For two years the story out of China was that prices only fall. Moonshot just tripled its own rate card with Kimi K3, and the uncomfortable part is that the model is good enough to charge it.

What the upgrade costs the people already on Kimi

Most price coverage compares a new model to its rivals. The more useful question here is what happens to the invoice of a team that is already running Kimi and clicks upgrade. Take an agentic month of 40M input tokens and 8M output, no caching. On K2.6 that bills $70. On K3 the identical traffic bills $240. The upgrade is a $170 line item, and nobody has to change a single prompt to trigger it.

One thing genuinely softens the blow. Artificial Analysis measured K3 producing about 21 percent fewer output tokens than K2.6 across its evaluation suite while scoring 13 points higher, so the model is less chatty per unit of work. Fold that in and the 8M output drops to roughly 6.3M, which pulls the K3 month to about $215. Better, but the honest headline is still a threefold increase rather than a fourfold one. Efficiency gains of 21 percent do not absorb rate increases of 375 percent.

Caching is where this gets interesting rather than merely expensive. Moonshot claims cache hit rates above 90 percent in coding workloads, and at $0.30 against $3.00 the discount is steep enough to pull a K3 month down to $143. That is still double a cached K2.6 month, because the cache discount applies to both models. Caching narrows the ratio from 3.4x to 2.0x. It does not erase it, and any post telling you K3 is cheap if you just cache properly is quietly comparing a cached new model against an uncached old one.

K3 and Sonnet 5 have the same sticker. A sticker is not a bill.

Here is a coincidence worth a minute of your time. From September 1, when Anthropic's introductory rate expires, Claude Sonnet 5 costs $3 input and $15 output. Kimi K3 costs $3 input and $15 output. Identical rate cards, down to the cached-input line at $0.30 apiece. A pricing page would show these two models as a tie.

Whether that tie survives contact with an invoice is a different question, because a rate card prices tokens and your work arrives as text. How much text becomes one token is a property of the model's tokenizer, and Anthropic documents this about its own lineup with unusual candour: the tokenizer used by Sonnet 5 and the other recent Claude models "produces approximately 30% more tokens for the same text" than the one in Sonnet 4.6 and earlier. Same vendor, same document, 30 percent more billable units, from nothing but a tokenizer swap.

Be careful what you do with that number, though, and we nearly got this wrong ourselves. Anthropic measures the 30 percent against the older Claude tokenizer, not against Kimi's. It is not a licence to mark Sonnet 5 up 30 percent against K3. Nobody has published a K3-versus-Sonnet-5 tokenizer comparison that we can find, we have not run one, and until somebody does, the honest answer on which model turns your documents into fewer tokens is that we do not know. What Anthropic's disclosure establishes is the mechanism rather than the magnitude: identical rate cards can bill differently for identical work, and a tokenizer change alone moved one vendor's effective price by tens of percent.

Read that table for what it is: an equal-token-count comparison, with the assumption doing the real work stated out loud rather than buried. On that basis Sonnet 5 undercuts K3 by $80 a month for the next six weeks and then draws level to the cent on September 1. The part you cannot read off either pricing page is which model needs fewer tokens for your particular text, and that unknown is large enough to decide the September question on its own. So run a few hundred of your real prompts through both, count the tokens each one actually bills, and put those counts into the cost calculator. That measurement takes an afternoon and it beats any arithmetic we or Anthropic can do for you from a rate card.

The case for paying it

A tripled rate card would be a scandal if K3 were an incremental update. It is not. Artificial Analysis has it at 57 on the Intelligence Index, ranked fourth of 189 models as of today, trailing Fable 5 at 60 and GPT-5.6 Sol at 59 while beating Opus 4.8 at 56. On Arena's Frontend Code board it is outright first with 1679 points, ahead of Fable 5, which is a 17-place climb from where K2.6 sat. Moonshot is candid in its own launch post that K3 trails Fable 5 and Sol overall, which is a refreshing thing for a lab to write about its own release.

The number that reframes the price is cost per finished task. Artificial Analysis spent $2,690 running K3 through its suite and puts it at $0.94 per task against Sol's $1.04, Opus 4.8's $1.80 and Fable 5's $2.75. So K3 is no longer cheap in absolute terms, but it is the cheapest way to buy a top-four intelligence score, at roughly half of Opus and a third of Fable. Moonshot did not stop competing on price. It moved up a weight class and undercut the models it found there.

