In partnership with
What's Actually Happening
DeepSeek just handed developers the keys to its own app. As of today, the features behind the DeepSeek chatbot (web search with citations, link reading, file uploads, a safety filter and voice) are switches you can flip in the DeepSeek API.
That turns a weekend of stitching together search, document parsing and speech vendors into a few request parameters, at search prices that undercut the big Western APIs by a wide margin.
Also tonight: Edge0 releases a 145B model built to run on one GPU, IBM previews small Granite models trained to shrug off prompt injections, Jensen Huang resigns as Nvidia's CEO, and four tools worth trying this weekend.
DeepSeek opened the features behind its consumer app to developers today. With a few request parameters, the DeepSeek API can now search the web and cite its sources, read links users paste into a chat, pull in PDF, Word, PowerPoint and image files, run a safety check on inputs and outputs, and title conversations automatically. Speech-to-text, text-to-speech and OCR are available as standalone APIs.
It runs on deepseek-flash, which DeepSeek now serves with DeepSeek-V4.1-Flash-1017, across three endpoints: an Anthropic-style Messages API, a Responses API and Chat Completions, so existing OpenAI or Anthropic SDK code mostly needs a new base URL. Turn thinking on and the model can search and read pages across several reasoning steps, with replies coming back as structured blocks for thinking, search steps, page reads and citations. DeepSeek's pricing page lists Search and Fetch at $0.75 per 1,000 successful requests, on top of flash token prices that start at $0.15 per million input tokens off-peak.
Our take: this is DeepSeek turning its app into a kit. Building a ChatGPT-style assistant used to mean wiring up a search provider, a document parser, a speech vendor and a moderation layer; now one API key covers all of it. The caveats are in DeepSeek's own guide: with search, link reading, safety or titles switched on, the service overrides settings like temperature; the safety check is described as basic risk control; and anyone serving the public still owns their own compliance. The open-source demo is for demonstration only and keeps your API key in the browser.
Will you build with DeepSeek's new app API this week?
Go deeper: Full pricing · Open-source chat demo
🔎 Why It Matters: Search Just Got Cheap
Web search is the add-on that quietly blows up AI app budgets. The big Western model APIs have generally charged around $10 per 1,000 searches; DeepSeek is listing search and page fetches at $0.75 per 1,000. For an agent that searches dozens of times per task, that gap is the difference between a feature you ration and one you leave on by default. Bundling file parsing, OCR and voice under the same key removes three more vendor contracts from a typical build.
It also shows where the competition is heading. Chinese labs have spent the past year undercutting on raw tokens; now DeepSeek is competing on the whole application stack, alongside its standalone Search and Fetch API and its DeepSeek Harness preview for agent builders. The trade-off hasn't changed: your users' queries and files go through DeepSeek's platform, and its guide reminds developers that obligations like content review, model registration and AI labelling are theirs. Read the platform terms before you ship it to customers.
A Word From Today's Sponsor
The voice-to-text app that built its own speech model
Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.
It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.
Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.
Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.
Let’s Talk MODELS! 🤖
Edge0 released Edge0-145B and a smaller Edge0-120B on November 12, pitching the bigger one as frontier-class AI that runs on your own computer. It's a latent mixture-of-experts model built on Qwen3.8-Flash-Next: 145 billion parameters in total, but only about 10 billion do work on each step, and groups of 64 experts share two large matrices to save memory. Edge0 says it peaks at about 29GB of memory and decodes at 80+ tokens per second.
The model card is more specific about hardware: the Windows build needs an NVIDIA GPU with at least 16GB of video memory, 20GB of system RAM and CUDA 13, and the 3-bit download is about 41GB. On an RTX 4090, Edge0 measured 91.7 tokens per second in agent work. On its own benchmarks, it scores 86.5 on Terminal-Bench 2.1 and averages 72.8 across six tests, ahead of its Qwen base model, though it trails on harder reasoning tests like HLE and GPQA-Diamond.
Our take: a model this size on a gaming GPU is a real milestone for private, offline agents, and there are no token bills. Two caveats: the benchmarks are Edge0's own, with a technical report still to come, and "open" here means open weights under the Edge0 Community Model License, which allows personal and research use but needs a separate license to sell it or run it as a commercial service.
