A budget model that costs about 60% less per task than OpenAI's GPT-5.6 Luna now trails it by a single point. That's the short version of DeepSeek V4 Flash "0731", and it's the strongest argument yet that the AI price war has moved off the billing page and onto the benchmark table.
Put another way: you get nearly the same output for about 40 cents on the dollar — and that's measured after OpenAI already slashed its own prices by 80%.
What happened
DeepSeek shipped V4 Flash "0731" as an upgrade to the budget model it launched in April 2026. The Artificial Analysis Intelligence Index — a single score that ranks AI models the way a standardized exam ranks students — puts the new build at 50 points, ten more than the April version. In this league, ten points is a jump, not a nudge.
The awkward part for OpenAI is the company it's now keeping. GPT-5.6 Luna, OpenAI's own budget model, sits exactly one point ahead. So you're paying roughly 60% less per task for a model that's a rounding error behind on a composite benchmark. On that same index, DeepSeek takes the top spot for price-to-performance ratio.
Why the bill is smaller: caching and fewer tokens
Two mechanics do most of the work. First, DeepSeek offers a 98% cache discount, above the 90% that's become the industry standard. Caching is like a barista who remembers your order: the second time you ask for the same thing, the shop doesn't start from scratch, so it charges you almost nothing. At 98% off, reused context costs a fraction of the full rate — and long, repeated prompts are exactly what business users run all day.
Second, the new model burns 12% fewer tokens than its predecessor. A token is a chunk of text, roughly three-quarters of a word, so fewer tokens means a shorter meter running. Neither trick makes the model smarter. Together they make the same work cheaper, which is the entire pitch.
Better at real work, not just quiz questions
DeepSeek says the model improved in every category it tested, with the biggest gains in agentic tasks — jobs where the AI takes several steps on its own instead of answering one question and stopping. Think of the difference between a calculator and an intern: one returns an answer, the other works through a process.
On GDPval, a benchmark built around messy real-world office work, it climbs from 1,189 to 1,559 Elo points. Elo is the chess-style rating system, where higher means it wins more head-to-head comparisons. It also hallucinates less often — meaning it makes up fewer facts, which is the failure mode that costs people the most time.
The engine underneath didn't change
Same architecture: 284 billion total parameters with 13 billion active, plus a one-million-token context window. Parameters are the model's stored knowledge; active parameters are the small slice it actually switches on at any moment. Picture a huge library with a compact reading room — big on knowledge, modest on running cost. The context window is the desk you can load, and one million tokens is roughly a stack of novels you can put in front of it in one go.
The weights are available under an MIT license on Hugging Face, which means you can download, run and modify the model yourself rather than renting it by the call.
What it means for you
You don't need to care about Elo points. You need to care that a task you'd pay a dollar for elsewhere now costs about 40 cents — and that the quality gap is a single point on a composite index.
At home. Household admin is death by a thousand small tasks: comparing insurance letters, drafting a complaint email, planning a trip. Cheap per-task AI means you stop rationing it. You ask, you get a draft, you fix two lines, you're done.
At work. The million-token window is the practical part. You can drop an entire report, contract or meeting transcript in and ask for the summary or the risk list. Before, that meant chopping documents into pieces and losing the thread; now it's a single paste.
In business. Cost per task is the number that decides what you can automate. Support replies, product descriptions, lead research — when each one gets 60% cheaper, work that was too low-margin to bother with suddenly isn't. Generate a hundred drafts, keep the five good ones.
For study. Tutoring works best when you can ask stupid questions without embarrassment and without a meter ticking. A cheaper model means you can ask about the same concept five different ways until it clicks, and the 12% token saving adds up because study sessions are long.
For creativity. Ideas are a numbers game. Cheap iterations mean you explore thirty directions instead of three, and the best one is usually somewhere near the thirtieth. A model that hallucinates less also gives you fewer confident, useless detours.
For side income. If you're building a small tool or service on top of an API, your margin is your model bill. A 60% cut changes which ideas are worth shipping — a product that costs you 40 cents a run is a business; the same product at a dollar a run is a hobby.
And if you'd rather skip the API entirely, there are free tools that cover a lot of the same ground — we keep a running list at mykreatool.com.
How to try it right now
1. Start free with the weights. Go to Hugging Face and search for DeepSeek V4 Flash. The MIT license means downloading and using it costs nothing. Look for "0731" in the model name — the April build is still floating around, and it scores ten points lower.
2. Check the version label before you test. If you're using a hosted chat that lists the model, confirm the string says 0731. Testing the old build and blaming the new one is the most common mistake people make with a release like this.
3. Run the same prompt side by side. Take one real task from your week — not a riddle — and run it on V4 Flash and on GPT-5.6 Luna. Judge the output yourself. A one-point index gap is invisible on most day-to-day work.
4. Be realistic about hardware. 284 billion total parameters isn't laptop material, and the 13 billion active ones still need real GPUs. Self-hosting buys you privacy and control, not a smaller bill, unless you already have the machines.
5. Use the cache on purpose. Keep a stable, long block of instructions or reference documents at the front of your prompts. Reused context is what the 98% discount rewards, and you throw it away if you rewrite your system prompt on every call.
6. Watch your token counts. Compare tokens used per task before and after the switch. That's the number that actually shows up on the invoice.
Upsides and what changes
The obvious upside is arithmetic: 60% off per task, on a model one point off the leader. The less obvious one is competitive pressure. OpenAI already cut prices by 80% and still gets undercut — that tells you where the floor is heading for everyone else.
The 98% cache discount raises the bar for the whole market, since 90% was the norm until now. Expect rivals to match it, and expect cache-aware prompt design to become a normal skill rather than a party trick.
The MIT license is the third shift. Weights you can download and run change the privacy conversation — teams in regulated industries that couldn't send data to an API suddenly have an option — and they make lock-in harder to defend when a contract comes up for renewal.
The real message here is that a one-point gap is no longer a moat. If the gap is a point and the price gap is 60%, the buyer starts asking why they're paying the premium.
Limitations
Honest version: one point behind is still behind, and a composite index is a lab exam, not your job. Benchmarks like GDPval test structured office tasks; your weird PDF from an old office scanner may behave very differently. The 60% saving is per task, so a model that needs two attempts to get something right can erase it — the index doesn't measure retries or the cost of a bad draft you have to fix. The 98% cache discount only fires when you actually reuse context; one-off prompts pay full price. The 12% token saving is measured against DeepSeek's own previous version, not against GPT-5.6 Luna. Hallucinations dropped but didn't disappear, so anything involving numbers, citations or legal language still needs a human read. And free weights aren't free compute: 284 billion parameters is a serious hardware bill if you decide to self-host.
Conclusion
DeepSeek V4 Flash "0731" is the clearest sign yet that the budget tier is where the real fight is: 50 points on the Artificial Analysis Intelligence Index, one point behind GPT-5.6 Luna, roughly 60% cheaper per task, with a 98% cache discount, 12% fewer tokens and a one-million-token window under an MIT license. For most people the takeaway isn't "migrate everything today." It's "stop paying a premium you can't measure."
Your one action for today: take the most repetitive AI task in your week, run it once on V4 Flash and once on whatever you use now, and write the output quality next to the cost. Ten minutes of work, and you'll know whether that one-point gap is worth 60%.
Comments 0