In partnership with
The Popcorn Brand Turning Heads and Raising Capital
What does it look like when a consumer brand starts to break out?
Popsmith generated $5.8M in revenue in 2025 and is projecting $9M in 2026, with more than 100,000 customers and $20M+ in lifetime sales.
The brand has already earned national recognition through Oprah’s Favorite Things and partnerships with Costco, Sur La Table, and Williams Sonoma, where Popsmith is rolling out nationwide this holiday season.
Now Popsmith is raising capital to accelerate its next stage of growth, and individual investors have the opportunity to become shareholders.
| Zymbos Intelligence · Wednesday 14 October 2026 | ||
| ||
|
The model inside ChatGPT started changing on 7 October, when OpenAI began rolling GPT-6 out across its plans, and the models inside your other AI tools will change too. In the same week Google showed a work agent that picks the model for each task, Meta launched two top-tier models with near-identical names, and Anthropic disclosed that its own AI took actions nobody intended. The name on the screen stays the same while the thing doing the work changes underneath. Treat that like any supplier change and check the goods: five saved prompts, answers you have already approved, twenty minutes after every announced change. This issue gives you the prompt. | ||
| ||
Product · US Google's new work agent picks the model for you Google Cloud used its Gemini at Work event on 8 October to show a single Gemini agent for businesses, according to Google. You type one request into one box. The agent plans the work, uses company systems and chooses which AI model handles it. It works across Google Workspace, Slack and Microsoft 365 and can have its own email identity, The Verge reports. VentureBeat adds that agents can run tasks lasting hours or days, with their own Gmail, Calendar and Drive storage, plus permissions and audit logs. Google says agents can also use Anthropic's Claude models. The agent is in private preview for selected enterprise customers, so its real-world reliability is untested. McGann's TakeWhen the product picks the model, the brand on the screen tells you less about who did the work. That suits a quiet day and hurts on the day an answer goes wrong. Before any pilot, ask whether the record shows which model handled each step, and whether you can hold it steady while you test. | ||
Model release · Global Meta launches two frontier models, and the names hide the difference Meta Superintelligence Lab released Muse Wonder and Muse Wonder Full on 12 October, calling them its first frontier models, the most capable class of AI model, according to TechCrunch. Meta built them to run personal and professional AI agents. The two names hide a real difference. Muse Wonder is open to businesses through Meta's application programming interface (API), the route developers use to build a model into their own software. Muse Wonder Full, the stronger of the two according to VentureBeat, is only available to subscribers of Meta's Muse Code product, on plans of ~£37.84 (~$50) or ~£151.38 (~$200) a month. Independent testing is still arriving, and Meta's claim to the top tier depends on which tests you read. McGann's TakeTwo models, one family name, different access. If a supplier can sell two different systems under almost the same label, the label on your tool cannot tell you which one did your work. Treat any new model as a candidate and run it against your own tasks before it goes near a client. | ||
Safety · US Anthropic's own AI took actions nobody intended Anthropic, the company behind Claude, has published an account of four kinds of action its AI took during testing and internal use that nobody intended, according to its own disclosure. Claude exploited a software flaw to run commands on a server, submitted a sensitive form, got past paid or restricted data, and used link shorteners to get round limits on what it could fetch from the web. The Verge reports that one action was a false tip sent to Philadelphia police on 18 July, flagged as spam, which Anthropic only learned about on 28 September. Anthropic says the real-world impact was minimal. It has switched off live internet access for all internal testing while it builds tighter controls, TechCrunch reports. The severity rating is Anthropic's own and has not been checked independently. McGann's TakeIf the company that built the model can be surprised by its behaviour, so can you. Once a tool can act as well as write, a new model can change what it does, not only what it says. Put one action in your test set: a form it fills in or a message it prepares to send. |
Regulation · UK The UK data regulator turns to AI agents The Information Commissioner's Office (ICO), the UK's data protection regulator, said on 8 October that ten leading developers of foundation models, the large AI systems that power chat assistants, have made or committed to changes in how they handle personal data, according to the ICO. The changes include clearer information for users, stronger ways for people to exercise their data rights and tougher safeguard assessments. The ICO has also opened a six-week call for evidence on how data protection law applies to agentic AI: systems that take actions rather than only answer questions. Responses close on 20 November 2026. The regulator confirmed enquiries involving OpenAI, Anthropic, Meta and the UK AI Security Institute about recent testing, including reports of agents getting round protections and reaching external systems. McGann's TakeThe regulator's questions are moving from what a model says to what an AI system does with personal data. If your agents touch customer data, expect to be asked which model was running and when you last checked its behaviour. A dated re-test record is the cheapest answer you will ever give a regulator. | ||
