$3 per million input tokens, $15 per million output. That’s Kimi K3.
Read that again, because it’s the whole story. Moonshot AI — the Beijing lab that built its name on cheap open-weight models — is now charging exactly what Anthropic charges for Claude Sonnet 5. Same dollars, same tier, same shelf. Not DeepSeek money. Not GLM money. Sonnet money, from the company whose last model was the budget option.
For two years, Chinese open-weight labs competed on price the way airlines compete on routes. DeepSeek went to $0.44/$0.87. GLM-5.2 went to $1.40/$4.40. Moonshot’s own K2.6 undercut everyone. The implicit promise was simple: same quality band, half the cost, switch and save. K3 breaks that contract. It arrives at Sonnet parity and says: we’re not the cheap option anymore, we’re the good option. That’s either the bravest pricing decision in open-weight history or a very expensive bet on benchmarks nobody has independently verified.
The bet matters because Moonshot just raised at a $31.5 billion valuation — up from $4.3 billion in January, fifteen times in eighteen months. That kind of money buys you the right to ignore the discount shelf for a cycle. But it also means the model has to actually be worth it. And here, the picture gets complicated.
Two point eight trillion parameters, ninety-eight percent asleep
K3 is big. 2.8 trillion parameters total, but only 16 of 896 experts activate on any given token. That’s a Stable LatentMoE design — the same family of tricks that made DeepSeek V3 and GLM-5.2 work, pushed further. You run a sliver of the model per token and you get frontier-grade output without paying the full parameter bill.
But the architecture doing the heavy lifting isn’t the MoE. It’s Kimi Delta Attention, KDA, and it’s the reason K3 can offer a flat 1,048,576-token context window without charging more for long inputs. Here’s the idea in plain terms: in a normal transformer, every attention layer is full attention. Every token looks at every other token. Memory grows quadratically with sequence length, and by the time you hit a million tokens of context, the KV cache is eating your whole GPU budget. KDA replaces three out of every four attention layers with a linear attention variant that compresses the KV state. The result, according to Moonshot’s own research published at arXiv, is up to 75% KV cache reduction and up to six times the decoding throughput at one-million-token context.
The architecture is the pricing story. You can’t ship flat-rate 1M context on a vanilla transformer — the math doesn’t work, the inference cost explodes, the per-token economics collapse. KDA is the only reason Moonshot can promise “same price at 4K context and 1M context” without bleeding cash on every long-context call. Whether it actually matches full attention on quality is a separate question the paper tries to answer in the affirmative. The community will spend the next month checking.
So when you see the $3/$15 sticker and the 1M context and think “that’s aggressive,” understand the aggression is structural. Moonshot isn’t subsidizing the price. It’s betting the architecture lets it serve long contexts cheaply enough to match Sonnet’s sticker without Sonnet’s margin.
The benchmarks, with the caveat written into the section title
Moonshot’s official numbers, max reasoning effort: DeepSWE 67.5, Program Bench 77.8, Terminal-Bench 2.1 88.3, GPQA-Diamond 93.5. By those cells, K3 edges Claude Opus 4.8 on every published eval. Trailing Fable 5 and GPT-5.6 Sol, but ahead of Opus on the coding agent benchmarks that people actually run in production.
The HN thread — 541 points, 278+ comments, one of the hotter AI threads this month — split exactly down the middle. One camp: “If the numbers hold up scrutiny, this is scary good.” Another camp: “GLM 5.2 is roughly the same quality at a much cheaper price.” Both can be true. The numbers can hold and the price can still be wrong.
Here’s the problem with the benchmark sheet, and it’s the same problem every vendor-reported launch has: no independent SWE-bench Verified yet. No third-party replications. No OpenAgents or Terminal-Bench leaderboard corroboration. Moonshot’s max-effort numbers are Moonshot’s max-effort numbers. We’ve been here before. Every lab reports their best run, their best seed, their best evaluation configuration. The real number is almost always lower. Sometimes a lot lower.
