What is Kimi K3? Moonshot's new model, tested in Chat Thing [Chat Thing](https://chatthing.ai/) Open main menu - Features - Use Cases - [Pricing](https://chatthing.ai/#pricing) - Resources [Log in](https://app.chatthing.ai/app/auth/login) [ Book a demo](https://chatthing.ai/demo) [Sign up](https://app.chatthing.ai/app/auth/register) [← Back to blog](https://chatthing.ai/blog)Changelog # Kimi K3 is live in Chat Thing: what it is, and is it actually good? ![Zef](https://res.cloudinary.com/djyjvrw5u/image/upload/v1710941716/IMG_2278_d3b5e4fa69.jpg) Zef 17 Jul 2026 ~ 9 min read --- ![Kimi K3 is live in Chat Thing: what it is, and is it actually good?](https://res.cloudinary.com/djyjvrw5u/image/upload/v1784238623/hero_a9dfc94c16.png) --- ### On this page 1. [What is Kimi K3?](#what-is-kimi-k3) 2. [The benchmarks behind the hype](#the-benchmarks-behind-the-hype) 3. [Is Kimi K3 good? We put it head-to-head](#is-kimi-k3-good-we-put-it-head-to-head) 1. [Test 1: the refund maths trap](#test-1-the-refund-maths-trap) 2. [Test 2: the furious customer](#test-2-the-furious-customer) 3. [Test 3: strict JSON](#test-3-strict-json) 4. [Test 4: tool calling (power-ups)](#test-4-tool-calling-power-ups) 5. [Test 5: vision, with a customer screenshot](#test-5-vision-with-a-customer-screenshot) 6. [Speed, and the cost of thinking](#speed-and-the-cost-of-thinking) 4. [What Kimi K3 costs in Chat Thing](#what-kimi-k3-costs-in-chat-thing) 5. [So, should you use it?](#so-should-you-use-it) 6. [Taking it for a spin](#taking-it-for-a-spin) Moonshot AI released Kimi K3 today, and you can already pick it for your bots in Chat Thing. Normally a new model lands, we add it, and everyone quietly gets on with their day. This one has enough noise around it that we thought it deserved a proper look, so we also ran it head-to-head against the other leading models in Chat Thing on real support-bot prompts. The results are further down, and one of them genuinely surprised us. ## What is Kimi K3? Kimi K3 is the new flagship model from Moonshot AI, the Chinese lab behind the Kimi K2 family. It's a trillion-parameter-scale mixture-of-experts model, with reports putting it at around 2.5 trillion parameters, and it arrived just three months after Kimi K2.6. Moonshot has open-sourced its previous flagships, and K3 is widely reported to be the largest open-weight model to come out of China so far. The headline specs: - **A 1 million token context window.** Enough to hold a large knowledge base, long documents and a full conversation without losing the thread. - **Multimodal.** It handles images, so screenshots, photos of error messages and diagrams are all fair game. - **Built for agentic work.** Tool calling is a first-class citizen, which in Chat Thing terms means full power-up support. - **Always-on reasoning.** K3 thinks before it answers on every message. You can't switch that off, which matters for speed and cost, and we'll come back to it. ## The benchmarks behind the hype Moonshot's announcement numbers are punchy, and early coverage has framed K3 as another "DeepSeek R1 moment": an open model landing uncomfortably close to the closed-source frontier. The charts from [Moonshot's announcement](https://www.kimi.com/blog/kimi-k3) tell the story better than a list can. On coding, K3 tops **Program Bench** (77.8) and **SWE Marathon** (42.0, ahead of Claude Opus 4.8's 40.0), sits second only to GPT-5.6 Sol on **Terminal Bench 2.1** (88.3 vs 88.8), and second only to Fable 5 on **FrontierSWE** (81.2). For an open-weight model, being sandwiched between the two most expensive closed flagships on nearly every chart is remarkable on its own. ![Moonshot's coding benchmarks for Kimi K3: DeepSWE, Terminal Bench 2.1, FrontierSWE, Program Bench and SWE Marathon, compared against GPT-5.6 Sol, Fable 5, Claude Opus 4.8, GPT-5.5 and GLM-5.2](https://res.cloudinary.com/djyjvrw5u/image/upload/v1784238792/benchmarks_main_b668dedd80.png) The agentic results are arguably stronger. K3 takes first place on **BrowseComp** (91.2), **Automation Bench**, and **SpreadsheetBench 2** (edging out Fable 5 by a tenth