The benchmark site OpenDesign Arena this week ran 13 AI models through the same batch of design tasks — building web apps, dashboards, mobile screens and landing pages. The top scorer was OpenAI's GPT-6 Astra, released September 3, averaging 82.7 points. Close behind, DeepSeek's V4.1 Flash scored 81.2 — a gap of 1.5 points.
The price gap is on another scale. Per the leaderboard data reported by Decrypt, Astra averaged $1.61 and 11.1 minutes per finished design; V4.1 Flash spent $0.023 and 5.3 minutes. Anthropic's Claude Fable 5.1 scored 80.3, took 12.8 minutes and cost $3.66 per design — the most expensive of the top three.
Among the remaining models tested — including Grok 4.6, Qwen 3.8-Max, Kimi K3, GLM-5.3 Flash and Gemini 3.8 Flash — every one scored lower than DeepSeek's entry and cost more to run. Of the 13 models, only GPT-6 Astra scored higher, and only by a point and a half.
[1]What the leaderboard actually measures
OpenDesign Arena scores out of 100: 30 points for whether the output meets the brief, 70 for layout, hierarchy, color and style fit. It only scores outputs that render as a working webpage — anything blank, broken or truncated gets a zero and is not retested.
That makes it a measure of reliable, deliverable, everyday design output, not general reasoning or coding skill. OpenDesign, the operator, is a company rather than an academic institution, and the board is self-reported data that no third party has reproduced. The question it answers is narrower than most leaderboards ask: which model should a working web designer use tomorrow.
A second set of numbers is worth attention. The board's delivery rate — the share of outputs judged ready to hand off without revision — came in at 60% for Astra, 57.7% for V4.1 Flash and 56.7% for Fable 5.1. In other words, what a designer actually feels is a 2.3-percentage-point delivery gap, not a 70-fold price gap. The price difference is spectacular and the quality difference is subtle; both numbers are true.
[1]Where the cheap comes from
DeepSeek's technical report explains the cost structure. According to VentureBeat's analysis, V4.1 Flash carries 552 billion backbone parameters but activates only 8 billion per token while reading a prompt and 16 billion while generating a response — a design DeepSeek calls a Causal Encoder-Decoder, splitting 40 Transformer layers into a 20-layer encoder and a 20-layer decoder. With Compressed Sparse Attention 2 and FP4 KV caching, the global KV cache falls to 890 bytes per token, roughly a quarter of the previous V4-Flash.
But "Flash" no longer means small. The previous V4-Flash backbone was 284 billion parameters; this one is 552 billion, a jump of about 94%, and the report separately lists 196 billion parameters in Engram conditional-memory modules. One Hacker News commenter asked whether it is "not really flash anymore" — sparse activation and cache design lower serving costs, but the near-doubling of total weights raises the hardware bar for self-hosting, even though the weights are MIT-licensed.
DeepSeek itself added a reservation. The technical report states that the new architecture creates robustness boundaries it has "not fully characterized": sparse-selection errors and the approximate state reconstruction of SWA Bounded Replay could degrade capability in untested edge cases — particularly sparse retrieval over very long contexts and cache-resumption boundaries. That is the publisher's own language, not a critic's conjecture.
[2]A repeating pattern
This is not DeepSeek's first pass at near-flagship quality for pocket change. Weeks earlier, its V4 Pro model landed within 5% of Claude Fable 5 on a separate benchmark comparison at a fraction of the price. V4.1 Flash is also becoming the new center of gravity for DeepSeek's API: documentation says legacy V4 Flash requests are already served by the new model, and V4 Pro will route to V4.1 Flash after September 14 until a future V4.1 Pro arrives. Meanwhile, DeepSeek is recruiting engineers in Beijing to build its own Code Harness — aiming to own the full agentic stack rather than just supply the model underneath it.
[1][2]What buyers actually care about
Cheap changes the negotiating structure; it does not necessarily change the purchasing order. VentureBeat's Pulse data suggests only 31% of surveyed enterprises name cost per million tokens as a primary infrastructure success metric, while 51% chose uptime and reliability and 39% chose developer productivity.
The real pressure therefore lands on whoever sells enterprise products at a premium. In the same week, OpenAI has been pitching tailored versions of ChatGPT to financial institutions and large enterprises — and the selling points of its financial-services edition and data agent are not token prices but data licensing, permissions and workflow. Open weights plus aggressive pricing will not immediately take those contracts, but the combination will be on the table at every renewal.
OpenDesign's board answers a narrow question: which model to use for web design tomorrow. What it cannot answer is the unknown DeepSeek flagged in its own report — how an architecture engineered for savings fails at the edges. A 1.5-point quality gap is nearly invisible in most everyday tasks; a 70-fold price gap is visible in every procurement contract; and the robustness bill always arrives last.
[2]