In partnership with
The cheap end of AI got cheaper this weekend. DeepSeek released a free-to-download coding model with a web search tool priced far below rivals, Disney picked up most of a bankrupt chatbot startup's staff, and Tesla finally showed a Roadster that jumps. We checked the model card, the court filing coverage and the tool pages before writing a word.
Turn AI into Your Income Engine
Ready to transform artificial intelligence from a buzzword into your personal revenue generator?
HubSpot’s groundbreaking guide "200+ AI-Powered Income Ideas" is your gateway to financial innovation in the digital age.
Inside you'll discover:
A curated collection of 200+ profitable opportunities spanning content creation, e-commerce, gaming, and emerging digital markets—each vetted for real-world potential
Step-by-step implementation guides designed for beginners, making AI accessible regardless of your technical background
Cutting-edge strategies aligned with current market trends, ensuring your ventures stay ahead of the curve
Download your guide today and unlock a future where artificial intelligence powers your success. Your next income stream is waiting.
TODAY IN AI
3 things that happened while you were busy
1. DeepSeek's new Flash model beats Opus 5 on one coding benchmark, by its own numbers.
DeepSeek released V4.1-Flash-1017 on Saturday, an update to its cheaper Flash tier, with weights free under an MIT license. On DeepSeek's own model card it scores 77.0 on DeepSWE v1.1 against 74.0 for Claude Opus 5, while Opus 5 stays well ahead on Humanity's Last Exam, 56.3 to 42.6, and on Terminal-Bench 4.0. The model has 763B parameters and a 1M-token context window. See the full benchmark table on Hugging Face. Independent reproductions of the coding score have not appeared yet, so treat it as a company claim.
2. Disney will hire most of Character.AI's staff after the startup's bankruptcy.
Character.AI filed for bankruptcy last week, and a legal filing reported by Bloomberg says Disney agreed to make job offers to at least 60 employees, the majority of the workforce, at an estimated $60M a year in compensation. Disney had already agreed to buy some assets for $20M, and former Character.AI CEO Karandeep Anand recently became Disney's first CTO. Read TechCrunch's report. The deal is a stalking horse bid, which means another company could still outbid it at auction.
3. Tesla shows a Roadster that jumps over other cars with compressed-air thrusters.
At an event outside Waco, Texas, Tesla's redesigned Roadster used 10 cold-gas thrusters fed by 600 liters of compressed air to lift itself over a Cybertruck and spin in midair, according to engineering VP Lars Moravy. Tesla claims 0 to 60 mph in one second and 300 miles of range on every trim, and Elon Musk said production starts in 2028, nine years after the car was first shown. Reservations cost $50,000 or $250,000 depending on trim. Read TechCrunch's account of the demo.
A 4X ROAS. But 22% of Orders Still Lost Money.
A 4X ROAS sounds pretty darn good.
And yet, across Walmart accounts managed by pdPlus last Q4, 22% of ad-attributed orders came from SKUs whose post-fee margin couldn’t cover the cost of the click.
A great holiday season on paper isn’t always a profitable one. Earnings can quietly disappear across ad spend, discounts, inventory decisions, creator payouts, and returns that hit after the confetti settles.
That’s where 9 Plays to Defend Your Margin This Holiday Season comes in. Nine ecommerce experts break down where margin gets lost across the P&L and how to fix it:
Inside, you’ll learn how to spot:
Healthy ROAS hiding money-losing SKUs
Discounts on products shoppers aren’t price-comparing
Ad dollars going toward inventory at risk of stocking out
Creator payouts capped by first-order math
Returns that turn into refunds instead of exchanges
Plus learn four more plays for keeping more of what you earn this Q4.
TODAY'S TUTORIAL
Stop your AI agent drowning in its own tool output
Agents fail quietly when a tool returns a huge log or search result and the whole thing lands in the conversation. In a test with AWS's open-source Strands framework, eight log fetches left a default agent holding about 252,000 tokens of history, and one built-in plugin cut that to about 9,000. Read the write-up. Here is the same discipline, applied to any agent you build or buy.
1. List the loud tools. Write down every tool your agent calls that can return a large result: logs, file reads, web search, database queries.
2. Store, do not paste. Have each loud tool save its full output to a file or store and return only a short summary plus an ID the agent can use to fetch details later. In Strands, the ContextOffloader plugin does this for you.
3. Cap repeat calls. Set a maximum number of calls per tool per task so a confused agent cannot retry the same search 14 times.
4. Give it a finish line. Add a clear "done" condition so the agent stops once it has an answer instead of looping.
5. Measure before and after. Run the same task both ways and compare token use, cost and answer quality. Fewer tokens is only a win if the answers hold up.
Good to know: The Strands test used a scripted stand-in for the model, so it measures how much text stays in context, not whether the answers got better.
FROM THE FRONTIER
The price of routine AI work is heading toward zero
The price tag. DeepSeek's Search and Fetch tools cost $0.75 per 1,000 successful calls, with failed and empty calls free, per DeepSeek's own announcement. The Neuron notes rivals such as Brave, Tavily and Perplexity list search at $5 per 1,000 and up. Calls made through DeepSeek Harness, Claude Code and Codex stay free until billing starts on October 31 at 16:00 UTC. Read DeepSeek's announcement.