of a point), and runs second on JobBench and the two Elo-style agent leaderboards. On visual reasoning it's second behind Fable 5 on both CharXiv and Zerobench, ahead of Opus 4.8 and the GPT-5.x line. ![Moonshot's general-agent and visual-agent benchmarks for Kimi K3: GDPval-AA v2 Elo, JobBench, AA-Briefcase Elo, SpreadsheetBench 2, Automation Bench, BrowseComp, CharXiv and Zerobench](https://res.cloudinary.com/djyjvrw5u/image/upload/v1784238793/benchmarks_extended_9f0831d269.png) Two caveats worth keeping in view. First, these are Moonshot's own launch-day numbers, and independent verification hasn't landed yet. Second, read the small print at the top of the charts: every model was run with thinking effort maxed out. That's a fair comparison, but it means the benchmark K3 is the slow, expensive, thinks-for-minutes K3, which matters when you're deciding what to run a production bot on. Which is exactly why we did our own testing. ## Is Kimi K3 good? We put it head-to-head Benchmarks are one thing. What we actually care about is how a model behaves when it's running a support bot: does it follow policy, does it stay honest, does it handle a furious customer with grace? So we gave five models the identical setup our customers use every day: a support-bot system prompt for a fictional analytics SaaS, a small policy context, and the same five tests: three chat prompts, a tool-calling task and a vision task. Same settings Chat Thing uses in production for each model. The lineup: **Kimi K3**, **Claude Sonnet 5**, **GPT-5.6 Terra**, **Gemini 3.5 Flash** and its predecessor **Kimi K2.6**. ### Test 1: the refund maths trap We asked a deliberately awkward billing question: an annual plan bought 3 months ago, renewed from an older monthly plan, cancel today, what do I get back and what happens to my data? Every model correctly navigated the policy (past the 14-day window, renewals non-refundable, credit for unused whole months). Every model also did the genuinely important thing: the policy doesn't say what happens to data after cancellation, and all five said so rather than making something up. But Kimi K3 was the only model that did the maths. The others said "you'll be credited for 9 unused months" and stopped there. Claude Sonnet 5's version was typical of the pack: > "Because it's an annual plan and you're cancelling mid-cycle, we'll credit you for the unused **whole months** remaining. With 9 months left, that means you'd get **9 months credited** back." GPT-5.6 Terra said essentially the same in three sentences flat, Gemini 3.5 Flash matched it with a slightly warmer tone, and Kimi K2.6 kept pace too. All correct, all helpful, none of them told the customer what 9 months is actually worth. K3 did: > "Annual Growth works out to $790/year ($79 × 10 months, with 2 months free), so that credit would be roughly **$592.50**. One thing to note: it's an account credit rather than money returned to your card, and our billing team can confirm the exact figure." That's correct ($790 ÷ 12 × 9), it's hedged appropriately, and it's the answer a customer actually wants. This is the reasoning layer earning its keep. ### Test 2: the furious customer Next, an all-caps customer demanding compensation after a slowdown during their big launch. This is a tone test and an honesty test in one: the policy has no SLA and no compensation path, and the worst thing a bot can do is promise one. All five models passed on honesty. Nobody invented a compensation policy. K3's reply stood out for being warm without grovelling, and for one nice line of self-awareness: > "That said, I don't want to just hide behind policy here. If you'll let me, I'd like to escalate this to the team so someone can look into what happened with the script during your launch." Claude Sonnet 5 was similarly excellent here ("You deserve a real answer, not just a policy quote from me"), and honestly this test was close between the two. ### Test 3: strict JSON We asked for a plan comparison as pure JSON with a fixed schema. All the current-generation models produced valid, correctly-typed JSON. The fun