Sponsored by
BREAKING STORY
Mistral opens the weights on its trillion-parameter Large 4
Mistral has released open weights for Mistral Large 4, the 1-trillion-parameter model it first previewed earlier this month (TechCrunch, 29 Oct). The Hugging Face model card describes a multimodal mixture-of-experts model with 49 billion active parameters per token, with Low, Medium and High reasoning modes.
The licence is a modified MIT licence that allows commercial use, but anyone running a model-as-a-service business needs a separate commercial licence from Mistral. API pricing is $0.70 per million input tokens and $2.10 per million output tokens ($0.07 cached input). Mistral calls it the best open-weight model for coding, claiming 48% on Artificial Analysis' Coding Agent Index, roughly level with GLM-5.3 while using 47% fewer tokens per task, and 41% on Terminal-Bench 4.0 (Mistral; SiliconANGLE, 29 Oct).
My take: This is a model Mistral ranks near the top for coding that you can own and run yourself, and its API output price is about a fifth of the $10 per million output tokens Google charges for Gemini 4 Carbon at its introductory rate. Price it out against your current API bill this week. Read the licence before you build a hosted service on it, though, because reselling it as a service needs a deal with Mistral.
A CRM so smart, it updates itself.
HubSpot's CRM is so smart, it now updates itself. Calls get logged and summarized the moment they end. New leads get researched and a first outreach drafted, automatically. Deals move forward on their own, with next steps flagged before you have to ask.
Even the one repetitive task you've been meaning to fix can now run itself, no code required. Customers using it are already seeing the difference.
This isn't the CRM you have to work hard to figure out. It's the one already working before you log in.
MARKET TRENDS
Agents now start more than 40% of pull requests on GitHub
GitHub's Octoverse 2026 (28 Oct) reports that "by September 2026, more than 40% of pull requests on GitHub started with an agent's first draft, up from 17% in March." That came to 110 million pull requests in September alone. The report also says 61% of Copilot users who selected a model themselves used more than one, and no single model carried more than 25% of requests (The Next Web, 28 Oct).
How to make money from it: Agents are writing a lot of code, and someone still has to check it. Offer code review and "agent clean-up" to non-technical founders whose apps were built by agents. Package what you know as a review checklist or a short course on vetting agent-written pull requests. If you already mix models, document that workflow; most Copilot users who pick their own model do too.
GROWTH HACK / DIY
Ask first when the first prompt is vague
In an AI builder, the first output is the activation moment, and newcomers write the vaguest prompts. "Make me a booking app" leaves out who books what and whether anyone logs in, so the first build misses and the user leaves. The fix: when the first prompt leaves something important open, ask two or three quick multiple-choice questions before building. Let people skip, and send template prompts straight to the build.
On 5 October, Base44 shipped this. Its changelog says the AI now asks before it builds when a first prompt "does not say what the app does or who it is for". Users can pick answers, write their own or skip, and gallery and store prompts build straight away. GoDaddy's Airo AI Builder added clarifying questions in March. Neither has published results.
3 steps
Find the details that matter. Review 50 to 100 first prompts that ended in a regenerate or exit, and list the three to five missing details behind the misses.
Gate on ambiguity. A small model call checks whether a high-impact detail is missing. Only then show two or three questions with three or four options each, a "write your own" box and an obvious Skip.
Test it. Run it for half of new users and compare how often first builds are kept, skip rate, time to first output and credits per activated user.
What you'll need: a cheap ambiguity check, a chip-style question UI, a bypass flag for template prompts, and event tracking.
AGENT LAB
When your agent crashes mid-task, don’t pay for the same work twice
The pattern. When a serverless function dies mid-turn, most agents lose the turn. You have paid for the model calls that already ran, and you don't know whether the refund went out. In Making every harness durable (19 Oct 2026), Vercel's fix is a journal. Every model or tool call writes an intent record before it runs and an outcome record when it finishes. A queue message backs each turn, so a new worker can pick up a crashed one.
How it works. The new worker replays from the journal. Finished calls are read back, not re-run, which Vercel says means you are never billed twice for them. Each tool is labelled safe or never. If a crash interrupts a safe lookup, the lookup runs again. An interrupted never tool, such as a refund, does not run again. Instead, the model is told it may or may not have happened so it can check first. You don't need Vercel. Any database plus a queue with leases will do.
Try it. Build a SQLite journal table and two tools: lookup_note (safe) and send_email (never, pointed at a test inbox). Kill the worker after each step and restart it. You want exactly one email every time. Then relabel send_email as safe and watch the duplicate arrive.
