Moonshot dropped Kimi K3.1 Pro on their API yesterday, and the pricing is aggressive: $2 per million input tokens, $10 per million output. That’s half what you’d pay for Claude Opus 5.5, matching GPT-6.1 Sol exactly.

The real news is the weight release. Moonshot says full model weights ship by November 4. If they deliver, this becomes the first genuinely competitive open weight model at this price point and performance tier.

Two versions, different speeds

The base K3.1 Pro comes with a “Highspeed” variant that runs 2-3x faster at double the price. The exception: cached input tokens cost the same whether you use the fast or standard version. If you’re doing heavy prompt caching, the speed boost is essentially free.

According to Moonshot’s benchmarks, K3.1 Pro performs comparably to both GPT-6.1 Sol and Claude Opus 5.5. Take vendor benchmarks with the usual grain of salt, but the pricing suggests they’re confident about the comparison.

Who should care

If you’re building production AI features and watching your API bills, this matters. Cutting inference costs in half without sacrificing quality changes the economics of a lot of products.

The open weights release matters more. We’ve had plenty of capable commercial models. We haven’t had many capable open models in this performance tier. If K3.1 Pro delivers on both the benchmarks and the November 4 promise, it’s a legitimate option for teams that need to run models on their own infrastructure.

The usual caveats

The weights aren’t out yet. Moonshot could delay, or the released version could perform differently than the API. The benchmarks are Moonshot’s own numbers. Independent testing will matter.

But the API is live now, the pricing is real, and two weeks isn’t long to wait for weights. Worth watching.

Quick Study: Claude draws your photos

Simon Willison built a canvas-based photo painting tool by prompting Claude Opus 5.5 High to generate an artifact. Upload a photo and it paints a version in a few seconds using simulated brushstrokes.

The implementation uses Sobel edge detection with no external libraries beyond Google Fonts. It’s a nice example of what you can build with artifact generation: specific enough to be useful, simple enough to generate in one shot.

Not groundbreaking, but it’s a clean demonstration of using Claude’s artifact mode for actual tools instead of just markdown documents.