{"slug": "moonshot-ai-s-kimi-k3-undercuts-gpt-and-claude-while-proving-its-numbers", "title": "Moonshot AI's Kimi K3 Undercuts GPT and Claude While Proving Its Numbers", "summary": "Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model, on July 16 via API and published its weights on July 27, making it the largest open-weight model to date. On the GDPval-AA v2 benchmark, Kimi K3 scored 1,687, third behind Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8), but ahead of Claude Opus 4.8 (1,600); on FrontierSWE, it scored 81.2, beating GPT-5.6 Sol and trailing only Claude Fable 5's 86.6. Priced at $3 per million input tokens and $15 per million output tokens, Kimi K3 costs roughly a third of GPT-5.6 Sol and a fifth of Claude Fable 5, signaling a shift from ultra-cheap Chinese AI to capability-focused competition.", "body_md": "*Moonshot AI's new Kimi K3 beats or nearly matches GPT-5.6 and Claude Fable 5 on several hard benchmarks, and it costs a fraction of what either one charges per token.*\n\nThe Beijing lab put Kimi K3 live over its API on July 16, then published the full model weights on July 27, making it, at 2.8 trillion total parameters, the largest open-weight model anyone has released. You don't need to take Moonshot's word for how good it is, either. The weights are public, and the model is already running on OpenRouter, Together AI and other third-party hosts, so anyone can run the same prompts Moonshot ran and check the answer themselves.\n\nThat distinction matters more than it usually would. DeepSeek's V4 Pro 0813 reached general availability on August 12, with the company claiming gains of up to 49.9 percentage points over its prior model. Four days later, no third-party evaluator had reproduced those numbers. Independent testing that has come in since found real weaknesses in software engineering and reasoning, and on Code Arena, DeepSeek's V4 Pro sits at rank 35, behind GLM 5.1 and behind Moonshot's own earlier Kimi K2.6. Bold claims from Chinese labs have had a bad few weeks. Kimi K3 arrives into that skepticism with something DeepSeek doesn't have yet: open weights and third-party hosting that let anyone verify the score instead of trusting the press release.\n\nOn GDPval-AA v2, a benchmark that tests real-world tasks across 44 occupations and nine industries, Kimi K3 scored 1,687. That's third place, behind Claude Fable 5 Max at 1,815 and GPT-5.6 Sol Max at 1,747.8, but ahead of Claude Opus 4.8 at 1,600. On FrontierSWE, a coding benchmark, K3 scored 81.2, beating GPT-5.6 Sol, GLM-5.2, Opus 4.8 and GPT-5.5, and trailing only Claude Fable 5's 86.6. Tom's Hardware reported that Kimi K3 outright beat Claude Fable 5 on the Frontend Code Arena leaderboard, a narrower but still notable win. It isn't the best model in the world. It's the best open-weight model, and it's within striking distance of the two most expensive frontier systems on the market.\n\nThe architecture explains part of the price gap. Kimi K3 activates only 104 billion of its 2.8 trillion parameters on any given token, routing through what Moonshot calls Stable LatentMoE: 896 experts total, 16 active per token, with a quantile-based routing method the company says guarantees even load across experts with zero of them going unused. A hybrid attention mechanism, KDA, swaps out standard quadratic attention in most layers while keeping full attention where it actually matters for accuracy. None of that is background trivia. It's the reason Moonshot can serve a 1-million-token context window at a fraction of what Anthropic or OpenAI charge for far smaller active compute per query.\n\n[Kimi K3 Shows the Cheap Chinese AI Era Is Coming to an End](https://startupfortune.com/kimi-k3-shows-the-cheap-chinese-ai-era-is-coming-to-an-end/)\n\nMoonshot AI's new Kimi K3 model launched at $3/$15 per million input/output tokens, three to four times what its predecessor charged, and near the per-task cost of GPT-5.6 Sol. The pricing marks a clear break from the ultra-cheap positioning that made DeepSeek famous and signals that top Chinese labs are now competing on capability rather than... - [chinese AI pricing increase 2024](https://startupfortune.com/kimi-k3-shows-the-cheap-chinese-ai-era-is-coming-to-an-end/) - [moonshot kimi K3 cost comparison](https://startupfortune.com/kimi-k3-shows-the-cheap-chinese-ai-era-is-coming-to-an-end/)\n\nAnd the price gap is real. Kimi K3 costs $3 per million input tokens on a cache miss, $0.30 on a cache hit, and $15 per million output tokens, with no long-context surcharge. Claude Fable 5 charges $10 and $50 for the same. GPT-5.6 Sol charges $5 and $30. Run the output math and Kimi K3 is roughly a third the cost of GPT-5.6 Sol and a fifth the cost of Claude Fable 5, for a model landing second or third on several of the benchmarks that matter most to developers actually shipping code.\n\nFrankly, that's the story here, not the parameter count. A 2.8-trillion-parameter model is an impressive number to put in a headline, but developers don't pay for total parameters. They pay per token, and on that measure Moonshot just forced Anthropic and OpenAI to defend prices that suddenly look a lot harder to justify. This is also the third major Chinese lab in a year to make the same move: DeepSeek did it with V3 and R1, GLM did it with its 5-series, and now Moonshot has done it with a model whose scores anyone can actually check. Whether Kimi K3 holds up under sustained production traffic, rather than a leaderboard run, is the test that's left. But unlike DeepSeek's still-unverified August claims, that test is already running in the open, on infrastructure anyone can rent by the token.\n\n**Also read:** [A Federal Judge Ruled Judges Are Immune Even If AI Wrote Their Rulings](https://startupfortune.com/a-federal-judge-ruled-judges-are-immune-even-if-ai-wrote-their-rulings/) • [CoreWeave Posts a Blowout Quarter Right as AI Bubble Fears Peak](https://startupfortune.com/coreweave-posts-a-blowout-quarter-right-as-ai-bubble-fears-peak/) • [Binance Lets AI Agents Trade Your Crypto With No Cap on Losses](https://startupfortune.com/binance-lets-ai-agents-trade-your-crypto-with-no-cap-on-losses/)", "url": "https://wpnews.pro/news/moonshot-ai-s-kimi-k3-undercuts-gpt-and-claude-while-proving-its-numbers", "canonical_source": "https://startupfortune.com/moonshot-ais-kimi-k3-undercuts-gpt-and-claude-while-proving-its-numbers/", "published_at": "2026-08-20 15:43:43+00:00", "updated_at": "2026-08-20 16:15:17.032070+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-products", "ai-research"], "entities": ["Moonshot AI", "Kimi K3", "DeepSeek", "OpenRouter", "Together AI", "Tom's Hardware", "Anthropic", "OpenAI"], "alternates": {"html": "https://wpnews.pro/news/moonshot-ai-s-kimi-k3-undercuts-gpt-and-claude-while-proving-its-numbers", "markdown": "https://wpnews.pro/news/moonshot-ai-s-kimi-k3-undercuts-gpt-and-claude-while-proving-its-numbers.md", "text": "https://wpnews.pro/news/moonshot-ai-s-kimi-k3-undercuts-gpt-and-claude-while-proving-its-numbers.txt", "jsonld": "https://wpnews.pro/news/moonshot-ai-s-kimi-k3-undercuts-gpt-and-claude-while-proving-its-numbers.jsonld"}}