Issue 002 · 15 October 2026
Agents got cheaper, glasses got smarter, and AI finished the night sky
This week AI stopped asking for permission and started finishing work: a frontier model built for long multi-step jobs, a browser agent 76 times cheaper to run, smart glasses with a privacy catch, and a map of the sky that no telescope ever completed.
Last week the story was that AI could act. This week the receipts arrived: cheaper agents, agents inside your network gear and on your face, and one agent that closed a fifty-year gap in astronomy. The useful question is no longer what a model can say. It is what it can finish, and what it costs to let it try.
1. Gemini 4 Argon arrives, built for long multi-step work
Why it matters:
Google's new frontier model, Gemini 4 Argon, is pitched at long jobs rather than one-shot answers: code migrations, legal and finance analysis, cybersecurity defense. Google says it ranks first on the Vals Index, which scores models on work in finance, coding, legal and tax. Launch price is $2 per million input tokens and $10 per million output tokens. A token is the unit models bill by, roughly three quarters of a word. Text the model sees repeatedly can be cached at 95% off. Access: developers and enterprises through the API (the plug that lets one piece of software call another) in AI Studio, AI Ultra subscribers in Antigravity, and Ultra and Pro subscribers in the Gemini app and AI Mode in Search. Google's own example: in Chrome, Argon separated core page handling from a 6-million-line legacy folder inside a 40-million-line codebase, work expected to take months, finished in weeks. Google also reports agentic code submissions up more than 35% and rollbacks down nearly a third. Those are Google's numbers about Google.
How to use this:
If you already pay for Gemini Pro or Ultra, you have access in the app today. Give it one long, boring, multi-step task you have been postponing: a contract set to compare, a messy spreadsheet to reconcile, a migration plan to draft. Time the old way first. Ask your team what the same work costs in tokens, not vibes. Do not migrate your stack over one benchmark table.
The Bottom Line:
A 200-page document set costs well under a dollar of input at list price. The real saving is calendar time on work measured in weeks, and that only shows up if you measure before and after.
2. Asana cut its browser agent's model costs 76 times
Why it matters:
Asana runs a browser agent, software that clicks through websites, fills forms and gathers information without a person driving. OpenAI published a case study on how Asana made it cheaper. The waste was specific: the agent cached its fixed instructions and tool definitions but not the growing history of page text and screenshots, so every request re-sent that history at full price. It also dropped older screenshots and trimmed text at nearly every step. After a 144-run study across four frontier models, the optimized workflow averaged about $0.47 in estimated model costs and about four minutes per run: 76 times cheaper and five times faster than the original setup. The engineering work Asana estimates would have taken one to two months by hand took about a week. This is a vendor case study with estimated costs, not an audited bill. Treat the multipliers as direction, not gospel.
How to use this:
If you run any automation that reads pages or documents step by step, ask your team one question: are we paying to re-send the same growing history on every call? Caching and smarter retention are usually the first fix. Measure cost per completed run before and after. The 76 times figure will not transfer to your workflow, but the failure mode probably will.
The Bottom Line:
About 47 cents per run instead of the old production cost, a 76 times reduction by Asana's estimate. A week of engineering replacing one to two months of manual investigation.
3. Meta's Muse Wonder, tested instead of announced
Why it matters:
Meta launched Muse Wonder and Muse Wonder Full with claims of frontier agentic performance. Claire Vo, who runs the How I AI podcast, did what most launch coverage did not: she ran it headlessly through Claude Code against nearly 20 tasks covering prototyping, agentic workflows, coding, writing and creative work, then scored the results. Her read: Meta's "personal and professional agent" framing mostly holds, Muse ranked first on three agentic tasks, and its design output is not good. It is fast on one-shot work and slow on long agent runs. "Headlessly" means running the model without its own chat window, driven by another program, so you see the raw behavior. One reviewer's bench is not a lab benchmark. It is still more evidence than a launch post.
How to use this:
Before you switch models because a launch post said "frontier," find one person who tested it on your kind of work. If your work is agentic, Muse is worth a trial. If your work is design or brand, this review gives you a reason to wait. Watch the 24-minute episode at 1.5 speed before you spend a month migrating.
The Bottom Line:
No dollar figure. The saving is the month you do not waste migrating to a model that is worse at the thing you actually do.
4. Alibaba's Qwen glasses put AI on your face, with a recording catch
Why it matters:
Alibaba launched two screenless AI glasses in China on October 13. The Qwen N1 and N1 Pro use Qualcomm's Snapdragon AR1 chip, a 50-megapixel camera that records 4K video, and a 272mAh battery you can swap without turning the glasses off. Prices: 2,999 yuan for the N1 and 3,499 for the Pro, discounted at launch to 2,218 and 2,600. The Pro adds health sensors (heart rate, blood oxygen, skin temperature, heart rate variability, stress), eye tracking and iris payment verification through Alipay. Alibaba has published no clinical validation for the health readings, so do not treat them as medical. The catch is in the terms. eWeek found that once you turn on environmental sensing, ongoing recordings do not stop on their own; you must give an explicit command to end them. The AI Inspiration Album is on by default and can capture color photos automatically when it decides a moment is worth keeping. Alibaba says those photos only reach cloud AI after you upload them yourself. The terms cover Qwen glasses generally, not each model. China only for now, with no US price or date.
