Tinker

AI didn't break the internet. It sent the bill. Millions of attack attempts. No organization says it was breached. Two vendors withdrew their claims. Inside: 8 AI predictions that missed, and a checklist if you run a site. Has your host patched yet?

For teams that kept regulated data away from the API, this changes the conversation. The records stay on your side and an automated agent does the review, so the open question is which workloads legal will now approve.

Hy4 is on OpenRouter at $0.417 per million input tokens, half off at launch. The official release replaces Hy4 Preview, with 1M tokens of context and image input for the first time. What you get: 1. 1,048,576 tokens of context, about 750,000 words in one go, and 65,536 output tokens 2. Image input alongside text, new for this family 3. Tool calling and structured outputs through a JSON schema 4. Benchmarks: 34.6 on the Artificial Analysis Intelligence Index, 81.0% on Vibe Code Bench v1.1, 70.6% on the Vals Legal Index, 39.4% on Terminal-Bench 4.0 Output is $1.251 per million, cache reads $0.021. One provider hosts it, at 50 tokens per second. The AA-Omniscience non-hallucination rate is 10.9%, so check grounded answers before you trust it in an agent loop. Not available through EU in-region routing. At this price and this context length it is the one to try first. Point your agent at a repo and run the loop before you move anything.

OpenAI's coding agent runs on your own disk, in your terminal, not in a browser tab. Codex CLI, four steps: 1. Install: npm, Homebrew cask, or the one-line script in the docs. Windows gets PowerShell. Prefer a binary? The release page has macOS arm64, macOS x86_64, Linux x86_64 and Linux arm64. 2. Run codex, pick Sign in with ChatGPT. Plus, Pro, Business, Edu and Enterprise cover it. An API key works too, but needs extra setup. 3. Run codex inside the project you want it to touch. It reads and writes the files on disk, not a copy in the cloud. 4. Editor: VS Code, Cursor or Windsurf. Desktop window: codex app. It's an alpha build. Don't pin anything to it yet.

Asking how a change can be verified gets the agent to name the check before it writes code. Real signups and real clicks catch bugs that a passing unit test never sees.

Ripple-65B-Preview: 1M token context at $0.20 per million input tokens and $0.60 output. It is on OpenRouter now, served in BF16 by a single provider. 65B parameters with native tool calling: it accepts tools and tool_choice for function calling, and response_format for JSON output without schema enforcement. Context window is 1,048,576 tokens, up to 131,072 completion tokens. Cache reads bill at $0.05 per million. Throughput is 117 tok/s at 1.96s latency. Uptime over the past 3 days is 83.41%, and there is no second provider to route around an error. EU in-region routing is not available. Cheap for the context. The uptime is a preview number. Run your agent loop against it before you move production traffic.

A 65 MB text-to-speech model now tops the CPU streaming benchmark for time to first audio. Phonon shipped two weeks ago and has passed 200,000 downloads. On Hugging Face's Open TTS Leaderboard it ranks first on intelligibility in German and Spanish. 1. Two sizes: 31m for English at 65 MB, 40m for English, French, German, Spanish and Portuguese at 75 MB. 2. Phonon 31m is Apache 2.0, free including commercial use. The 40m model is CC BY-NC, priced by user count if you ship it commercially. 3. Top three on intelligibility in every language evaluated, against models up to 100 times its size. 4. Runs on phones, laptops, single-board computers and in a browser tab. No server, no network, the text stays on the device. 5. In Transformers from v5.21, with a ptts runtime covering CPU, CUDA, Vulkan, Metal, WebGPU, Core ML and Snapdragon. At 65 MB and Apache 2.0, it is the one to try for a voice that works offline. Load gradium/phonon-31m-en in Transformers and generate a wav before you design around it.

Claude Marketplace just added new partner agents and products, and it now has over 25 in total. You can browse all of them today. Eligible enterprise customers can also pay for part of them from their committed Anthropic spend. What's new, in plain words: 1. Checkmarx One finds weak spots in code and helps teams fix them before the code ships. 2. Netlify puts apps and sites online, with a preview link for every change and no servers to set up. 3. Runway turns a written brief into video, images or audio, and edits footage from a plain sentence. 4. Zoom answers questions from meeting summaries and documents, showing each person only what they may see. 5. Asana AI Teammates handle multi-step work and send decisions that need judgment to the person who owns the task. Good news for the curious: you don't need to be a developer to follow this. Open a few listings and see what each would do for your team. Where work with Claude is heading just got exciting, and it's fun to watch.

Breaking!🔥OpenAI released GPT-Image-2.6, in two versions: Flare and Sunburst. Both are in ChatGPT and Codex from today, on every plan, desktop, mobile and web. The two things I care about: it does GIFs now, and it accepts reference images. On speed they say 50% faster. We'll see if that holds once you throw a real prompt at it. Pricing: $8/$5 input, $30 output, same for text-to-image and image-to-image.