result: Kimi K2.6 returned `"pageviews": "100k"` as a string where every newer model, K3 included, returned proper numbers. Small thing, but it's exactly the kind of small thing that breaks a downstream integration, and it shows the generation-on-generation cleanup. ### Test 4: tool calling (power-ups) Power-ups are how Chat Thing bots actually _do_ things, so tool calling matters more to us than almost any benchmark. We gave each model a `create_support_ticket` tool and an urgent message: a customer whose checkout widget had been down for two hours, losing orders. Full marks across the board. All five models called the tool, none of them just talked about calling it, and every single one extracted the right arguments: the customer's email, `priority: "urgent"`, `product_area: "widget"`. The differences were at the margins. GPT-5.6 Terra fired the tool silently in 0.7 seconds. K3 took ~3 seconds but paired the call with a genuinely warm message to the customer, and wrote the most detailed ticket summary of the five ("Checkout widget completely down for ~2 hours; customer actively losing orders — needs immediate attention"). If your bot's tickets go straight to a human team, that extra detail is worth having. ### Test 5: vision, with a customer screenshot Support bots get sent screenshots constantly, so we sent each model one: a fake analytics dashboard with a 429 rate-limit error banner and an events chart flatlining at zero from 13 July, along with "our analytics stopped recording this week and I don't understand why". Every model correctly connected the error banner to the flatline and gave the right fix (upgrade, or wait for the monthly reset). Gemini 3.5 Flash even greeted the customer by name after spotting her email in the nav bar. But K3's answer was the strongest of the five, because it went one step past diagnosis: > "Worth a quick check first: your event volume was climbing steadily before the cutoff. If that wasn't from real traffic growth, common culprits are the snippet installed twice, a dev/staging site pointing at your live site ID, or a new event firing in a loop. Fixing one of those might mean you don't need a bigger plan at all." That's the reply of a support agent who's seen this ticket a hundred times, and no other model offered it. It also flagged that the dropped events are gone for good, which is exactly the expectation-setting a customer needs. Two honest caveats: that answer took 31 seconds of thinking, and our first attempt hit a provider error and needed a retry. One amusing footnote from the same test: Kimi K2.6 diagnosed the problem correctly, then leaked raw function-call XML into its customer-facing reply, hallucinating a ticket tool we never gave it. K3 showed no trace of that behaviour anywhere in our testing. ### Speed, and the cost of thinking The honest trade-off. Because K3 reasons on every message, it was the slowest of the modern models in our test: around 3 seconds to first answer versus about 1.5s for Claude Sonnet 5, ~1s for Gemini 3.5 Flash and a blistering 0.6s for GPT-5.6 Terra. On the hard billing question it spent over 2,000 thinking tokens before writing a word. That thinking is why it was the only model to do the maths, but you pay for it in latency and tokens. We also saw a couple of transient provider errors during testing, which is par for the course on launch day and worth knowing if you switch a production bot over this week. ## What Kimi K3 costs in Chat Thing Chat Thing meters usage in message tokens, and each model carries a multiplier. Here's how K3 slots in against the models we tested it against: | Model | Input multiplier | Output multiplier | Context | Vision | Power-ups | | --- | --- | --- | --- | --- | --- | | Kimi K3 | 6 | 30 | 1M | ✅ | ✅ | | GPT-5.6 Terra | 5 | 30 | 1.05M | ✅ | ✅ | | Claude Sonnet 5 | 4 | 20 | 1M | ✅ | ✅ | | Gemini 3.5 Flash | 3 | 18 | 1M | ✅ | ✅ | | Kimi K2.6 | 1.36 | 6.82 | 262k | ✅ | ✅ | So K3 is priced like the frontier model it's benchmarking as: in the same bracket as GPT-5.6 Terra, a step above Claude Sonnet 5. And remember the reasoning tokens: because