TRENDING AI RESEARCH
TokenEconomizer: cut your coding agent's bill by 18–31%
What it says. A UC Berkeley and Microsoft paper (v1, 15 Oct 2026) found that coding agents waste tokens in repeatable ways when you use them on the same repo day after day. In their tests the agent was shown 23 tools every turn but no repo used more than 7, and failed runs were 17% of tasks but 32% of spend. An outside controller that adds short repo notes, hides unused tools and interrupts failure loops cut average token cost by 18.4–30.9% across Claude Code, Codex and Copilot CLI, at about the same success rate. Web-dev and security-audit savings were smaller, at 6.4–21.2%.
Why it matters. None of this needs a new model. In the paper, pruning tools was the biggest single saving. The catch: most repos cost more at first and only became cheaper by about the sixth task, and the test streams were not real user sessions.
How to use it.
Disable MCP servers and tools you have not used in that repo this week.
Put your test command, source layout and skippable setup steps in CLAUDE.md or AGENTS.md, one fact per line.
Add a rule: if a fix fails twice, stop and summarise.
Or try Microsoft's open-source plugin.
TOOL REVIEW
Spacesheep: publish what your AI agents make and revise it in the comments
Spacesheep, which launched on Product Hunt on 29 October, turns the docs, dashboards, mockups and small apps your agents make into shareable pages called "spaces". You connect it to Claude, ChatGPT, Claude Code, Codex or Cursor through an MCP server. Each space can be private, link-only or public. Viewers comment on a specific line, and tagging @spacesheep in a comment asks its AI to make the fix, with every version kept.
Price (as of 30 Oct 2026): Free; Pro $4.50/mo; Max $9/mo with Claude Opus tokens; Team $10 per member/mo; Enterprise by quote.
For: solo builders who send AI-made reports, dashboards or prototypes to clients and use more than one agent.
Verdict: A cheap, well-designed way to share agent output and keep iterating on it, but it's brand new and the AI platforms are building their own sharing features. Try the free plan on one real deliverable before paying.
WHAT GOT FUNDED YESTERDAY
Wabi (wabi.ai) · consumer AI · US. Founded by Eugenia Kuyda, who previously built Replika. Wabi is a personal AI assistant that builds mini-apps, such as trivia games or book-club trackers, for friends to share. It says the next era of personal AI "happens between people". It raised $55M at a $1.25B valuation, led by IVP, with a16z, Forerunner, the Amazon Alexa Fund, Tobi Lütke and Marc Benioff.
Mintive (mintive.com) · AI infrastructure · US. Founders Jiquan Ngiam (CEO), Vijay Vasudevan (CTO), Bozhi See and Phillip Bensaid. Formerly MintMCP, it gives enterprise AI agents their own identity, permissions and audit trail. It says it treats each agent like an employee and is model-agnostic. A $7M seed was led by Cowboy Ventures, with Coatue, Operator Collective, Jeff Dean and Andrej Karpathy.
Ravical (ravical.com) · accounting AI · Belgium. Founders Joris Van Der Gucht (CEO), Ken Bastiaensen (CTO) and Benjamin Vandermarliere (CPO). It turns senior accountants' know-how into shared processes for the whole firm. It says personal assistants make individuals better "but the firm doesn't". It says 100+ firms use it. The $17M seed was led by MMC Ventures, with Lakestar, Pitchdrive and Shapers.
Braid (nextcommunications.io) · embedded connectivity · Netherlands · first reported 28 Oct. Founders Jignesh Dave (CEO) and Mitesh Popat. Formerly Next Communications, it offers one API that lets banks, airlines and other brands sell mobile data inside their own apps, across 600+ carriers. It says it is profitable and had never raised outside money. Its $38M Series A was led by Octopus Ventures, with KPN Ventures.
Crux Labs (crux-labs.ai) · AI agent evaluation · US. Founders Julie Kheyfets (CEO) and Krishna Kulkarni (CTO). Its agents learn from a customer's own experts, write custom tests and run them on all live agent traffic. It says it is the first platform to turn enterprise judgement into continuous agent improvement. It raised over $6M, co-led by Benchstrength and Misneach, with Analog Ventures and Pareto Holdings.
QUICK HITS
Nvidia released Nemotron 3.5 Transcribe, a free open-weight model that covers 40 languages and tells up to eight speakers apart (TechCrunch, 28 Oct).
Trakkr traced 1,285 statistics cited by five AI assistants and found only 40% led to a study saying the same thing (28 Oct).
LlamaIndex published a draft Document Parsing Protocol with 140 rules for what a correctly parsed page looks like (28 Oct).
That's today. Hit reply and tell me what you're building. I read every email.
Dev
