DeepSeek shipped the official V4-Flash on July 31 — open weights, MIT license, and a technical report, all on day one.
The striking part isn't that the score is high. It's that the score and the price arrived together: 50 on the Artificial Analysis Intelligence Index, at $0.14 / $0.28 per million tokens.
The day before, OpenAI had cut GPT-5.6 Luna by 80%. In a single day, the token price war opened.
It got better without getting bigger
The key facts, per DeepSeek's release and technical report:
- DeepSeek-V4-Flash-0731 is the official release, superseding the preview, with substantially enhanced agentic capabilities.
- The architecture didn't change: the same efficient MoE core (~284B parameters, ~13B active), a speculative decoding module attached, 1M context.
- The price didn't change either — same as the preview.
Which means it got stronger not by scaling up, but through post-training.
The result: the smaller Flash beats DeepSeek's own larger V4-Pro preview on Terminal-Bench 2.1 (82.7), DeepSWE (54.4), Toolathlon (70.3), and Cybergym (76.7).
Artificial Analysis puts it at 50 on the Intelligence Index — a 10-point jump over the April V4 Flash, 6 points ahead of V4 Pro, and among the top 3 open-weights models.
The price is the payload
Per million tokens (input / output):
| Model | Input | Output |
|---|---|---|
| DeepSeek V4-Flash | $0.14 | $0.28 |
| GPT-5.6 Luna | $0.20 | $1.20 |
| GPT-5.6 Terra | $2 | $12 |
Artificial Analysis reports DeepSeek completing equivalent benchmark tasks at roughly 105x lower cost than Fable (some independent runs put the range at 105–137x).
One comparison making the rounds: Opus 5 shipped just 7 days earlier. On one task, Opus 5 got it right in one shot where DeepSeek took three — but DeepSeek did it in ~900 lines versus ~3000, and the run cost one cent.
Be honest: cheap tokens ≠ cheap outcomes
Here's the counterpoint, and it matters.
A developer hid 105 bugs across two real codebases. V4-Flash fixed 8 — last place. But the same run cost $0.61, versus $68.08 for Fable 5. Note the unit: cost per task, not per million tokens.
That points at a real shift in measurement:
Cheaper per token can be misleading. If a model needs more turns, the total cost per task can end up higher.
So the right question isn't "which model is cheaper per token." It's "what did it cost to actually get this done, and in how many turns?" Watch the verifiable outcome, not the proxy metric.
Open weights + MIT is a distribution move
The most strategically loaded part isn't performance or price — it's how it shipped: weights open on day one, MIT license, technical report included.
The timing wasn't lost on observers: in the same week Washington floated AI controls and Congress introduced a kill-switch bill, a frontier-class model was published as free software.
Stack that on OpenAI's 80% cut the day before — "an 80% price cut isn't a discount, it's a weapon" — and the shape of the fight is clear: frontier intelligence is being pushed toward zero, and open weights plus near-zero pricing is redrawing who controls distribution.
Three takeaways
- Getting better no longer requires getting bigger. Same structure, same price, post-training pushed Flash past its own Pro.
- Measure cost per task, not price per token. Cheap tokens don't guarantee cheap results.
- Open weights + near-zero pricing is a declaration of war on distribution. Who gets to price "access to the frontier" is no longer settled by a handful of labs.
Two implications for anyone building. First, when the best model changes weekly and prices can drop 80% overnight, you shouldn't be hand-picking models — you want a routing layer that trades off cost and capability for you. That layer gets more valuable as the price war escalates.
Second, when intelligence trends toward free, the scarce thing is no longer intelligence. It's judgment, and the taste to point it somewhere that matters.
Based on DeepSeek's official release (DeepSeek-V4-Flash-0731 weights, MIT license, technical report), Artificial Analysis benchmarks and cost comparisons, and independent developer tests discussed publicly. Data and specs per official and third-party sources.
Top comments (0)