K3 thinks on every message and thinking counts as output, its real per-message cost runs higher than the multiplier alone suggests. Kimi K2.6 remains the bargain of the family by a mile. ## So, should you use it? Our take after a day with it: - **Reach for Kimi K3** when your bot deals with genuinely hard questions: multi-step policy logic, calculations, careful reasoning over a big knowledge base, or customers who communicate in screenshots. It was the only model in our tests that went the extra step, twice, and that's exactly the behaviour you want on complex support tickets. - **Don't make it your always-on driver** for a high-traffic widget answering the same simple questions all day. The always-on reasoning makes it slower and hungrier than a workhorse like Gemini 3.5 Flash or Kimi K2.6, and those models handled the routine prompts perfectly well. - **If you liked K2.6 for the price,** stay put. K3 is not a replacement, it's a different tier. ## Taking it for a spin Open your bot, head to General settings, then Advanced, and pick **Kimi K3** from the model dropdown. Send it your hardest real support question, the one with the awkward edge case, and see what it does. That's a better benchmark than anything in the announcement. Model selection is available on Standard plans and above. If you're on a plan below that, you'll still get a great default model, you just won't see the picker. _Cover artwork and benchmark charts: from [Moonshot AI's Kimi K3 announcement](https://www.kimi.com/blog/kimi-k3)._ ### Related articles [![GPT-5.6 lands in Chat Thing, plus twelve more new models](https://res.cloudinary.com/djyjvrw5u/image/upload/v1783942921/hero_c5c962c5c0.png)](https://chatthing.ai/blog/gpt-5-6-and-thirteen-new-models) [

GPT-5.6 lands in Chat Thing, plus twelve more new models

](https://chatthing.ai/blog/gpt-5-6-and-thirteen-new-models) GPT-5.6 (Sol, Terra and Luna) is now live in Chat Thing, alongside twelve more new models. A 1.05M-token context window, vision and power-up support across the board, and it is unusually light on message tokens for how capable it is. [![Claude Fable 5, Opus 4.8, and a brand-new model picker](https://res.cloudinary.com/djyjvrw5u/image/upload/v1781277929/new_models_blog_post_og_image_60c61ed10a.jpg)](https://chatthing.ai/blog/claude-fable-5-opus-4-8-and-a-new-model-picker) [

Claude Fable 5, Opus 4.8, and a brand-new model picker

](https://chatthing.ai/blog/claude-fable-5-opus-4-8-and-a-new-model-picker) Claude Fable 5, the most capable model Anthropic have made, is live on Chat Thing, along with a stack of other new models and a rebuilt picker. A heads-up on what Fable costs, and which plans get to choose. [![Custom Domains: Your Bot, Your Brand, Your Domain](https://res.cloudinary.com/djyjvrw5u/image/upload/v1770724454/file_94_5d631697_29ed_4946_9cd4_efccfd7fe3f5_758939dc92.jpg)](https://chatthing.ai/blog/custom-domains-your-bot-your-brand-your-domain) [

Custom Domains: Your Bot, Your Brand, Your Domain

](https://chatthing.ai/blog/custom-domains-your-bot-your-brand-your-domain) Give your AI bots a fully white-labelled experience. Custom domains let you serve your bot from your own domain like chat.yourcompany.com instead of embed.chatthing.ai.
Previous
[← Chat Thing's MCP server is live: describe your agent, let AI build it](https://chatthing.ai/blog/chat-thing-mcp-server)
## Footer ### Ready to answer every question, day or night? Add an AI support agent to your site in minutes. [Get started free](https://app.chatthing.ai/app/auth/register) [Book a demo](https://chatthing.ai/demo) AI customer support that learns from your content and answers 24/7. ### Features - [Channels: Deploy your bot](https://chatthing.ai/pages/features/channels) - [Data sources: Add knowledge](https://chatthing.ai/pages/features/data-sources) - [Power-ups: Tool calling](https://chatthing.ai/pages/features/power-ups) - [AI Models: GPT-4, Claude, & more](https://chatthing.ai/pages/features/models) - [SDK: Create custom assistants](https://chatthing.ai/pages/features/chatbot-sdk) - [Bot evaluations: Test your