Anthropic's Project Glasswing Shows the Real Launch Playbook: Recruit 50 Partners First, Go Public With Their Results

The play

Anthropic didn't launch Project Glasswing by posting a blog and hoping people would try it.

They built a cohort first.

Before any public announcement, they recruited approximately 50 enterprise partners — companies that build and maintain critical infrastructure software. They gave those partners access to Claude Mythos Preview. The partners ran it on real code. Vulnerabilities started surfacing.

After one month, those partners had collectively found more than 10,000 high- or critical-severity vulnerabilities. Cloudflare found 2,000 bugs, 400 of which are high- or critical-severity, with a false positive rate that Cloudflare's team considers better than human testers. Mozilla found and fixed 271 vulnerabilities in Firefox 150 — over ten times more than they found in Firefox 148 with a previous model.

Then Anthropic published the update.

That's not a launch. That's a results report. And that distinction is the entire growth play.

Why cold launches underperform

When you announce a new product cold — no evidence, no partners, no numbers — you're asking the market to do two things at once:

1. Believe your claim

2. Try your product

That's a big ask. Most people solve the first step by ignoring the second step. They're interested, but not enough to run the experiment themselves.

The partner-first launch solves this. You remove the belief gap before you ask for the trial. The announcement says: people like you already tried this, here's what they found, here's what one of them is willing to say publicly.

That's a fundamentally different ask. It's no longer "trust us that this works." It's "look at what already happened."

How to structure the partner cohort

Who to recruit

Your lighthouse partners should be your ideal customers — the ones with the clearest version of the problem you solve, enough resources to run the product seriously, and enough credibility that their endorsement carries weight with the next 1,000 customers you want to reach.

You don't need 50. Even 5–10 well-chosen partners generating real results is enough for a strong launch. Anthropic's scale was unusual. Most early-stage products don't need that many.

What to offer

Free access is the obvious move. But the more important offer is involvement: partners who feel like they're shaping a product are more invested in its success than partners who are just using a free trial.

Give them a direct line to the team. Ask for feedback. Make them feel like co-creators of the early version, not test subjects.

What to ask for in return

Three things:

• Use the product on real work (not a demo environment)
• Track and share the results with you
• Be willing to be named publicly when you announce

The third one is the hard ask. Some partners will say no. That's fine. You need two or three who will say yes — that's enough for the announcement to have external confirmation.

The timeline

Give partners enough time to generate meaningful results. One month is usually the minimum. Ninety days is often better. The goal is outcomes that represent real usage, not first impressions.

How to write the announcement

The launch post should read like a results summary, not a product pitch.

Lead with the number. Not "we built a tool that finds security vulnerabilities" but "our partners found more than 10,000 high- or critical-severity vulnerabilities in the first month."

Then explain what the tool is and how it works. Then quote or link to the partners who are willing to go on record.

The structure is: outcome → explanation → external confirmation.

That sequence builds credibility in the right order. The outcome earns the read. The explanation provides the context. The external confirmation closes the loop.

Who this is for

This pattern works best for products with clear, measurable outcomes — security (vulnerabilities found), sales (pipeline generated), engineering (bugs fixed, time saved), finance (cost reduced).

If your product's value is harder to quantify, partner-first launches still work, but the "results report" framing is less clean. You'll rely more on qualitative testimonials than numbers, which means the partner selection matters even more.

The key constraint: you need a product good enough that real usage generates real results. This is not a strategy for products that aren't ready. Glasswing worked because the model actually found the bugs. If you recruit partners before your product is ready to produce results, you waste the launch window and burn the relationship.

Build first. Partner second. Announce third.

Source: https://www.anthropic.com/research/glasswing-initial-update