Kimi Allegretto ($39/month, 150 agent credits, 50 Agent Swarm runs with 4 concurrent, Kimi Code 5x, Kimi Claw browser agent, Kimi K3 model) vs Claude Sonnet 5 ($20/month Pro or $2/M API through August 31 then $3/M, best writing quality, 128K output, adaptive thinking, Anthropic FLI C+). Allegretto is purpose-built for agentic coding workflows and costs 95% more. Sonnet 5 is the writing and analysis default and is now cheaper than Allegretto before August 31.
Large Language Models04 Aug 20266 min read
Three August 3 stories. Moonshot confirmed kimi-k2.5 and all moonshot-v1 strings sunset August 31 — 28 days to migrate. Anthropic published its open-weights position paper July 27 — conditional release framework, not a ban. And Qwen3.8-Max: the model previewed July 19 is not an August launch, has no benchmark table, no per-token API pricing (the $2/$6/M figure on X is unverified), no open-weight date, and Alibaba's "second only to Fable 5" claim has no third-party verification.
News03 Aug 20265 min read
Moonshot AI has confirmed that kimi-k2.5 and the entire moonshot-v1 model series will be fully discontinued on August 31, 2026 — the same day Claude Sonnet 5 intro pricing ends. Both model strings are already unavailable to newly registered Kimi API users following the K3 launch on July 16. Existing users who still call kimi-k2.5 or any moonshot-v1 string have 28 days to migrate. Recommended migration target: kimi-k3 for intelligence-requiring tasks, or kimi-k2.7-code / kimi-for-coding for coding-specific workloads.
Large Language Models03 Aug 20265 min read
Kimi Allegretto ($39/month) vs Claude Pro ($20/month). Allegretto doubles the price of Moderato and adds Kimi Claw browser agent, 150 agent credits, 50 Agent Swarm runs with 4 concurrent subtasks, and Kimi Code at 5x. Claude Pro gives you Claude Opus 5 — the Arena.ai #1 Frontend Code model — with unlimited messages (soft limit applies), Projects, and voice with Gmail/Slack connectors. The right choice depends entirely on whether your work is agentic/coding-heavy or writing/analysis-heavy.
Large Language Models31 Jul 20267 min read
Three July 28 stories: Treasury threatens Moonshot sanctions over Fable 5 distillation — China warns all necessary measures — Nvidia/Meta/Microsoft/OpenAI oppose broad ban. OpenAI IPO: $852B, $2B/month, not profitable, September-November window. HF used Chinese GLM-5.2 to stop the OpenAI agent attack — US models with safety guardrails were less effective.
News28 Jul 20266 min read
Kimi Code ($19/mo, K3 model, SWE Marathon #1 at 42.0%) vs OpenAI Codex (ChatGPT Plus $20/mo, GPT-5.6 family, Terminal-Bench #1 at 88.8%). Different benchmark strengths, nearly identical subscription prices. Kimi Code leads long-horizon agentic runs. Codex leads terminal task execution. Here is when each is the right call — with pricing, benchmark table, and a switching note: you can run the Kimi K3 model through Claude Code's harness via the Anthropic-compatible endpoint.
Code Tools27 Jul 20268 min read
Kimi K3 open weights went live on July 27, 2026 at huggingface.co/moonshotai/Kimi-K3 under a Modified MIT license. The 2.8 trillion-parameter model is the largest open-weight model ever released. Download size: ~594GB MXFP4 safetensors. vLLM with KDA prefill cache support is available. Minimum hardware: 8× H100 80GB to load the model. Community GGUF Q4 quantised builds expected within 24 hours. Key facts: Modified MIT license allows commercial use. Important caveat: independent testing found a 51% hallucination rate not disclosed in Moonshot's benchmark charts.