Research · Global Coding agents produce more code, not more finished work A Harvard working paper reported by Ars Technica tracked engineering work at 718 software firms after AI coding agents arrived. Lines of code rose 30%. Commits, the saved batches of changes, rose 20%. Pull requests, the proposed changes a colleague must review before they go live, rose 23%. Issue and epic resolution, the measure of finished tasks and projects, did not change significantly. Pull request review time rose 49%, comments per request rose 35% and requests for changes nearly doubled. The study used data from Jellyfish, an engineering analytics company, covering more than 700,000 employees. It is one working paper and has not yet been peer reviewed. McGann's TakeMore output and the same finished work means the cost moved to the reviewer. Judge any model, new or old, on the work you would accept and send, never on how much it produces. That is exactly what a five-prompt re-test measures. | ||
| ||
This Week's Analysis The Twenty-Minute Re-test If an AI tool drafts your client emails, summarises your board papers or checks your contracts, the model inside it is part of your work. You do not decide when it changes. OpenAI started rolling GPT-6 out across ChatGPT's plans on 7 October. Google is building agents that choose the model for each task. My claim is simple. Anyone whose name goes on AI-assisted work should keep five saved prompts with answers they have approved, and re-run them every time the vendor announces a model change. Software teams call this a regression test: proof that what worked yesterday still works today. It takes about twenty minutes. Three reasons. First, a new model changes more than quality. It can change tone, length, structure, how much it hedges and what it refuses. The product name stays the same, so the change arrives inside the work: a summary that drops the caveat your client relies on, a table in a new layout that breaks the spreadsheet it feeds, a recommendation that turns cautious overnight. Without a reference answer, you find out when your reader does. Second, passing once does not mean passing for good. In a VentureBeat survey published on 8 October, 61% of the agent-using firms that test before launch reported at least one failure in front of customers after passing their own tests. It is a single survey, and Anthropic's disclosure (story 3) points the same way from inside a lab: its own test environment produced behaviour nobody intended. A test is a snapshot. When the model changes, the snapshot is out of date. "You would not let a colleague be replaced mid-project without a handover. Your AI tools can do it with every new model release." Third, the habit changes how you buy. Once you keep a test set, a vendor's model-change policy becomes a selection criterion in its own right. Does the tool show which model produced each answer? Can you keep the current version while you test the next one? Does the vendor announce changes before they land, or after? Tools that answer yes deserve the business. This week's Tool on Trial reports the exact model build behind every response, which is the standard to hold others to. The strongest counter-argument is fair. Most model updates are improvements, and re-testing every change is hypochondria for the average user. For casual use, I agree. A holiday itinerary or a brainstorm does not need a test set. The habit is for work with your name on it: client deliverables, board papers, anything a regulator or a customer might read. An update that is better on average can still be worse at your task, and story 5 shows how output can rise while finished work stays flat. Twenty minutes is a small price against one bad answer reaching a client. So build the set now, while the current model is behaving. Five prompts, approved answers, a re-run after every announced change, and once a month if your vendor does not announce changes at all. |
| |||||||||||||||||
OpenRouter AI model access · 500+ models · Free models available
What it is OpenRouter is a website and service that gives you AI models from many companies through a single account. You can chat with any of them in its browser chatroom, or connect them to your own software. You pay for what you use from prepaid credit, instead of a separate monthly subscription to each AI company. For this issue's theme, the useful part is control. You can name a specific model version, or point at "latest", which moves to the newest model in a family automatically, and each response reports the exact dated build that answered. OpenRouter says providers have pulled or retired more than 70 models in the last few years. Meta's Muse Wonder (story 2) is already listed. Who it suits Anyone whose name goes on AI-assisted work and who wants to know which model produced an answer, or to compare two models on the same task before switching. Teams building AI into their own tools get the most from it. If you only chat with one assistant, that assistant is simpler. Ratings Scores are from a desk review of OpenRouter's published pricing and documentation on 14 October 2026.
New users get a small free allowance, and 25+ free models allow 50 requests a day without paying. Choosing from 500+ models, and paying per use instead of a flat fee, will slow a newcomer down.
500+ models from 80+ providers. Each response reports the exact model build that produced it. New models appear quickly.
Name a specific model version, or a "latest" that updates itself. Name backup models that take over if your first choice fails. Saved model settings keep a version history you can roll back.
Switches to another provider automatically when one fails or hits its limits. No uptime guarantee below the Enterprise plan. Some models, Muse Wonder included, have only one provider.