How to use this:
If you are considering wearables for field work or meetings, test three things before you approve them for staff: how a recording starts, how a person stops it, and whether the recording light stays visible. Turn off automatic photo capture before anyone walks into a room with confidential material. Write the consent rule down. "The glasses did it" is not a defense.
The Bottom Line:
Replaces the phone in your hand for translation, meeting notes and reminders. The hidden cost is a recording policy your team has to enforce, which is real work and lands on whoever owns compliance.
5. DeepMind and Singapore built scam detection that stays on your phone
Why it matters:
In Singapore, scams are more than half of all reported crimes and have cost more than S$4 billion since 2019. Up to 80% of those losses now come from social engineering, where someone builds a conversation with you before asking for money. DeepMind and GovTech Singapore fine-tuned a small Gemma model, a lightweight model small enough to run on a phone, on local scam conversations: fake officials, fake banks, impersonated local figures. The model runs on the device. It reads the conversation with your consent, warns you in the moment, and neither the Singapore government nor Google can see the conversation, because nothing goes to a cloud. "Fine-tuned" means extra training that specializes a general model for one job. The rollout is deliberately phased. No independent measure of money saved has been published yet.
How to use this:
If you build an app with messaging, this is the pattern to copy: a small model on the device, explicit consent, and a warning at the moment of pressure instead of a fraud report after the money is gone. If you are the customer, ask which of your bank and messaging apps will use it and when. If you run a family group chat, this is the argument for turning on whatever local scam warnings your platform offers.
The Bottom Line:
The target is the S$4 billion lost since 2019. No published savings figure yet. The design choice worth copying: privacy and protection were solved together instead of traded off.
6. SolarWinds put an autonomy dial on its network agents
Why it matters:
SolarWinds launched Avior, an agentic layer that sits across its on-premises and cloud products. The first agent handles network engineering: it correlates alerts into incidents, spots configuration drift, audits devices against compliance policy, manages configuration backups, tracks capacity and flags obsolete gear. The interesting part is not the tasks. It is the dial. You choose assistive mode (the agent recommends), human-in-the-loop mode (an admin approves each change), or fully autonomous within limits you define. Safeguards include risk-based guardrails, audit trails, automated rollback, data masking and least-privilege access. It works with third-party systems such as ServiceNow. Available now through a guided onboarding program, with a database agent planned. No independent time savings published.
How to use this:
Steal the dial. If you deploy agents anywhere, define three modes and what each one is allowed to touch. Start in assistive mode on the noisiest, most repetitive queue you have, and only promote the agent to autonomous on low-risk actions after it earns a clean record. "The agent has admin rights" is how you get an outage and a headline in the same afternoon.
The Bottom Line:
Aimed at the tactical toil that eats network engineers' weeks. SolarWinds publishes no hours saved, so ask for a before and after on your own alert volume.
7. AI finished a map of the sky that no telescope ever completed
Why it matters:
Ultraviolet light from space is absorbed by the ozone layer, so it can only be seen from orbit. Over 50 years, space telescopes photographed parts of the sky in UV, and the largest dataset, NASA's GALEX mission from 2003 to 2013, imaged about two thirds of it in some 38,000 separate observations. GALEX skipped the brightest regions, including much of the Milky Way's plane, to avoid damaging its detectors. The result was a map full of holes, and filling them properly takes weeks of pixel-level calibration that no one gets promoted for doing. Astrophysicist Brice Ménard used Claude Science to orchestrate a team of AI agents that gathered the public UV surveys, put them on a common scale, merged them and filled the gaps. About a third of the final map is predicted rather than measured, and the map labels every pixel as measured or predicted with an uncertainty estimate. You can look at it. It is the first complete UV map of the sky.
What you can do with this:
Open the map and look at the band you can never see with your eyes. Then use the pattern on your own backlog: find the project everyone agrees is worth doing and no one has weeks for, and see whether an agent team can close it. Keep the measured and predicted labels. The discipline that makes this credible is the same discipline you want in any AI output: mark what was observed and mark what was inferred.
The Bottom Line:
Zero money saved. It shows what agents are now good at: the unglamorous, months-long, nobody-owns-it job. Also a genuinely beautiful thing to bring up at dinner.
That is the week. The shift is not that models got smarter. It is that agents got cheap enough to run at scale, contained enough to trust with a dial, and present enough to sit on your face and in your network gear. Pick one workflow, put it in assistive mode, and keep the approval step until the output earns its way out.
Which of these would you actually use this week? Reply and tell me. I read every response.