Kimi K2.7 Code vs Claude Opus 4.8: we tested the cheaper open model in Cyrus

Let's get the obvious out of the way: we know the only question anyone actually wants answered this month is "when Fable?" We hear you. We're refreshing the status page too. But while the internet waits for its favorite model to come home, something quieter and genuinely useful landed — a new open model called Kimi K2.7 — and the price-to-quality ratio is hard to look away from.

So we did what we always do with a new model: we pointed Cyrus at it and watched.

This is internal testing, not a benchmark suite. But the short version: the model is smart, it makes real intellectual leaps, and it costs a fraction of what the frontier closed models charge.

What is Kimi K2.7 Code?

Kimi K2.7 Code is Moonshot AI's coding-focused model in the Kimi K2 family, released on June 12, 2026. It's a mixture-of-experts model with roughly 32B active parameters out of about 1T total, a native multimodal architecture that accepts text and image input, and a 262K token context window. It's built to run end-to-end programming tasks reliably over long contexts — exactly the kind of multi-step, read-plan-edit-verify loop that agent work like Cyrus depends on.

How does Kimi K2.7 compare to Opus 4.8?

The headline isn't the capability gap. It's the price gap.

That's $3.50 versus $25 per million output tokens — roughly one-seventh the cost on the side of the ledger that usually dominates an agent's bill. As we put it while testing: less than a sixth of the price, without hardly any trade-offs. When the cheaper option is this close on quality, the math gets uncomfortable for the expensive option fast.

Opus 4.8 still has real advantages — a 1M-token context window, Anthropic's safety tuning, and the deepest judgment on genuinely hard, ambiguous tasks. If you're running long-horizon work across a huge codebase, that headroom matters. But for a large slice of everyday coding tasks, "smart enough, seven times cheaper" is a very compelling place to be.

The community's take

We're not the only ones noticing. Here's one developer's read after their own agent testing:

Worth stressing: that's one person's ranking, not a controlled evaluation, and "closer to Fable level" is a high bar to claim. Take it as a vibe check, not a verdict. But it lines up with what we saw — and the broader signal is clear. Search interest for "kimi k2.7 code", "kimi k2.7 benchmark", and "kimi k2.7 pricing" all spiked into breakout territory this month. People are price-shopping their coding models, and they're doing it in real time.

What we saw in our own testing

Running a real task through Cyrus, Kimi K2.7 didn't just autocomplete — it reasoned. At one point it recognized that to get a change into one repository, it first had to cut a release-candidate version of a separate package and pull that in. That's the kind of multi-repo, dependency-aware leap that separates a model you can hand a ticket to from one you have to babysit. We came away genuinely impressed for a model at this price point.

Trying K2.7 in Cyrus through opencode

Here's the part we're most excited about. We're testing out opencode runner support for Cyrus, which lets you drive sessions with models served through providers like OpenRouter — including open models like Kimi K2.7.

That support exists thanks to a pull request from GitHub user jappymondo at the German company Digimondo. We're really fortunate to have an open-source community building on top of and with Cyrus, and this is a great example of it. The work is in internal testing right now, and we're looking forward to releasing it soon.

The bigger picture: capable coding models are getting dramatically cheaper, and Cyrus is becoming a place where you can pick the right model for the job — frontier closed models when you need their judgment, fast and affordable open models when you don't.

And yes — we're still counting the days until Fable comes back. That's its own story.

If you're not yet using Cyrus, get started today and put your model choice to work on your real backlog.