Large Language Models27 Jul 20266 min read
Kimi K3 open weights go live tonight July 26 at 8PM ET (July 27, 00:00 UTC) on Hugging Face at huggingface.co/moonshotai. The download is approximately 594GB for the native MXFP4 safetensors format. License: Modified MIT (matches K2 series — confirm on the model card). vLLM with KDA prefill cache support ships alongside the weights. Minimum hardware: 8× H100 80GB to load the model. Community GGUF Q4 quants expected within 24 hours of weights landing.
Large Language Models26 Jul 20265 min read
Kimi K3 open weights release Sunday July 27, 2026 — two days away. The 2.8 trillion parameter MoE model will be the largest open-weight model ever released. This guide covers what you need to self-host: GPU requirements (minimum 64 accelerators recommended by Moonshot), quantisation options (MXFP4, Q4, Q8), estimated cost on AWS/Azure/GCP, and why self-hosting eliminates the China National Intelligence Law data residency concern that blocks enterprise use of the Moonshot-hosted API.
Large Language Models25 Jul 20267 min read
Moonshot AI launched Kimi K3 at Shanghai's World AI Conference on July 16, 2026. At 2.8 trillion parameters — the largest open-weight model ever built — K3 ranks #4 globally on the Artificial Analysis Intelligence Index (score 57), leads SWE Marathon at 42.0%, and tops Design Arena frontend coding at 1679 Elo. Pricing: $3/$15 per million tokens ($0.30/M cached input). Full 1M context window on Allegretto+ tier. Open weights by July 27. The three-tier Kimi subscription structure (Moderato, Allegretto, Vivace) and the China National Intelligence Law data residency consideration explained in full.
🔥 TrendingLarge Language Models21 Jul 20268 min read
Kimi Code CLI and Claude Code are the two most capable terminal-based AI coding agents in 2026. Claude Code leads on benchmark scores (Claude Opus 4.7 at 64.3% SWE-Bench Pro vs Kimi K2.6 at 58.6%) and enterprise reliability. Kimi Code CLI costs significantly less per token and supports native 300-agent parallel swarms. Here is the full comparison for developers choosing between them.
22 May 20268 min read
Kimi K2.6 ties GPT-5.5 on SWE-Bench Pro (both 58.6%) and leads on Humanity's Last Exam with tools (54.0% vs 52.1%) — at $0.95 vs $5.00 per million input tokens. GPT-5.5 leads on pure math (AIME 99.2% vs 96.4%), context window (1M vs 256K tokens), and native computer use. Here is the full head-to-head.
22 May 20268 min read
Kimi K2.6 is Moonshot AI's open-weight 1-trillion parameter model released April 20, 2026. It ties GPT-5.5 on SWE-Bench Pro (58.6%), leads on Humanity's Last Exam with tools (54.0% vs GPT-5.5's 52.1%), and costs $0.95/$4.00 per million tokens — roughly 80% less than GPT-5.5. Agent Swarm coordinates up to 300 sub-agents in parallel. Here is the full tested review.
22 May 20269 min read
Kimi K2.6 (API: $0.60/$2.50 per million tokens), Claude Code (Pro $20/mo, Max $100-$200/mo), and OpenAI Codex (included with ChatGPT Plus/Pro) all claim the top spot for AI-assisted coding in 2026. We ran benchmarks, tested real workflows, and broke down exactly where each tool wins and where it falls short.
02 May 202610 min read
Kimi K2.6 is Moonshot AI's open-weight agentic model released April 20, 2026 — a 1-trillion-parameter MoE model with 32B active parameters, a 256K context window, and multimodal input. It beats GPT-5.4 on SWE-Bench Pro (58.6% vs 57.7%) and undercuts Claude Opus 4.6 API pricing by roughly 8x. Here is what it actually does, what it costs, and who should use it.
23 Apr 20269 min read
Moonshot AI launched Kimi K2.6 on April 20, 2026 — a 1 trillion parameter open-source model that leads coding and agent benchmarks with 80.2% on SWE-Bench Verified and 54.0% on Humanity's Last Exam with tools. API access starts at $0.60 per million input tokens, undercutting Claude Sonnet by 80%. Here is everything developers need to know.
Code Tools21 Apr 20268 min read