The more interesting criticism came from a different angle entirely. One HN commenter pointed out that reasoning efficiency matters more than per-token pricing. “If Sol spends 10K reasoning tokens vs K3 spending 50K, Sol wins on cost effectiveness.” This is the argument that actually eats K3’s lunch. A model can be cheaper per token and still be more expensive per task if it thinks too much. K3’s benchmarks are at max reasoning effort. What’s the token cost per SWE-bench solve? Moonshot didn’t publish that. Until they do, $3/$15 is a sticker, not a bill.
And there’s the geopolitical noise that won’t go away. Multiple HN comments — not a majority, but persistent — said plainly that their organizations refuse to use Chinese models due to security and IP concerns. That’s not a quality critique, it’s a procurement policy. K3 can be the best model in the world and still not land in half the enterprise accounts in the West. Moonshot’s pricing doesn’t fix that. If anything, Sonnet parity makes the comparison sharper: why pay Anthropic prices for a model your legal team won’t sign off on?
The table that actually matters
This is the one place a table earns its keep, because pricing is a side-by-side question:
| Model | Input $/M | Output $/M | |——-|———–|————| | DeepSeek V4 Pro | $0.44 | $0.87 | | GLM-5.2 | $1.40 | $4.40 | | Kimi K3 | $3.00 | $15.00 | | Claude Sonnet 5 | $3.00 | $15.00 |
K3 is identical to Sonnet. It’s roughly three times GLM-5.2 on input, three and a half times on output. It’s seven times DeepSeek V4 Pro on input and seventeen times on output. Seventeen. Read that again.
The question the entire HN thread is arguing about reduces to: is K3 seventeen times more useful than DeepSeek V4 Pro on your workload? Almost certainly not. Is it three times more useful than GLM-5.2? Maybe, on the longest agentic tasks, on the hardest multi-hour coding sessions where one-million-token context genuinely matters. On everything else — short queries, summarization, structured extraction — no chance.
Moonshot knows this. Which is why the framing in their launch wasn’t “we’re cheaper than Claude.” It was “we match Claude, and we’re open weight.” The play isn’t to win on price. The play is to win on capability parity plus the open-weight escape hatch. Pay Sonnet prices now, get the weights later, self-host when your traffic justifies it.
Except the weights aren’t here yet.
The nine-day credibility gap
Here’s the part that should make you nervous. Moonshot’s launch materials originally included a paragraph saying open weights would arrive “in coming days.” Modified MIT license, downloadable from HuggingFace. Then that paragraph got quietly removed from the docs. The HN thread caught it.
As of July 18, there are no K3 weights on HuggingFace. What’s there instead is a file called “K3-Q4-GGUF” — 1.91 terabytes, looks official, is not. It’s a community merge of K2 variants packaged to look like K3 quantized. Anyone who downloads it expecting frontier weights is getting a recycled older model. The fact that this fake exists at all tells you the demand is real. The fact that Moonshot hasn’t shipped the real thing two days after launch tells you something else.
The official line is now “weights by July 27.” Modified MIT. That’s eleven days from the July 16 launch, nine days from today. It’s also the real deadline for everything K3 claims to be.
If the weights drop on schedule and they’re real and they run at the promised quality, K3 becomes the most capable open-weight model ever shipped. At that point the pricing argument changes completely — you can always self-host, skip the API, eat the $3/$15 only when you don’t want to deal with inference infrastructure. The open-weight community gets a genuine frontier-grade base model. DeepSeek and Z.AI have to respond. The whole discount-tier equilibrium shifts.
If the weights slip — even by a week, even with a plausible excuse — K3 stops being an open-weight play and becomes just another closed frontier model with a Chinese accent. Sonnet prices without the Sonnet ecosystem. Sonnet prices without the Anthropic usage tier, the Bedrock integration, the enterprise compliance paperwork. At that point you’re paying frontier money for a chatbot with a nice context window, and the entire “we’re the open-weight alternative” pitch collapses into marketing.
The open-weight movement’s credibility is riding on July 27. Not metaphorically. Literally. DeepSeek, GLM, MiniMax, Qwen — every Chinese lab that’s promised open weights has either delivered or gotten burned. Moonshot has eleven days from launch. The community is watching, and the GGUF faker already proved they’ll fill the vacuum with garbage if Moonshot doesn’t fill it with the real thing.
