{{first_name | Hey}}, welcome back.
Two AI products launched in the same week of September. TypeSafe AI shipped a model called Jev on the 15th. AgentCloak shipped a privacy tool on the 18th.
Founder Diogo Almeida had 1,500 followers before he posted. Doomers is the agency TypeSafe hired to run the launch. Its case study puts the video at 38,473,968 views and his account at 134,021 followers. Its launch page, read a day later, says 39M and 139,000. Take the smaller pair if you want the conservative version.
Vercel routes model traffic for a large share of AI apps, and it publishes its own adoption data. Within a day of Jev arriving on its AI Gateway, nearly 13% of paid teams had used it, double the share the GPT-5.6 family reached.
AgentCloak went out on a wire at 10am EDT on Friday the 18th. By September 24 its Chrome extension had 416 users. The product is free and the company behind it has raised $50 million.
What Jev does
Jev never writes a sentence. You send it context and a typed question. It returns the option it picked plus a separate confidence score, so ask whether a support ticket is urgent and you get back urgent at 0.88 confidence in about a tenth of a second.
The three fields a Choice call returns, from TypeSafe's docs here. The values are an example.
TypeSafe came out of stealth on September 15 with $40 million led by DCVC at a $200 million valuation, per Forbes. Almeida spent four years at OpenAI working on ChatGPT, and the launch post opens with "After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?"
Jev charges $0.042 per million input tokens and nothing at all for output, against $2.00 and $12.00 for GPT-5.6 Terra. Charging nothing for output is a positioning decision, and your CFO will want an answer ready the first time a competitor prices that way against you.
The demo that carried the video has Jev playing Doom from the game's state data, and the timing is the part people screenshotted.
The two timings TypeSafe put side by side on its own homepage, relayed by The Register. Independent analysis notes the runs measure different workloads.
What the crowd did with it
More than 150,000 people joined the waitlist inside 24 hours of the video going up.
Then developers around the world got access and started publishing their own use cases. Three of the stranger ones:
Sarvagya Kulshreshtha ran one candidate profile against 400 companies in 12 seconds, for a twentieth of a cent.
Nader Dabit built a spreadsheet column that works out what you meant and fills each row in about 100 milliseconds.
Read those first two again and the wording is nearly identical. Both open "JEV is INSANE" and both hit "All for just $" at the same point. Doomers says every paid post it places carries X's paid partnership label, and neither of these has one, so the likeliest read is that the format itself spread. A template that copies is its own kind of distribution.
A community tracker counted 714 posts from 604 authors between September 15 and 19, mostly in English, with 122 in Chinese and 74 in Japanese. That is well past anything a list of 80 to 100 accounts reaches on its own. On day five TypeSafe scrapped the waitlist and opened access to everyone.
Where Jev finished the week
Two days after the post, Doomers put Jev sixth among 2026's product launches by views on X. The list is the agency's own, so treat it as a claim rather than an audit. The names around it are the part worth looking at.
2026 product launches ranked by views on X, as published by Doomers. Top 10 of 13 rows.
The other launch that week
Every chief information security officer (CISO) is fielding a version of the same question about what happens to customer data once an employee pastes it into ChatGPT. AgentCloak Desktop swaps the private values for synthetic ones before the prompt is sent, then puts the real ones back into the answer on screen, and it costs nothing.
Detection runs on Rampart, a 14.7MB model the US National Design Studio published in June. By its own model card, Rampart catches 98.42% of private terms in Latin scripts and about 13.7% in non-Latin ones. Anyone with offices in Asia will want to raise that second figure in procurement.
AgentCloak is an InCountry product, and InCountry has raised $50 million across five rounds. Founder Peter Yared says he has built and sold six enterprise companies, and Crunchbase lists five, so the team has launched before. Morningstar, citybiz, FinancialContent and MarTech Series carried the release, mostly verbatim. No tier-one tech outlet covered it, and we couldn't find a single Hacker News thread about the company.
The Chrome extension was already live at version 1.1.2 on September 10, eight days before the press release announced it, so the wire was announcing something that had been shipping quietly for over a week. The release also names six chatbots, while the company's own FAQ says the desktop app supports ChatGPT Desktop only. WindowsForum headlined its write-up with that gap the same day.
3 plays you can run on your next launch
Make the demo work with the sound off: TypeSafe's video runs just under three minutes and shows the product doing the whole job end to end. Mute your own and see whether a stranger can still tell you what the product does. The OpenAI line got people to press play, and the demo then had three minutes to hold them. AgentCloak has a demo too, in Peter Yared's post and on its product page. The wire release that carried the launch to every trade outlet links the download page and the homepage and nothing else, so most people who met the product met it as text.
Give it two phases, and know which one you're failing: The launch agency seeded the post to between 80 and 100 engineers, founders, builders and creators in the first hour. That's phase one, and it buys a kickstart from people whose own audiences trust them. Phase two is the ordinary user who sees those posts and tries it themselves. If phase two doesn't produce the same reaction, no amount of seeding saves you.
