Midjourney bought a consumer product team and a daily-habit app, not a horoscope business. It has never shipped a standalone app of its own, and Co-Star has run one at scale since 2017. The deal closed in spring 2026 and was announced on 23 July; terms were not disclosed.
What changed hands
Co-Star founder Banu Guler became Midjourney’s chief design officer and is building a cross-functional design, product and frontend team covering all of Midjourney’s efforts. Roughly two dozen Co-Star employees came with her. Midjourney founder David Holz said the app stays entirely under Guler’s control and direction, with expanded resources behind it.
Co-Star has about 4.3 million monthly active users according to Sensor Tower data cited in reporting on the deal. It is free with paid chart readings on top. Its planetary position data comes from NASA’s Jet Propulsion Laboratory, and its daily text is written by a mix of human astrologers and language models.
The acquisition sits inside a wider expansion Midjourney announced a month earlier: eight projects split evenly between hardware and software, including a full-body ultrasound scanner intended for spa locations, plus a first standalone image-generation app. Holz has said roughly half a dozen more announcements are coming over the following six months.
The two products share a pipeline
The pairing reads as a joke until you diagram both systems. A birth chart generator and a diffusion model share the same three-stage architecture: a deterministic input layer that nobody disputes, a generative layer that turns that input into language or pixels, and a final stage where the user supplies the meaning.
| Stage | Co-Star | Midjourney |
|---|---|---|
| Fixed input | Birth time and place resolved against JPL ephemeris data | Prompt tokens, seed, parameters, model weights |
| Reproducible? | Yes — same birth data, same chart, every time | Yes — same seed and settings, same image |
| Generative layer | Astrologers and language models turn geometry into sentences | Denoising turns a latent into an image |
| Output character | Specific enough to feel personal, loose enough to fit | Recognisably what you asked for, never quite what you pictured |
| Where meaning lands | In the user | In the user |
The deterministic layer is the trust anchor
Astrology’s calculation layer is real astronomy. Planetary positions for a given moment are computable to arcsecond precision, and Co-Star computes them from the same ephemeris data spacecraft navigation uses. Everything contested happens after that, in interpretation.
Image models have the same split. Seeds, weights, and sampler settings are deterministic; run them again, and you get the same file. What the model does with a phrase like “melancholy” is not something anyone can derive from first principles. Both products present a hard, checkable substrate and then add something soft on top of it — and in both cases, the hard part does reputational work for the soft part.
The generative layer is a prompt problem
A birth chart is a structured object: positions, aspects, houses. Turning it into a paragraph a 24-year-old will screenshot is the same task as turning a prompt into an image someone will post — a constrained input, an enormous output space, and a quality bar that is aesthetic rather than factual.
Co-Star has been doing this every day for years with editorial oversight in the loop. That is unusual. Most consumer AI products ship raw model output and let the user judge it. Co-Star’s blunt push notifications are a house voice maintained across millions of personalised variants, which is closer to a style guide enforced at inference time than to a chatbot.
The user finishes the output
This is where the systems converge hardest. Bertram Forer demonstrated in 1949 that people rate a generic personality description as highly accurate when told it was written for them personally. His students averaged 4.26 out of 5 on a paragraph every one of them had received identically.
Generated images work on a related reflex. An AI portrait of “you as a medieval cartographer” is not a likeness; it is a plausible thing you agree to see yourself in. Both products manufacture recognition. That is the shared design competency Midjourney is buying — a team that has spent years tuning how much specificity an output needs before a stranger claims it as theirs.
Both companies sell the same thing: a machine-produced artefact personal enough that the user finishes it themselves.
Where the analogy stops
The architectures rhyme. The epistemics do not.
An image model’s output traces back to training data through a mechanism that can be inspected, measured and improved. Astrology’s interpretive layer has no proposed causal link between planetary geometry and personality that has survived testing. Noticing that two systems share a pipeline shape says nothing about whether either one’s claims hold.
The audiences differ too, and Pew’s May 2025 survey is the useful corrective here. About 27% of US adults say they believe in astrology, roughly unchanged since 2017. About 30% consult astrology, tarot or a fortune teller at least once a year — but around 20% do it mainly for fun versus about 10% for helpful insight, and only about 1% say they rely on it heavily for major decisions.
That is the market Midjourney actually acquired: mostly entertainment, with a daily open rate. Belief is not the retention mechanism. Ritual is.
The transferable mechanic
Co-Star’s loop is: fixed personal input collected once, freshly generated output on a schedule, output shaped to be compared with someone else’s. Nothing in that structure is astrological. It applies to any product with a per-user parameter set and a generative backend — which describes most of what Midjourney is now building.
What Midjourney lacked
Midjourney’s distribution has been a Discord server and a web app. That filters its user base to people willing to learn a chat-command interface inside a platform built for gaming communities. It has no onboarding, no notification surface, no App Store presence and no sessions that start without deliberate intent.
Co-Star supplies all four, plus a team that has shipped against App Store review, mobile retention curves and freemium conversion. Buying that is faster than hiring it, and it arrives already knowing how to work together.
The stated rationale is less mechanical. Holz said he met Guler over five years ago while thinking about how people reflect and make decisions, and that they spoke frequently for years afterward. Guler framed the deal as a shared position that computers can reveal our humanity to us rather than replace us. Read that as positioning if you like; it still explains why a design lead from an astrology app is a coherent hire for an image company. Both products hand the user a mirror.
The first joint product
Bloomberg reported an astrology-focused image generator as the first product from the combination — birth chart data as the parameter set, generated imagery as the output. It is the smallest possible test of the thesis: take a per-user input Co-Star already holds, run it through Midjourney’s models, and see whether the result gets shared.
The risk is on the other side of the trade. Part of Co-Star’s audience is actively hostile to generative image tools, and some users said publicly they were leaving over the acquisition. A birth-chart image generator points that tension directly at the most engaged part of the user base.
Common questions
No. Midjourney says the app continues to operate under Guler’s control, with additional resources. Neither company has announced changes to existing Co-Star features or pricing.
Undisclosed. Midjourney is self-funded and does not publish financials, so no figure has surfaced from either side.
The opposite, in terms of surface area. The standalone image app is the flagship item on the roadmap; the hire exists to build it and the products around it.
An agency delivers designs; Co-Star delivers an operating product with millions of daily users, a distinctive editorial voice maintained at generative scale, and a team that has already solved retention on a per-user generated feed.
Yes, and for years. Its daily text has been produced by natural-language systems working alongside human astrologers, on top of astronomical position data. The acquisition changes ownership, not the underlying method.
What to watch
Two signals will settle whether this was a product acquisition or a talent acquisition with a mascot attached: whether the standalone Midjourney app ships with Co-Star’s onboarding and notification patterns visibly inside it, and whether Co-Star’s monthly active users hold through the next two quarters of Sensor Tower data.