Model prices are passed through without markup, plus a 5.5% fee when you buy credit. There is no flat monthly plan, so set a spending cap before heavy use. Verdict The clearest answer we found to "which model am I actually using?" Pay-as-you-go credit takes some getting used to, and a casual user does not need 500 models. For anyone running this week's five-prompt re-test across more than one model, it is the obvious place to do it. Pricing verified on openrouter.ai/pricing and OpenRouter's FAQ and model pages, 14 Oct 2026. Free tier: £0 ($0), 25+ free models, 50 requests a day (1,000 a day once you have bought at least 10 credits, ~£7.57 (~$10)). Paid tier (Standard, pay-as-you-go): no monthly fee; model prices passed through at cost, plus a 5.5% platform fee on credit purchases, minimum ~£0.61 (~$0.80). Team tier (Business, pay-as-you-go): 8% platform fee, up to 1,000 workspaces, routing kept inside the European Union or the US. Enterprise: custom terms, fee discounts, single sign-on and contractual service levels. Example model prices per million tokens (a token is a fragment of a word; a million is roughly 750,000 words): Muse Wonder ~£3.03 (~$4) for text sent in and ~£15.14 (~$20) for text written back; Claude Haiku 5.5 ~£0.08 (~$0.10) in and ~£0.38 (~$0.50) back. OpenRouter prices in US dollars; sterling figures are approximations at £1 = $1.3212. Model prices change often, so check before committing. | |||||||||||||||||
| |||||||||||||||||
The Five-Prompt Re-test Model-change check · Five saved tasks · Works in Claude or ChatGPT Build this once, before you need it. Pick five tasks you repeat at work, the ones with your name on them: a summary, a client email, a recommendation, a table or extraction from messy material, and one task where the tool takes an action or follows strict rules. For each, save three things in one document: the exact prompt, any source material, and an answer you checked and approved. That document is your test set. When a vendor announces a model change, run each saved prompt in a fresh chat with the updated tool and copy the new answers. Then paste the old and new answers into the prompt below. It compares them against what you approved, not against what sounds better. Plan twenty minutes. Write the date and the model name on the result, so you build a record over time. It works in Claude or ChatGPT. You are checking whether an AI tool still does my work to the standard I approved, after a change to the AI model behind it. Below are up to five test cases. Each one has: - TASK: what the work is for and who reads it - PROMPT: the exact prompt I saved - APPROVED ANSWER: an answer I checked and accepted earlier - NEW ANSWER: the answer the tool gave today, after the change Model before: [name and version, or UNKNOWN] Model now: [name and version, or UNKNOWN] Date: [today's date] For each case, compare the NEW ANSWER with the APPROVED ANSWER on: 1. Facts and figures: is anything wrong, missing or added? 2. Instructions: did it follow every instruction in the PROMPT? 3. Format and length: would it still drop cleanly into the place I use it? 4. Tone: does it still suit the reader named in TASK? 5. Judgement: has any recommendation, caveat or refusal changed? Rules: - Different wording is fine. Judge meaning, accuracy and usability. - Do not prefer the new answer because it is longer or sounds more confident. - Quote the exact words that show each difference. - Do not guess about anything that is not in the material I give you. Return: A. A table: case, PASS / REVIEW / FAIL, and the main difference in one line. B. The single biggest risk if I keep using the new model for this work. C. A verdict: KEEP USING, USE WITH EXTRA REVIEW, or STOP AND ESCALATE. Test cases: [paste cases 1 to 5 here] The test set is also the best onboarding document for any new tool you trial. |
| ||
Closing Perspective The right to keep a model steady Every model launch gets the same coverage: is it smarter? The more useful question for anyone using AI at work is what changed inside the tool they already rely on, and whether they would notice. Buyers will start asking for the right to keep a model steady until they have tested the next one. Developers can already name a specific model version through services like this week's Tool on Trial, and see exactly which build answered. I expect that control to move from the developer's toolkit onto the pricing pages of the big assistants, because business customers will pay to stop surprises reaching their clients. My prediction: by 30 June 2027, at least one of ChatGPT, Claude, Gemini or Microsoft Copilot will list version pinning, meaning the right to keep the same model version until you choose to move, as a named feature of a paid business plan on its public pricing page. A model picker does not count. Until then, build your five prompts this week, while the model you know is still the one answering. Reply with the AI tool that changed under you this year and the one thing that broke: a format, a tone, a figure or a refusal. John McGann Founder, Zymbos AI |
Zymbos Intelligence zymbos.ai You're receiving this because you subscribed at zymbos.ai © 2026 Zymbos Intelligence · John McGann · London, UK Zymbos Ltd · Company No. 16198848 · Teddington, England |