most real work.

There is a catch, and it is time. In one side-by-side comparison the DeepSeek run was about 26 times cheaper but took roughly three times longer to finish. That is a real tradeoff, not a rounding error.

So treat it like batch work. Anything a human is waiting on, like a chat reply or a live call, needs speed. Anything that runs overnight does not care if it takes nine minutes instead of three. Document processing, classification, summarising yesterday's tickets, enriching a spreadsheet: all of that can run slow and cheap while you sleep.

The part I find most interesting is the boring part. DeepSeek is charging 75 cents per 1,000 search-and-fetch API calls. OpenAI and Anthropic charge around ten dollars for the same thousand. That is roughly 7.5% of the price.

Fetching a web page is not a thinking task. It is plumbing. There is no deep reasoning happening when a model opens a URL and reads the text back. If a chunk of your bill is search and fetch, that line item is a gift sitting in plain sight.

The honest conclusion for a small business: most of what you ask AI to do does not need frontier intelligence. Sorting emails, pulling fields out of a PDF, writing a first draft of a reply. These were solved problems two model generations ago. You are paying premium rates out of habit.

What I would do is split the work by task, not by brand loyalty. Pick one strong model for the hard reasoning steps and route the rest to something cheap. If you are building AI agents that run dozens of steps per job, the cheap steps are where the money hides.

And test it yourself before you switch anything. Run your actual tasks, not somebody's leaderboard. Your work is weirder than any benchmark.

Google's Willow chip and the magic state milestone

Now the other end of the spectrum. Google says it reduced magic state errors by 40 times using its Willow quantum chip.

Quick plain-English version. Quantum computers need special prepared ingredients called magic states to run anything genuinely useful. Making those ingredients cleanly is hard, and they come out full of errors. Google demonstrated a filtering method that throws away the bad ones, with about 80% of attempts surviving the check.

Here is the number that explains why anyone cares. Projections say 90 to 95% of a future quantum machine's qubits would be spent just manufacturing magic states. Almost the entire machine exists to feed the tiny part doing the actual work.

So most of a quantum computer's job right now is keeping the quantum computer honest. Cut that overhead and you do not just get a faster machine. You get a machine that needs far less hardware to be useful at all.

Why it matters beyond physics labs: the thing everyone quietly worries about is encryption. A lot of security today rests on the idea that breaking it would take an absurd amount of time. A working quantum machine changes that arithmetic. Not today, not next year, but the direction of travel is clear.

The useful upside is real too. Chemistry and physics simulations that are impossible now become possible, and that is where the genuinely new science comes from.

What should you do about it? Almost nothing. This is not a procurement decision for a 20-person company in 2026. Nobody is selling you a quantum server rack and nobody should.

The one thing worth keeping in the back of your head is data you are storing for the long haul. Contracts, medical records, anything that still needs to be secret in fifteen years. At some point your software vendors will start talking about post-quantum encryption, and when they do, it is a fair question to ask rather than a sales pitch to ignore. Just do not let anyone sell you a quantum-proof subscription this week.

Cheap now, strange later

These two stories sit on completely different clocks. DeepSeek's pricing is something you can act on by Friday afternoon. Google's magic states are something a handful of labs will argue about until the hardware catches up.

But they point the same way. The expensive, scarce thing keeps becoming the cheap, available thing. Intelligence per dollar is falling fast enough that the bottleneck is no longer the model. It is whether you have written down what you actually want done.

That is the unglamorous truth under all of this. If your process is vague, a smarter model will produce vague output faster and for less money.

My prediction: one of the big labs responds to this pricing within a few weeks, cuts rates, and calls it a commitment to accessibility. I will take the discount anyway.