The multiplier. Open weights mean any cloud can host the model and compete on price. The Neuron reports OpenRouter already lists about a dozen providers for it. That pressure lands on bulk work first, such as routine coding and agent loops, rather than on the hardest reasoning, where Opus 5 still leads by a wide margin on the same model card. Compare the scores yourself.
The politics. Cheap Chinese models arrive with baggage. The Neuron points to a September advisory from the NSA, FBI and CISA accusing DeepSeek and Alibaba of distilling American models at industrial scale, and to a Chinese draft rule that would restrict exports of weights for the very largest models. Several governments also restrict DeepSeek on official devices, so confidential material should stay out of its hosted tools. Read The Register on the advisory.
The other ledger. While DeepSeek cuts prices, Anthropic pledged 5% of its shares to a new Public Benefit Fund, subject to stockholder approval, with the board able to add 2% for every 50% rise in company value above a baseline. The fund cannot be used for lobbying or political giving. Read Anthropic's announcement.
The reframe. For years the question was which model is smartest. The question for most businesses is now which model is good enough for the boring 80% of the work, and that answer keeps getting cheaper every month.
MONEY FORWARD
Three ways to turn this week's AI news into income
1. Cheap automation for small shops. With a capable open coding model hosted by a dozen providers, small scripts for invoices, inventory checks or report formatting cost pennies to run. Offer local businesses a fixed-price setup and a small monthly fee for upkeep. See DeepSeek's tool pricing.
2. Product photo editing at volume. Krea 3 lists about $0.06 per 2K image edit and accepts up to eight reference images, which suits background swaps for online sellers. Charge per catalog, not per image, and keep a human check for anything customer-facing. See Krea 3's pricing claims.
3. Scam-call screening setup. CallSitter is open source, but it needs a phone number provider, a Telegram bot, a tunnel and API keys. Older relatives and small clinics would pay someone to install and maintain it. See the setup requirements.
GET THE EARNINGS GUIDE
Want the step-by-step version of ideas like these? Get The Aiwire earnings guide at theaiwire.co. These are ideas, not income promises, and results depend on your skills, market and effort.
IN THE KNOW
What people are actually watching and sharing
Publishers pull up the drawbridge. Axios found 42% of 102 top US publishers now block AI training crawlers, against 25% that block AI search bots, as reported by The Neuron. Read the Axios analysis.
Samsung borrows TSMC's lines. Seoul Economic Daily reports Samsung will use TSMC's 3-nanometer process for the base layer of Nvidia's custom memory chips, which could raise bandwidth by up to 30%. Read the report.
No power until the 2030s. Microsoft pulled out of Nscale's planned Essex data center after grid officials said electricity would not arrive until next decade, according to SiliconANGLE citing the Telegraph. Read the SiliconANGLE story.
Flies live longer without magnetism. University of Nottingham researchers found fruit flies shielded from most of Earth's magnetic field lived about 20% longer, per Science Focus. Nobody knows why, or whether it applies to people. Read the Science Focus summary.
Apple's kitchen screen. Apple's HomeView hub pairs a 7-inch touchscreen with a conversational Siri, facial recognition and FaceTime. See the product page.
TRENDING TOOLS
5 free or low-cost tools worth a look this week
1. Krea 3. An image editing model that renders 2K natively and finishes an edit in under eight seconds, per Krea, at about $0.06 per image. Its speed comparisons are pre-generated, not a live benchmark. See Krea 3.
2. CallSitter. An open-source phone answerer under Apache-2.0 that wastes scammers' time, declines sales calls and takes messages, with a $20-a-day call spending cap built in. You need Telnyx, Telegram and OpenAI accounts. See the repo.
3. llmspeed. A free web simulator that estimates tokens per second for a local model on a given GPU, Mac or laptop before you buy hardware. The figures are estimates, some projected from specs. Try llmspeed.
4. Rubi. An MIT-licensed knowledge graph for coding agents that maps your repo and compacts context early, with Claude Code, Claude Desktop and Codex tested end to end. Its cost savings are the author's own benchmarks. See Rubi on GitHub.
5. DocuMemory. Turns AI chats into compact decision notes that Claude Code, Cursor and any MCP client can reuse. There is a 15-use free trial, then $12 a month. See DocuMemory.
PROMPT STATION
Plan a cheaper model switch in seconds
This prompt works in Claude or ChatGPT and turns a vague "should we switch to the cheap model" into a test plan with clear pass marks. Use it before November 1, when DeepSeek's free tool calls end and the real bills begin.
COPY AND PASTE THIS PROMPT
You are an AI cost analyst. We currently use [CURRENT MODEL] for [TASK, e.g. drafting support replies] at roughly [MONTHLY VOLUME] requests a month. We are considering [CANDIDATE MODEL] at [CANDIDATE PRICE]. Build a two-week switch test: 1) pick 30 real examples that cover easy, typical and hard cases, 2) define a pass mark for quality, speed and safety, 3) list which data must never leave our systems and whether this model's hosting allows that, 4) give a go or no-go rule with numbers, 5) estimate monthly savings if we switch only the easy and typical cases. Keep it under 300 words and flag every assumption.
Tip: Swap [TASK] for code review, product descriptions or meeting notes, and [CANDIDATE MODEL] for DeepSeek V4.1 Flash, Gemini Flash or an open model on your own server. For a tougher test, paste three of your worst real examples and ask it to predict which one the cheap model will fail.