Kimi K2.7 Code vs. Claude Fable 5: Landing Pages for 94% Less

Together AI ran an experiment comparing Kimi K2.7 Code with Anthropic's Claude Fable 5 across 12 landing pages and published the results on June 17, 2026. Kimi K2.7 Code came in roughly 94% cheaper than Fable 5 and barely trailed it on output quality for most pages. The write-up is by Hassan El Mghari.

How the comparison worked

Both models got identical prompts for landing pages across several categories: a developer tool that turns SQL queries into charts, a page for a B2B SaaS startup offering task management and team workflows, and a page for a 1920s-style Art Deco rooftop speakeasy bar in gold and emerald. With a simple first prompt and no extra context, both models produced pages that looked "recognizably AI-generated."

• In total, 12 landing pages were compared across both engines
• On average, Kimi K2.7 Code was about 16 times cheaper than Claude Fable 5
• Kimi K2.7 Code was about 8 times cheaper than Claude Opus 4.8
• Prompt categories included a SQL-to-charts tool, B2B SaaS, and an Art Deco rooftop bar
• All versions, including those generated by Opus 4.8, Fable 5, and Kimi K2.7 Code, are published on the OVSC site

What changed the outcome

To push design quality past the baseline, the Together AI team connected both models to its own MCP server stocked with design references. The server fed the models screenshots of well-designed landing pages, individual UI elements, and other visual examples. According to the authors, this reference MCP server proved the most effective way to give the models stronger creative direction than a text prompt alone.

With that context added, the quality gap between Kimi K2.7 Code and Claude Fable 5 narrowed noticeably, while Kimi still cost several times less.

What it means

Together AI's experiment shows that open models like Kimi K2.7 Code have moved past being a mere "cheap alternative" to closed flagships. Given the right context (here, a design MCP), they deliver results of comparable quality at a fraction of the cost, and the gap between open and closed models on tasks like interface generation keeps shrinking faster than expected.

FAQ

How much cheaper is Kimi K2.7 Code than Claude Fable 5?

According to Together AI, Kimi K2.7 Code was on average about 16 times cheaper than Claude Fable 5 and about 8 times cheaper than Claude Opus 4.8, with comparable output quality.

What exactly did the experiment compare?

The team compared 12 landing pages that Kimi K2.7 Code and Claude Fable 5 generated from identical prompts, ranging from a developer tool and a B2B SaaS service to a themed site for a rooftop bar. All versions are posted on the OVSC site along with a breakdown of cost, tokens, and generation time.

What helped the models produce better results?

A custom MCP server with design references, including screenshots of strong landing pages and individual UI elements, which gave the models stronger creative direction than a text prompt alone.