Here's the number that matters: 50.

That's where DeepSeek's newest model, V4 Flash 0731, landed on the Artificial Analysis Intelligence Index when it shipped on July 31 — a 10-point jump over the previous V4 Flash, which sat at 40. Fifty puts it one point behind GPT-5.6 Luna (51), dead level with Google's Gemini 3.6 Flash (50), and six points ahead of DeepSeek's own bigger, pricier V4 Pro.

Now the part that stings for everyone selling tokens: it costs $0.14 per million input tokens and $0.28 per million output. That's the price of a model that a week ago would have been called mid-tier — now delivering near-frontier scores.

The thesis: the cheap seats just moved up to the front row, and the pricing of every "small, fast" model on the market now looks negotiable.

🧠 Why This Matters

For most of the last two years, the deal was simple. If you wanted the smartest answers, you paid flagship prices. If you wanted cheap, you accepted a model that fumbled multi-step tasks. V4 Flash 0731 blurs that line hard.

It's a 284-billion-parameter mixture-of-experts model that only fires 13 billion parameters per token, which is why it can be this cheap and this fast while carrying a 1-million-token context window (OpenRouter lists it at exactly 1,048,576 tokens). On the agentic-work benchmark GDPval-AA v2, its Elo climbed to 1559, up from 1189 — the kind of leap that shows up when a model stops getting lost halfway through a task.

Independent developer Simon Willison, who tested it the day it dropped, didn't hedge:

"This may currently be the best value-per-intelligence model out there."— Simon Willison, simonwillison.net

When the person who has benchmarked half the models on the planet says that out loud, procurement teams start rewriting spreadsheets.

📊 Deep Dive

The Intelligence Index isn't one test — it's a blend of reasoning, coding, math, and knowledge evals rolled into a single number. Fifty is a serious score for something this small. Here's how V4 Flash 0731 stacks up against the models it's now crashing into:

  • DeepSeek V4 Flash 0731 — Index 50 · $0.14 in / $0.28 out · 1M context
  • Gemini 3.6 Flash — Index 50 · matched on score
  • GPT-5.6 Luna — Index 51 · $0.20 in / $1.20 out (after this week's price cut)
  • GLM-5.2 — Index 51 · one point ahead
  • DeepSeek V4 Pro — Index 44 · the bigger sibling it now beats by 6
  • Kimi K3 — Index 57 · the open-weights frontier, still 7 points clear

The eye-catcher is that comparison to GPT-5.6 Luna. OpenAI cut Luna's price by 80% on July 30 — from a launch price of $1.00/$6.00 down to $0.20 input and $1.20 output — a move widely read as a direct response to Chinese labs undercutting it on cost. V4 Flash 0731 landed the very next day and still comes in cheaper on both ends. As XenoSpectrum put it plainly:

"V4 Flash, at $0.14 for input and $0.28 for output, undercuts Luna on both metrics."— XenoSpectrum

On output tokens — where most real workloads rack up the bill — that's roughly a 4x price gap for a one-point difference in score.

⚠️ The Catch

A benchmark score is not a production deployment. A few things worth keeping in view before you rip out your stack.

First, one point on an index is easy to over-read. Fifty versus 51 is a rounding error on most tasks, but the shape of the errors differs model to model — the only way to know if V4 Flash handles your workload is to run it on your own evals, not the leaderboard's.

Second, Willison noted output quality improved sharply only when he pushed the reasoning level from default to high — which means more thinking tokens, which means the real-world cost sits above that headline $0.28. Cheap per token isn't the same as cheap per finished task.

Third, this is a hosted API from a company based in China, and for a lot of enterprises that alone dictates where data can and can't go. Weights are downloadable, so self-hosting is on the table, but that's a very different cost and ops conversation than pasting in a key.

🎯 What Happens Next

Watch the response times. OpenAI already blinked on price two days before this launched; the question is whether Google, Anthropic, and the rest follow at the small-and-fast tier or hold the line on the premium models where margins actually live.

Watch adoption on the coding and agent platforms too. A model that scores 1559 on an agentic benchmark at $0.28 output is exactly what tool-builders reach for when they're running thousands of automated steps and every fraction of a cent compounds.

🧩 Bigger Picture

Strip away the horse-race and the pattern is clear: the frontier keeps getting more expensive to reach, while the cost of yesterday's frontier keeps collapsing. A 50 on the Intelligence Index would have been near the top of the chart not long ago. Today it's the budget option.

That's the real story in the numbers — not who's ahead by a single point, but how fast "good enough for almost everything" is getting cheap. When near-frontier intelligence sells for cents per million tokens, the moat stops being the model and starts being everything around it: the product, the data, the distribution.

The leaderboard crown still changes hands every few weeks. The floor, meanwhile, just keeps rising — and it's the floor that ends up in your app.

Sources