A 2.8-Trillion-Parameter "Largest Open Model" Arrives at the Same Price as Sonnet 5 — How Kimi K3 Broke the "Chinese Models Are Cheap" Assumption
Moonshot AI released Kimi K3 on July 16. With 2.8 trillion parameters and a 1M-token context, it lands near the top of coding leaderboards — but at $3/$15, its pricing now matches Sonnet 5. The weights are slated for July 27, and the "Chinese models are cheap" assumption has collapsed.
On July 16, 2026, Moonshot AI released its flagship model, Kimi K3. Billed as an "open 3-trillion-class" model, it lands near the top of coding leaderboards. For developers, though, the real story isn't the performance — it's the price tag and the release timing.
From 18th to 1st: A Leap Up the Frontend Leaderboard
K3 is a mixture-of-experts (MoE) model with 2.8 trillion total parameters. It uses a sparse design that activates 16 of its 896 experts per token. The context window is roughly 1.04 million tokens (1,048,576), aimed squarely at long-running agentic work.
The coding rankings are what catch the eye. On arena.ai's frontend leaderboard, K3 jumped from the previous generation K2.6's 18th place to 1st. Here are the main published scores.
| Benchmark | Score | Context |
|---|---|---|
| Frontend Code Arena (arena.ai) | 1st | Top of 6 out of 7 categories |
| Terminal-Bench 2.1 | 88.3% | Autonomous terminal operation |
| GPQA Diamond | 93.5% | Reported as the best open model at launch |
| BrowseComp | 91.2% | Agentic research tasks |
| MCP Atlas | 84.2% | Tool use via MCP |
That said, it isn't dominant across the board. On the independent Artificial Analysis index, it placed 4th out of 189 models. Developer Simon Willison's hands-on testing echoes this: K3 generally comes out ahead of Claude Opus 4.8 and GPT-5.5, but falls short of Claude Fable 5 and GPT-5.6 Sol. Moonshot itself concedes that overall usability doesn't quite reach Fable 5 or GPT-5.6 Sol.
The Price Tag Now Matches Sonnet 5
This is the biggest shift. K3's API pricing is $3.00 per million input tokens and $15.00 per million output tokens. On a cache hit, input drops to $0.30.
| Item | Kimi K3 |
|---|---|
| Input (per 1M tokens) | $3.00 |
| Input, cache hit | $0.30 |
| Output (per 1M tokens) | $15.00 |
| Scope | Flat across the full ~1.04M-token context |
That $3/$15 is the same level as Claude Sonnet 5's standard pricing. It also represents a 3x to 3.75x increase over the K2 generation. The pitch for Chinese open models has been "comparable performance for a fraction of the price." With K3, Moonshot has let go of that premise itself, choosing to compete on substance rather than price.
"Open" Is an IOU Until July 27
There's one more caveat that matters for adoption decisions: the weights are not available at launch. In its own WeChat announcement, Moonshot set July 27 as the date for the full weight release.
Until then, K3 is effectively a hosted-only model. Local or on-premise testing — and any procurement decision that presumes access to the weights — has to wait for that date. For now, you can start via the OpenAI-compatible API (api.moonshot.ai/v1, model ID kimi-k3) or through OpenRouter. If you just want to swap it into an existing agent to try it out, changing the endpoint and model ID is enough.
Before You Try It: The Tokens That Quietly Add Up
Hands-on reports point to several factors that make the bill hard to predict.
- Reasoning tokens are heavy: In Willison's testing, a 3,417-token response came with 13,241 reasoning tokens, costing 25 cents in a single call.
- Reasoning volume is unstable: Consumption reportedly varied by 275–351% under identical conditions. On OpenRouter, there are cases where reasoning hits the output limit and the request fails outright.
- Input tokens are counted differently: One report describes a short instruction worth about 10 tokens being billed as 95. A hidden system prompt of roughly 85 tokens is suspected, but unconfirmed.
- Assessments of output volume disagree: Moonshot says output tokens are down 21% versus K2.6. Third-party measurements, meanwhile, found roughly twice the output of comparable models.
Moonshot also names limitations of its own: stable operation requires passing the entire history, and the model tends to over-interpret ambiguous instructions by anticipating intent. Both matter if you plan to run it unattended for long stretches.
Behind the Wider Field of Options
For developers, this means one more option at the top tier. The 1M context and the Terminal-Bench 2.1 numbers in particular make it a realistic candidate for long-running agentic work. The safe approach is to start narrow with caching enabled, look at the actual bill, and expand from there.
At the same time, there are concerns worth noting. If the price advantage of open models shrinks, so does the escape hatch teams have relied on for cost. And as long as the weight release remains an announced plan, the assumption that "it's open, so I can switch away freely" hinges on whatever license ships with it. Now that the price tags have converged, the selection criterion is shifting back from "because it's cheap" to "will this actually finish the job faster?"
References: Simon Willison — Kimi K3, and what we can still learn from the pelican benchmark / VentureBeat — China's Moonshot AI releases Kimi K3 / OpenRouter — Kimi K3 / Trilogy AI — Kimi K3 Is Live: Pricing, Benchmarks, and the Wait for Public Weights / kie.ai — Kimi K3 Pricing