Read that table in the right direction. It does not say K3 is the best model, it says K3 is the cheapest entry into the group of models that are arguably the best. Index scores are a blunt summary and these leaderboards move weekly, so treat the ranking as a snapshot dated July 17 rather than a standing fact. Moonshot's own benchmark table is worth reading directly: 81.2 on FrontierSWE against Fable 5's 86.6, 88.3 on Terminal-Bench 2.1 against Sol's 88.8, and 42.0 on SWE Marathon where it leads both.

Two things the launch post does not lead with

K3 is verbose, and verbosity is billed. Artificial Analysis clocked it generating 130 million tokens across their evaluation against a 63 million average, which is more than twice the field. The 21 percent efficiency gain over K2.6 is real, and K3 is still one of the wordiest models on the board in absolute terms. Part of that is structural: Simon Willison points out that OpenRouter currently exposes only the maximum reasoning effort level, with more promised later, so you cannot yet dial the thinking down for simple work. When output costs $15 per million and the model insists on filling it, that setting matters more than the rate card does.

The second one is more serious. Artificial Analysis found K3's hallucination rate rose to 51 percent from K2.6's 39 percent, even as accuracy improved from 33 to 46 percent. A model that is both more right and more confidently wrong is a specific hazard for unsupervised agent work, where nobody reads the intermediate steps. That is a single third-party measurement rather than a replicated finding, so weight it accordingly, but it is the thing we would test first before routing anything unattended.

Worth noting what is not a caveat: the 1M context window is billed flat, with no long-prompt surcharge. That is a real edge over Grok 4.5, where crossing 200K prompt tokens doubles the rate on every token in the request, and over the long-context cliff we have written about on the OpenAI side. If your workload is genuinely long-prompt, K3 keeps arithmetic simple in a way most of its price band does not.

The July 27 asterisk

Moonshot's launch post commits to publishing full K3 weights by July 27, ten days out, with a technical report after. Take that seriously, because previous Kimi releases have followed through, but today it is a promise and not a download. Until then K3 is API-only, through Kimi.com, the mobile apps, platform.kimi.ai and OpenRouter.

The weights change the cost conversation more than the rate card does. A 2.8-trillion-parameter MoE that activates 16 of 896 experts is not something you casually self-host, and the hardware bill for serving it will dwarf a $240 API month for anyone who is not running serious volume. But it does put a ceiling on Moonshot's pricing power: if the API rate ever gets uncomfortable, a competitive serving market can form under it, which is precisely what nobody can do to Anthropic or OpenAI. That option has value even for teams who will never exercise it.

Where K3 lands in the field

Priced against the whole board on that same 40M/8M month, K3 sits squarely in the middle. It is now more than two and a half times the cost of GLM-5.2 and thirty times DeepSeek V4 Flash, and it is comfortably under Opus 4.8, Sol and Fable 5. The budget tier of Chinese models did not disappear. Kimi just left it.

So who should actually move? If you are paying Sol or Fable money for frontend and agentic coding, K3 is the clearest pilot on this list, and Arena's frontend result plus the $0.94 per task give you a real reason to route a slice of traffic at it this week. If you were on K2.6 because it was cheap, do not upgrade on reflex. Your bill triples, and the honest test is whether the extra 13 index points fix work that was actually failing before.

And if the budget tier was the point, the tier still exists, just not with a Kimi badge on it. GLM-5.2 at $1.40/$4.40 and DeepSeek V4 Flash at $0.14/$0.28 are still there. Put your own numbers into the calculator and check the full board on the pricing page before you let a version bump reprice your month.

Sources

• - Kimi K3 pricing, Moonshot: platform.kimi.ai
• - Kimi K2.6 pricing, Moonshot: platform.kimi.ai
• - Kimi K3 launch post and benchmark table: kimi.com/blog
• - Kimi K3 on Artificial Analysis (index, cost per task, verbosity): artificialanalysis.ai
• - Anthropic pricing and the 30 percent tokenizer note: platform.claude.com
• - Kimi K3 on OpenRouter (reasoning effort levels): openrouter.ai
• - The Decoder on K3 and the end of cheap Chinese AI: the-decoder.com
• - Simon Willison on Kimi K3: simonwillison.net
• - Related: Kimi K2.6 pricing and benchmarks
• - Related: Claude Sonnet 5 pricing
• - TokenCost pricing page: tokencost.app/pricing