Kimi K2.7 Code と Claude Fable 5 の比較：ランディングページ作成コストが 94% 削減

まとめ

We ran 12 landing pages through Kimi K2.7 Code and Claude Fable 5. Kimi cost 94% less and scored within a few points on nearly every page. Open-source models aren't just cheaper, they're genuinely competitive on quality. And the gap is closing faster than people realize.

‍

We ran an experiment where we had Kimi K2.7 Code and Claude Fable 5 each produce 12 landing pages for a side‑by‑side comparison. Overall, Kimi K2.7 Code cost about 94% less than Fable 5 and yielded similar-quality output, especially after we gave Kimi the right context with a design MCP.

We published our findings on the OVSC website, along with all variants generated by Claude Opus 4.8, Claude Fable 5, and Kimi K2.7 Code. On average Kimi was ~16x cheaper than Fable and ~8x cheaper than Opus.

The OVSC website lets you explore all the landing pages along with breakdowns of total costs, token usage, and generation time.

To understand how we ran this experiment, we started by establishing a baseline and seeing what the model could produce from the prompt alone.

The prompts

We started with a small set of landing-page prompts across a few different categories, including B2B SaaS, a rooftop speakeasy, and a developer tool for SQL queries. Here's a sample of the prompts we used:

• Build a landing page for a developer tool that turns SQL queries into charts.
• Build a landing page for a rooftop speakeasy cocktail bar - art deco, gold-leaf and emerald, 1920s glamour.
• Build a landing page for a B2B SaaS startup - a team project-management & collaboration tool (tasks, timelines, team workflows, integrations).

We gave the same prompts to both Kimi K2.7 Code and Claude Fable 5.

Here are the pages that these models created when asked to *“Build a landing page for a developer tool that turns SQL queries into charts.”*

Unfortunately, both models made landing pages that felt recognizably AI-generated.

Even so, this gave us a useful baseline for what each model could produce from a simple prompt alone. To push the designs further, we needed to give the models stronger creative direction, and the most effective way we found to do that was with a custom MCP server.

Design inspiration MCP server

We set up a custom MCP server that provided screenshots of well-designed landing pages, along with individual UI elements and other visual references. Because Kimi K2.7 Code is multimodal, we could include those images directly in the prompt alongside text.

That changed the results significantly. Instead of generating a layout from a short prompt alone, Kimi could work from concrete examples, pick up on the visual language, and apply those patterns to a new page. In practice, the results had stronger hierarchy, better typography, and more intentional composition.

Here's a before and after of the Rooftop Speakeasy landing page:

With design inspiration, Kimi produced pages that loaded faster, avoided broken-image placeholders, and used far more readable typography.

Once the design improved, the next thing we wanted to explore was cost.

Costs per landing page

One of the advantages of using an open-source model like Kimi K2.7 Code is cost. For example, this landing page for a B2B SaaS cost just 4 cents with Kimi. The same prompt cost $1.09 with Claude Fable, making it almost 27 times more expensive.

On average, the landing pages we generated with Kimi K2.7 Code were roughly 16 times less expensive than those generated with a proprietary model like Claude Fable 5.

With generative coding agents you rarely generate just one version of a landing page. More often, you generate many variations so you can explore different design directions, copy, and page elements. You then iterate on the ones that show promise, editing and refining through repeated cycles of experimentation and adjustment. With all the back and forth, the price difference adds up quickly, even for something as simple as a SaaS landing page.

If you were to generate 100 pages with Kimi K2.7 Code, you would save around $94 compared to using a proprietary model like Claude Fable 5.

Lower cost was a clear advantage, but we also wanted a way to compare the quality of the results.

Comparing the results

After generating the landing pages, we wanted a systematic way to compare Kimi and Fable. We were not just looking at the code itself, but at the overall quality of each page, including positioning, visual direction, content structure, craft, responsiveness, and technical execution. To do that, we gave GPT-5.5 a rubric to review and score the screenshots and source code from each page and assign a final score from 0 to 100.

Here are the scores for each landing page:

‍

Claude Fable scored higher in both examples, but the gap was relatively small. Kimi remained competitive on design, structure, and overall page quality, while costing much less to run. For this kind of workflow we felt that trade-off was reasonable.

Final thoughts

Open-source models like Kimi K2.7 Code are already capable of generating useful landing pages, but our experiment showed that prompts alone are only part of the equation. Without better context, both Kimi and Claude Fable tended to produce polished but generic results.

The biggest improvement came from giving Kimi visual inspiration through a custom MCP server. Once it could work from screenshots and design references, the pages became more readable, more structured, and more visually intentional.

Combined with the lower cost, that makes open-source models a practical choice for this kind of workflow. If you can give the model stronger inputs and iterate cheaply, you can get surprisingly far.

You can try open-source models like Kimi K2.7 Code in the Together AI Playground.