OpenDesign’s new model benchmark has a clear winner, and it is not the model everyone expected. GPT-6 Astra, OpenAI’s latest release, tops the OpenDesign Arena leaderboard with a score of 82.7 out of 100 on real-world design tasks. But DeepSeek’s V4.1 Flash comes close—scoring 81.2, or 98% of Astra’s mark—while charging just $0.023 per finished design compared to Astra’s $1.61.
The test covered 13 models, including Claude Fable 5.1, Grok 4.6, and Qwen 3.8-Max. Eleven of them scored lower than DeepSeek V4.1 Flash and cost more to run. Only GPT-6 Astra scored higher. The result raises a question designers have been asking since the AI era began: how much performance do you really need, and at what price?
The Benchmark’s Design Test
OpenDesign Arena runs models through a set of everyday design tasks: building web apps, dashboards, mobile screens, and landing pages. The scoring system is simple. Thirty of the 100 points check whether the output actually meets the brief. The remaining 70 grade design quality on layout, hierarchy, color, and style fit.
The test is designed to answer a narrower question than most AI leaderboards ask. It asks: Which model should a working web designer actually use tomorrow? That focus changes what counts as a win. A model that scores highly on abstract reasoning tests might fail on a live website’s homepage.
A model’s output only gets scored if it renders as a working webpage in the first place. Anything blank, broken, or cut off scores zero and does not get retested. That means the benchmark measures reliable, everyday design output, not general reasoning or coding skill.
Astra’s Lead and DeepSeek’s Price
GPT-6 Astra, released on September 3, already carries a reputation for doing a bit of everything. It can lay out a circuit board, draft a tax return, and build a 3D scene. Early testers flagged it as a weaker writer than the model it replaced. Its price and pace on OpenDesign’s chart fit that same generalist profile: slower and pricier than DeepSeek’s cheaper entry, but still the highest scorer in the field.
Astra took 11.1 minutes to finish each design task and cost $1.61 per finished design. DeepSeek V4.1 Flash completed the job in 5.3 minutes and charged $0.023. That is a gap measured in both time and dollars.
Claude Fable 5.1 came in at 80.3, took 12.8 minutes, and cost $3.66. Every other model OpenDesign tested—Grok 4.6, Qwen 3.8-Max, Kimi K3, GLM-5.3 Flash, and Gemini 3.8 Flash among them—scored lower than DeepSeek V4.1 Flash and cost more to run. That is 11 of the 13 models tested.
Only GPT-6 Astra beat it outright, and only by a point and a half.
Where the Savings Come From
DeepSeek’s technical report for V4.1 Flash explains the design choice behind its low price. The model carries 552 billion parameters total—the internal settings a model tunes during training to store what it has learned. But it wakes up only 8 billion of them to read an incoming prompt and 16 billion to write the response.
DeepSeek calls this a Causal Encoder-Decoder design. It is the same trick behind the model’s fast completion times. The model stays small in practice even though its total capacity is huge.
“The model activates just 8 billion of its 552 billion parameters to read a prompt.”
That line sums up the whole approach. Most models use nearly all their parameters for every task. DeepSeek V4.1 Flash uses a fraction, which is why its cost stays so low.
A History of Low-Cost Entries
This is not DeepSeek’s first pass at closing a capability gap on the cheap. Weeks earlier, the company’s V4 Pro model landed within 5% of Claude Fable 5 on a separate benchmark comparison while charging a fraction of Fable’s rate. DeepSeek has also been recruiting engineers in Beijing to build its own Code Harness, aiming to own the full agentic stack instead of just supplying the model underneath it.
What the Delivery Rates Show
The benchmark also reports a delivery rate: the share of outputs OpenDesign judged ready to hand off without revision. DeepSeek V4.1 Flash’s delivery rate came in at 57.7%. GPT-6 Astra’s delivery rate was 60%. Claude Fable 5.1’s was 56.7%.
That gap matters because revision costs money. A model that hands over a finished page 60% of the time saves a designer more time than a model that hands over a page 57.7% of the time. But the difference is small enough that the cost gap between Astra and DeepSeek V4.1 Flash still makes DeepSeek the better choice for most projects.
The Takeaway for Designers
The OpenDesign Arena results offer a clear choice for anyone building a website or app. If you want the absolute best design output, GPT-6 Astra is the model to use. If you want good design output at a fraction of the cost, DeepSeek V4.1 Flash is the model to use.
The 1.4% cost gap is not an abstraction. It is the difference between paying $1.61 per design and paying $0.023 per design. For a project with hundreds of designs, that adds up quickly.
Here is how the top three models compare:
- GPT-6 Astra: 82.7 points, $1.61 per design, 60% delivery rate
- DeepSeek V4.1 Flash: 81.2 points, $0.023 per design, 57.7% delivery rate
- Claude Fable 5.1: 80.3 points, $3.66 per design, 56.7% delivery rate
The rankings hold across most metrics. Astra leads on score, time, and delivery rate. DeepSeek V4.1 Flash leads on cost and beats Astra on delivery rate. Claude Fable 5.1 sits in the middle on every axis except price, where it falls far behind both leaders.
For most designers, the choice comes down to a single question. How much of a premium am I willing to pay for the top score? The answer to that question determines which model wins.
The OpenDesign Arena results show that the best model is not always the most expensive one. Sometimes the cheaper model is almost as good—and a lot easier to afford.
Source: decrypt.co