Try it: Edge0 app (v0.2.0) · 145B weights · Deployment guide
Small Models, Fewer Hijacks
IBM released two experimental models on November 12, granite-4.1-3b-llmon-preview and granite-4.1-8b-llmon-preview, built around LLMON, a markup language designed for models rather than people. Reserved tokens in the tokenizer mark which parts of a prompt are instructions, which are data and which are requirements, and an execution block ties each instruction to the inputs it's allowed to act on. IBM trained that behavior in from the base checkpoints rather than bolting on a filter.
The target is indirect prompt injection: instructions hidden in a web page, email or file that hijack an agent reading it. On the OPI benchmark, IBM says LLMON "practically eliminates the problem," avoiding 47% more attacks than the best frontier model it tested and 95% more than the best open model, with similar trends on ASB and StruQ, while matching Granite 4.1 Instruct on standard benchmarks.
Our take: prompt injection is the reason most companies won't let agents read the open web unsupervised, so structural fixes like this matter more than another point on a leaderboard. These are previews IBM says aren't for production, the results are IBM's own, and the approach only helps if your app actually marks up its prompts in LLMON.
Try it: 3B weights · 8B weights · LLMON toolkit · Attention visualizer
Top 5 In AI Research 🔬
The other stories worth your attention:
Jensen Huang resigned as Nvidia's CEO, effective immediately, after 33 years, with CFO Colette Kress named interim CEO while the board searches for a successor. The release gives no reason; see who the board named as its first-ever chair. (Nvidia, via SEC filing)
ChatGPT Wallet lets OpenAI's dots agents check out for you, using tokenized Visa, Mastercard and Amex cards so the agent never sees your card number, plus a Face ID or fingerprint approval on every purchase. It's limited to one plan in one country for now; see who can use it. (OpenAI)
OpenAI launched two security agents: Daybreak Red Team pen-tests web apps, and Daybreak Defense Factory turns flaws into merge-ready patches using GPT-6 Cyber. Red Team is in early access; read Greg Brockman's blunt warning to companies that don't adopt tools like these. (The Register)
Anthropic launched Claude for Manufacturing, a plugin of 17 skills for the plant office, from quoting jobs to diagnosing breakdowns, with connectors for Autodesk Fusion, Infor, Microsoft 365 and NetSuite. It's on Team and Enterprise plans; see the three-word setup prompt. (Anthropic)
Four nonprofits sued five AI labs to freeze frontier development, asking a San Francisco court for an emergency injunction against Anthropic, Google, Meta, OpenAI and SpaceX. See the California law they're using to make the case. (Al Jazeera)
🛠️ Tools That Are Hot Right Now!
🦙 llama.cpp 1.0 - run open models on your own machine with the engine behind much of local AI, now on a stable, versioned API so updates stop breaking your apps. Free and open source; see how many model architectures it supports.
🖥️ Zellij 0.46 - juggle several coding agents in one terminal workspace, with a new Attention Inbox that flags which pane is waiting on you. Free and open source; see what else landed in the release.
📋 Project OS - give Claude Code and Codex a shared project manager with task files, developer and QA roles, and your sign-off before any Git action. Free under MIT; see how it stops runaway review loops.
🎨 klad - paint side by side with your coding agent: you draw with a pen while it paints through a command line, and every stroke stays editable. Free for personal use and small companies (source-available, built from source, Mac first).
What's The Recap?
DeepSeek opened the features behind its app to developers, putting web search with citations, link reading, file input, a safety check and voice into its API, with Search and Fetch at $0.75 per 1,000 requests. Edge0 released a 145B open-weight model built to run on a single GPU, IBM previewed small Granite models trained to ignore prompt injections, and Jensen Huang resigned as Nvidia's CEO. The thread: the parts it takes to build a real AI app are getting cheaper, smaller and closer to home.
What Do You Rate Today's Newsletter?
Stay building. 🤖
Our newest video covers Anthropic slowing down, Gemini 4 Hydrogen, OpenAI's new model and the shake-up at Nvidia.
📬 Read by 14K AI readers. Know someone who'd like it? Send them here to subscribe.