A model doesn’t have to be #1 at everything to be useful. Kimi K3.1 Pro is a good example. It scored 55 on Artificial Analysis’s Intelligence Index, compared with 58 for the highest-scoring model in the chart above. So it isn’t leading that leaderboard, but it’s still competitive.[ppl-ai-file-upload.s3.amazonaws] The interesting part is its value: what it can help you do, and what you pay for it. What is it good at? 💻 Coding and bigger projects Kimi is built for coding and tasks that take multiple steps—not just answering one question. Think working through a project rather than generating a single snippet of code. Coding is one of the main areas Moonshot says it improved with this release.[kimi] 📚 Working with lots of information It supports up to a 1-million-token context window. That’s the amount of information the model can work with at once, making it interesting for large codebases or long collections of documents. In Kimi Code, access to that full context window depends on your plan.[kimi] 🎮 3D creation and research Moonshot also highlights improvements in 3D creation and financial research. Those are areas worth testing not a guarantee that it will nail every project on the first try.[kimi] Why use it instead of a bigger-name model? Price is one reason. For developers connecting it to apps or automations, the standard API costs $2 per million input tokens and $10 per million output tokens—half the listed token price of Claude Opus 5.5. Tokens are the small pieces of text an AI reads and writes.[simonwillison] That doesn’t automatically mean every finished task costs half as much. But it gives you a reason to compare it on your actual work instead of choosing based on leaderboard position alone.

What just dropped: Inception released Mercury 3 yesterday. It's a "diffusion" AI model, a different kind of AI than ChatGPT or Claude, and it writes over 1,000 tokens per second. The simple version (how it's different): Most AI models (ChatGPT, Claude, Gemini) write like a typewriter: one word at a time, left to right. Mercury works more like an AI image generator. It sketches a rough version of the whole answer at once, then sharpens it in a few passes until it's clean. That's why it's so fast. Why this one matters: A lot of people doubted this approach could ever get smart enough to compete. Mercury 3 is over 40x bigger than Mercury 1, and Inception says it matches or beats GPT-6 Luna and Claude Haiku 5.5 (the fast, cheap versions of the big models) while running 4-8x faster. Quick specs: - Speed: 1,068 tokens per second - Memory: 1 million tokens of context (you can paste in a LOT of documents) - Price: $0.20 per million input tokens, $0.70 per million output tokens - Launch deal: 50% off right now ($0.10 / $0.35). They haven't said when the discount ends. The free part (read before you sign up): - New accounts on the Inception API platform get 100 million free tokens. No credit card needed to start. - It's a one-time credit, not monthly, and it's shared across all their models. - When it runs out, you add a card to keep going. - The free tier has per-minute rate limits, which is plenty for testing. - The free tokens are for the API. The chat site is a separate way to try it. The honest part: - The benchmark wins and quotes come from Inception and its early partners. Test it on your own tasks before you switch anything. - Inception compares it to the fast, cheap tier of models. For your hardest reasoning work, your top model is probably still the better pick. Where it shines: Anything where waiting ruins the experience: voice agents, support bots, tutoring apps, and browser agents that take lots of steps. Small delays add up fast when an agent makes dozens of calls.

may have just reached self replication, franework i made for that is pretty easy, if you want to take a look at it yall can send a dm to me, will be commenting below later after opensourcing a small technical report on that soon

What just dropped: Eagle released EAGLE-4 today. It makes AI models write their answers way faster without changing what they say. The simple version (how it works): Normally an AI writes one word at a time, and every word takes the big model a full "think." EAGLE-4 adds a tiny helper model (called a "drafter") that quickly guesses the next several words. The big model checks all those guesses in one go and keeps the ones it agrees with. Same answer you would have gotten anyway, just way less waiting. What's new in version 4: one drafter now adjusts itself to your hardware and to how many people are using the model at once. No hand-tuning. The numbers they posted: - Qwen3.8-27B: 961.6 tokens/sec vs 240.5 normally (about 4x faster) on 4 NVIDIA B300 GPUs - 512 users at once: 28,340 vs 21,203 tokens/sec (1.34x faster) - AMD MI350X: 3.49x faster for a single request - On a regular phone (Motorola Razr 70 Ultra): about 3.6x faster on coding and 53% less energy per word The honest part (read this before you get hyped): - This is for people running their own models on servers. It plugs into vLLM and SGLang. It is not an app you download, and it will not speed up ChatGPT or Claude on your end. - It has only been fully tested on NVIDIA B300 data-center GPUs. Other NVIDIA cards (A100 and newer) and AMD cards are supported on paper, but not fully tested yet. - Right now there is one ready-made drafter, for Qwen3.8-27B. Text only, no images. - The phone results are a demo. The public GitHub release is for GPUs, and I did not see a phone setup in it. - It is "source-available," not true open source. It's free to use unless your company makes over $20M a year or you are building a competing product. Then you need a paid license. Why you should care even if you never install it: Speed like this is what makes AI agents (Claude Code-style tools) feel instant, and it's how AI companies cut costs. Expect the APIs and open models you use to get faster and cheaper as providers adopt it.

yo so I'm doing a live Q&A this thursday, oct 22 at 6pm denver time that's 5pm pacific and 8pm eastern I want to actually talk to you guys face to face a lot of you are founders and devs building stuff and I barely know what you're working on I've built a bunch of apps with nobody behind them so this is me building the people part first bring whatever you're stuck on AI tools, picking models, getting your first users, anything join here: https://www.skool.com/live/pFBJJsvz9bf if you want in, reply to this post with your name what you're building one question you want me to answer live I'm reading every reply before thursday so I can actually prep for you