Launching soon?
Phase one is the part that's hardest to assemble on your own. We put products in front of 100k+ subscribers, so if your date is close, tell us what you're shipping.
Let someone else publish your number: TypeSafe's homepage claims Jev is 193.6x faster and 444.6x cheaper than language models. Outside tests came back far lower. A UK events company running 50 real moderation decisions put it at roughly 5x faster than Mistral Small 4, and a TypeSafe employee measured 15.9% faster on one swapped workflow component. The 13% is the number still standing a week later, because Vercel measured it on infrastructure it controls and said how.
The risk to plan for
In the Doom demo, Jev reads the game's state as text. It never sees the screen, which makes the problem a lot easier than it looks. A competitive programmer called Psyho said so publicly 18 hours after launch and called the whole thing "expert marketing." He was right about the mechanics. That took 18 hours. Agree who replies to your own version, and in what words, before you ship.
Tool used
Doomers, a launch agency, ran the whole X strategy and picked the accounts to seed.
The three-minute film came from Skale, a video studio.
Vercel AI Gateway routes model calls for AI apps, and it produced Jev's first independent adoption number.
OpenRouter lists Jev, which is where a lot of developers first try a new model.
Cloudflare added Jev to its AI model catalogue inside three days, for teams running models at the edge.
LangChain published an integration guide two days after launch.
Business Wire, a paid press-release wire, carried AgentCloak to the trade outlets that picked it up. The founder's own post and the product page video were the only other places it ran.
Your launch lands tomorrow. What's carrying it?
Steal this prompt
Point it at a competitor's launch before you plan your own. It's written to make the model say what it doesn't know instead of guessing.
You are a competitive intelligence analyst. I am the CMO. Work from the material below unless I say otherwise.
COMPETITOR MATERIAL: [the launch post, the product page and two pieces of coverage, each with the date you captured it]. OUR CONTEXT: [what we are shipping, the category, the ship date, and the decisions still open on our plan with their deadlines].
Evidence rules. Quote the pasted material by default. Anything you go and find, tag RETRIEVED with its URL and capture date. With no quote for a claim, write NO EVIDENCE and move on. Do not estimate, and do not fill a gap with a plausible number.
Name every originating document or first-hand source the pasted coverage cites. A publisher is not an origin. If it all traces to one, say so. If my pasted set is too small to support a claim about coverage generally, say that and move to 2.
An amplifier is anyone other than the company who published about this launch within 14 days. List the ones you find, tag each RETRIEVED, and name the channels searched, including forum threads, newsletters, GitHub repos and blog posts, since that record outlives the social one. For each, say whether they are a customer, partner, investor, an analyst firm the company pays, or none of those, and how you established it. Quote any disclosure statement. Where there is none, write NO DISCLOSURE, and do not read that as independence. Say plainly that the list is not exhaustive.
Sort every claim into four lists: the company's, a third party that sells something whose value rises with this launch, a third party with nothing to sell, and relationship not established. Flag any company claim wearing a customer's clothes, such as a named customer statistic spoken by an executive rather than by the customer.
Give the sequence of public challenge and company response with dates, and say whether the company answered before or after. If it never answered, say so and give the date the challenge has stood since. If there was no public challenge, say so and move to 5.
List the cost centres we would incur to run an equivalent launch, naming the unit Finance already prices each one in, such as person-days by role, vendor invoices, seats or compute hours, and giving no quantity. Mark each visible in the evidence, implied by disclosed product scope, or inferred, where inferred is a category you name and never an amount. For anything not visible, name the centre, name the person or system holding the real number, and give the one question that would price it. Supply no figure yourself.
Answer in three parts.
One, under 350 words: the three findings that change a decision already open on our plan, one sentence each, naming the decision each changes. If I have not told you which decisions are open, write DECISION ASSUMED and state the assumption in six words. Then the decision you want out of the meeting, and what doing nothing costs us as a dated consequence rather than a number. Draw this part from the pasted material and from anything you tagged RETRIEVED, but write the tags nowhere in it. Close it with one line naming any finding that rests on an assumption rather than evidence.
Two, two things we could run in the next 30 days and one thing we should not copy. For each: the evidence above, the cost centre from answer 5 it draws on and who confirms the capacity exists, the owning team, the metric that should move, and what done looks like. Without an owning team, write OWNER REQUIRED and name the function that would assign it. Without a baseline, write BASELINE REQUIRED and name who owns that number.
Three, the five answers as an appendix, no length limit.
Get more such prompts in the Prompting Playbook (free for you)
Resource of the week
Obviously Awesome by April Dunford, on product positioning. TypeSafe came out of two years of stealth and had three minutes to explain a model that doesn't write text, to people who had never heard of it. Dunford's book is a process for exactly that problem. Over 100,000 copies sold.
Other Interesting Launches
PS. Did you see our AI Tools Directory? It’s a curated collection of 250+ useful AI tools for your profession, business domain, or specific use cases.
Stay curious, {{first_name | readers}}