What it’s actually like to run
OpenRouter is serving K3 right now, and the live metrics are the most honest data we have on the model in production. Not benchmarked. Actually running.
The good numbers: 88.2% cache hit rate. That’s excellent — it means K3’s prompt caching is working as advertised and most repeated context is hitting the discounted $0.30/M cached tier. Throughput at 29 tokens per second. Tool call error rate at 0.33%, which is genuinely clean. On those axes K3 looks like a model that’s been hardened for production, not a launch-day mess.
The bad number: 17.82% structured output error rate. That’s nearly one in five JSON-mode calls failing. For a model pitched at agentic workloads — agents live on structured tool calls — that’s alarming. If you’re building an agent pipeline that makes fifty structured-output calls per task, a 17.82% failure rate means you’re retrying constantly, your latency balloons, and your effective token cost is much higher than the sticker suggests because you’re paying for failed attempts.
This connects back to the reasoning-efficiency argument. A model can have frontier benchmarks and still be expensive to actually operate if it’s unreliable in the formats production agents need. The 0.33% tool call error rate is great. The 17.82% structured output rate is not. Those are different failure modes, and the gap between them says something about where K3 is polished and where it isn’t.
There are also two model variants at launch that most of the coverage glossed over. K3 Max is the chat-and-agent variant. K3 Swarm Max is the parallel-processing variant, designed for the 300-agent swarm tier in the $99/month Allegro subscription. The existence of a swarm-tuned variant tells you Moonshot is serious about the multi-agent use case — they’re not just selling a model, they’re selling an architecture where many K3 instances coordinate. Whether that actually works at scale is another open question.
The subscription angle nobody is discussing
Moonshot’s consumer pricing is tiered by context length and agent count: $19 for Moderato at 256K context, $39 for Allegretto at 1M context, $99 for Allegro at 1M plus the 300-agent swarm, $199 for Vivace. That’s a Claude Pro / ChatGPT Pro price band.
The interesting question, flagged by Developers Digest and AIReiter, is whether existing Claude or OpenAI monthly subscriptions beat pay-per-token K3 for daily coding work. If you’re already paying $200/month for Claude Max and K3 output is $15 per million tokens, the subscription wins only if you’re using fewer than roughly 13 million output tokens a month. Heavy daily coders blow past that. For them, pay-per-token K3 might actually be cheaper than their current Claude subscription — assuming the quality holds and the structured output failure rate doesn’t kill agent workflows.
AIReiter’s framing is the sharpest one I’ve read on K3: “The model is the CPU, the context window is the RAM, and your membership tier is the RAM configuration.” That’s exactly right. Moonshot isn’t selling a chatbot subscription. They’re selling compute configurations, the way cloud providers sell instance sizes. You pick the model, you pick the context window, you pick how many parallel agents, and you pay for the configuration you need. The API pricing is for people who want raw access. The subscriptions are for people who want a pre-bundled configuration.
What this means
K3 is a bet, and Moonshot has put the valuation behind it. The bet is that open-weight models can stop being the discount tier and become the quality tier. That pricing parity with Sonnet is justified by capability parity with Sonnet, and that the open-weight escape hatch — self-host when you’re ready — makes the sticker price worth paying for users who’d rather not be locked into Anthropic’s billing forever.
If the benchmarks hold up to third-party replication, if the weights land on July 27, if the structured output failure rate comes down with a few weeks of production hardening, K3 is the model that finally crowdfunded the open-weight frontier. The calculus shifts. The discount era ends not because Chinese labs got expensive, but because one of them got good enough to charge what the work is worth.
If any of those ifs break — weights slip, benchmarks don’t replicate, the structured output failure rate stays in the teens — K3 is the most expensive K2.6 ever made. Sonnet prices for Sonnet-adjacent capability, no ecosystem advantage, no compliance cover, no open weights to fall back on. A frontier-shaped object with a Chinese accent and no exit ramp.
July 27 is the test. Not the benchmark. Not the launch. The weights drop or they don’t. Everything else — the pricing, the architecture, the swarm variants, the subscription tiers — is contingent on that one date. Nine days. The entire open-weight credibility surplus Moonshot has been building for two years either compounds or evaporates. No press release can substitute for a HuggingFace upload.