Claude Sonnet 5.5 costs $10 per million output tokens. DeepSeek's newest model costs 60 cents off-peak. Airbnb leans on Qwen. Cursor built on Kimi. The next phase of AI is a price war, and China is set up to win it.

Finally! Lets see how it performs on local machines

And just like I said in my article , this is going to be come also a PERMANENT cost for business , to be able to defend properly. xitter.gq/altcoindaily/status/21…

If Codex stores the code itself, the agent stops being a visitor in someone else's repo. Branches, history and reviews could be designed around agents from the start.

One small open model can find objects, read text, and cut the subject out of a photo, all in one go. Tencent's WeChat Vision team released WeVisOne-VL in 2B and 4B sizes under Apache 2.0. Hosted demos are live for both, so you can try it today. What it does: 1. It draws boxes around the objects you name, points at them, and counts them. If the object isn't in the picture, it answers "None" instead of guessing. 2. It reads text in images, so you can pull words off signs and scans. 3. On dense detection, the 4B model scores 71.3 against 58.7 for LocateAnything-3B, per Tencent's own tables. The weak spot is saying no. It correctly rejects absent targets only 52.6% of the time on PR-Bench, and those numbers are self-reported. Still, if you've ever stitched three tools together to find things and cut them out, this one is worth an afternoon of poking.

Anthropic released a way to prove which AI model answered you, without revealing its weights. FlashProve is open source under Apache 2.0, but Anthropic calls it a research prototype that isn't maintained. 1. Proving Llama 4 Maverick, a 400B model, runs at about 110 tokens a second on one NVIDIA H100, roughly 82 words a second. 2. Each proof is about 18.78 MB and checks on an ordinary CPU in 17 seconds. 3. Proving costs about 70 to 95 times the GPU time of normal serving. 4. It only works with mixture-of-experts models for now, and requests are capped at 1,024 tokens. You can't use this in a chat app yet, and proving needs an H100. But the checker is public and runs on a normal Linux computer, so anyone can test the claim. It's a useful step toward knowing which model is actually answering you.

Tencent released Hy4 today, its new flagship open model, and the API price stayed where it was. 770 billion parameters in total, 49 billion active at once, 1 million tokens of context. First Hy model that reads images natively, charts and visual math included. It also does Computer Use: terminal commands and clicks inside desktop apps, trained on Blender, CAD and long jobs across several programs. Hallucinations fell from 15.76% to 5.45% against the August preview. GameDev +19.1%, GameLogic +9.8%, multimodal agent +36.8%, document tasks ~97% done. On 579 real financial tasks it leads the open models at 3.04, ahead of DeepSeek-V4.1-Flash at 2.97. On Tencent's own agent benchmark it matches or beats Claude Opus 5.5 at about one eighth of the output price. Still behind the top closed models on research math and very long coding sessions. $0.83 per million input tokens, $2.50 output.

An AI agent searched thousands of lung cancer compounds and proposed new designs in about an hour. You can try it today in Google's Antigravity app with its Science plugin. Google is also giving 10,000 scientists free access, if selected. The search started from two lung cancer drugs, crizotinib and lorlatinib. Lorlatinib is a ring with no free-twisting joints. Crizotinib has five. The most rigid compounds were the most potent. Work that used to take days of chemistry can now end in a shortlist. Every design still needs a lab test, but the lab starts with a sharper list.

Goldman Sachs just pulled AI coding tools from most of its traders and bankers. Only engineers keep Claude Code and GitHub Copilot now. Bloomberg reports it, citing internal guidelines from October. The reversal came fast: 1. In July, the bank's CIO Marco Argenti praised "vibe coding," where non-technical staff build a quick prototype with AI. 2. Now a spokesperson says the experimentation phase is over. 3. The reasons given are "productivity and risk management," and staff are reportedly frustrated. For the rest of us, the good news is that a prototype you build on a weekend belongs to you, and nobody needs to sign off on it. Keep poking at it at home, where you can experiment without a risk committee watching.

Minimax publish the plugin marketplace skills, along with the custom REA skill from the team! Try them out ! @MiniMax_Code

I can’t create skills anymore . I create them but they don’t show up in the menu. @grok also confirms he does not see any skill . I created several ones All of this happened after the usage limit issues of the other day . Anyone with the same problem ?

Claude Code won't help you rebuild an existing video game anymore. The restriction has been active for several days, and renaming the clone won't get around it. Addy Osmani, an Anthropic employee, confirmed on X that two sentences in the system prompt for Opus 5.5 do the job. What's still open to you: 1. Your own game, or a title in a familiar genre 2. Projects built with your own assets 3. Some users say the rule has caught original projects too It follows a wave of AI-built clones, including hundreds of unofficial browser ports of Halo and GTA Vice City. Osmani says the rules will keep changing. Your own game idea is still wide open, and that's the interesting part. Pick a concept that's entirely yours and see how far one afternoon gets you with Claude Code.