bot](https://chatthing.ai/pages/features/testing) - [Analytics: Topics, sentiment & insights](https://chatthing.ai/pages/features/analytics) - [Bot Gallery: Example bots](https://chatthing.ai/gallery) - [All features](https://chatthing.ai/features) ### Use cases - [AI Customer Support Bots](https://chatthing.ai/pages/use-cases/customer-support) - [AI Internal Knowledge Bots](https://chatthing.ai/pages/use-cases/internal-knowledge) - [AI Bots for Marketing](https://chatthing.ai/pages/use-cases/marketing) - [SDK: Create custom assistants](https://chatthing.ai/pages/features/chatbot-sdk) - [All features](https://chatthing.ai/features) ### Channels - [AI support for your website](https://chatthing.ai/pages/features/channels/website) - [AI support for Slack](https://chatthing.ai/pages/features/channels/slack) - [AI support for Discord](https://chatthing.ai/pages/features/channels/discord) - [AI support for WhatsApp](https://chatthing.ai/pages/features/channels/whatsapp) - [AI support for Telegram](https://chatthing.ai/pages/features/channels/telegram) - [AI support API](https://chatthing.ai/pages/features/channels/api) ### Data sources - [Chat with your website](https://chatthing.ai/pages/data-source/web) - [Chat with your files](https://chatthing.ai/pages/data-source/files) - [Notion chatbots](https://chatthing.ai/pages/data-source/notion) - [Chat with your youtube](https://chatthing.ai/pages/data-source/youtube) - [Chat with your RSS feed](https://chatthing.ai/pages/data-source/rss) ### Company - [Discord](https://discord.com/invite/DGCCR34Fsj) - [Docs](https://chatthing.ai/docs/getting-started) - [Blog](https://chatthing.ai/blog) - [FAQ](https://chatthing.ai/faq) - [Status](https://status.chatthing.ai/) - [Testimonials](https://senja.io/p/chat-thing/iq1tWz) - [Custom AI development](https://chatthing.ai/pages/ai-development-services) ### Free tools - [Agent Readiness Checker](https://chatthing.ai/tools/agent-readiness-checker) - [JSON-LD Generator + Checker](https://chatthing.ai/tools/json-ld-generator) - [llms.txt Generator & Validator](https://chatthing.ai/tools/llms-txt-generator) - [Support Ticket Deflection Calculator](https://chatthing.ai/tools/ticket-deflection-calculator) - [Knowledge Base Grader](https://chatthing.ai/tools/knowledge-base-grader) - [AI FAQ Generator](https://chatthing.ai/tools/faq-generator) - [AI Canned Response Generator](https://chatthing.ai/tools/canned-response-generator) - [Support Bot Prompt Generator](https://chatthing.ai/tools/support-bot-prompt-generator) - [All free tools](https://chatthing.ai/tools) ### Compare - [Intercom alternative](https://chatthing.ai/pages/alternatives/intercom) - [Zendesk alternative](https://chatthing.ai/pages/alternatives/zendesk) - [Tidio alternative](https://chatthing.ai/pages/alternatives/tidio) - [Crisp alternative](https://chatthing.ai/pages/alternatives/crisp) - [Chatbase alternative](https://chatthing.ai/pages/alternatives/chatbase) - [DocsBot alternative](https://chatthing.ai/pages/alternatives/docsbot) - [ResolveAI alternative](https://chatthing.ai/pages/alternatives/resolveai) - [ChatBotKit alternative](https://chatthing.ai/pages/alternatives/chatbotkit) - [CustomGPT alternative](https://chatthing.ai/pages/alternatives/customgpt) - [Botsonic alternative](https://chatthing.ai/pages/alternatives/botsonic) - [Ada alternative](https://chatthing.ai/pages/alternatives/ada) ### Join our newsletter Be the first to know about new features and updates. Join [Privacy](https://chatthing.ai/legal/privacy) [Terms](https://chatthing.ai/legal/terms) [DPA](https://chatthing.ai/legal/dpa) [GDPR](https://chatthing.ai/legal/gdpr) [Sub-Processors](https://chatthing.ai/legal/sub-processors) [Twitter](https://twitter.com/pixelhopio) [Discord](https://discord.com/invite/DGCCR34Fsj) [YouTube](https://www.youtube.com/@ChatThingAI) © 2026 Chat Thing. All rights reserved. Company number: 14789358. Built with by [Pixelhop](https://pixelhop.io?utm_source=chat+thing&utm_medium=website) ## Cookies We use cookies to run the site and understand how it's used. You can change your choice anytime. Accept Decline Customise