OpenAI’s Camera Buy Is Built for Tiny Sensors

OpenAI has bought Glass Imaging for more than $300 million, according to people familiar with the matter. The Los Altos startup sells neural software that makes tiny camera sensors yield cleaner pictures.

Neither company has confirmed the purchase, which was described on September 14, 2026. OpenAI also did not answer questions about how it will use the tech. The likely job is giving a small, cheap module enough sight that a pocket gadget can watch a room.

OpenAI Paid More Than $300 Million for Better Eyes

Glass Imaging was founded in 2019 by Ziv Attar, the chief executive, and Tom Bishop, the chief technology officer. Both previously worked on Apple’s camera stack, including Portrait Mode. The firm had been valued at about $100 million in a 2025 funding round, so the OpenAI price is more than three times that level.

Insight Partners led a $20 million Series A on May 12, 2025, with GV, Future Ventures, and Abstract Ventures joining. GV had led a $9.3 million extended seed in 2024, with Future Ventures, Abstract, and LDV Capital in that round. Those two named raises are the figures Glass put on the record. The company has not posted a confirmed comment on the OpenAI sale.

THE GLASS IMAGING DEAL

The step-up is large for a licensing shop that still sells to phone makers. It is small beside OpenAI’s 2025 hardware move, when it took in Jony Ive’s device startup io in a deal valued at $6.5 billion. That earlier buy brought the design team. This one buys the eyes.

The Software That Undoes a Tiny Lens

GlassAI is a neural image signal processor. It sits on the RAW file, the noisy mosaic that leaves the sensor, and it is trained on one camera module at a time. The company says its networks reverse lens and sensor imperfections instead of painting over them after the shot.

A normal phone ISP runs demosaic, denoise, and sharpen as a chain. Shivansh Rao, on Glass Imaging’s machine learning team, said a traditional ISP runs those as discrete steps, and each step discards information later steps can never recover. Glass trains the whole path end to end on RAW so the losses do not stack.

GLASSAI PROCESSING BLOCKS

• Neural ISP: Takes sensor RAW and undoes lens flaws, sensor effects, and noise on the device.
• Neural Zoom: Fuses a burst of RAW frames to pull zoom from whatever lens is already there.
• Neural Night: Denoises dark scenes while trying to keep a natural look, rather than a smeared night-mode glaze.

The pitch is restoration, not invention. Rao said GlassAI is trained to recover genuine detail the optics actually captured but a standard pipeline would lose, which is how the firm answers the charge that AI cameras fake texture. Bishop has described the same idea as correcting optical aberrations and sensor issues while cutting noise, without the made-up detail that generative upscalers add.

That matters most as pixels shrink. Glass’s Deep Optics team wrote that smartphone telephoto sensors have gone from about 1.0 μm in the iPhone 7 Plus era toward 0.5 μm in today’s 200-megapixel periscope modules. Wider apertures then add geometric blur. The lab calls this the telephoto physics wall. In its own simulation, neural restoration improved resolution (MTF50) by over 50% as pixels shrank from 0.75 to 0.35 μm, and the neural ISP reached 745 cycles/mm in the vertical direction at 0.35 μm. A conventional ISP largely stalled, because leftover blur stayed in the file.

Separately, Glass still sells a co-designed optics kit: a sensor twice the size of the largest phone sensors, folded into a thin module, with apertures up to 10x larger and up to 10x more light. That hardware family is not the Honor software already in stores. Both come with the company OpenAI is buying.

Honor Already Ships Zoom Without a Telephoto Lens

The proof is already on a shelf. In June 2026, GlassAI shipped on Honor’s 600 series for zoom. Honor does not put a dedicated telephoto in that phone. It crops the 200-megapixel main camera instead.

“We’re honored to partner with Honor and to have our Neural ISP technology featured in the 600 series,” Attar said at the time. Rao walked through why that crop is a brutal test. The 200-megapixel sensor uses a 16-in-1 Hex Bayer pattern, so native pixels are about 0.56 μm and are normally binned into 2.24 μm super-pixels. Zoom shoots at those native sub-micron pixels, which is where a standard ISP runs out of real detail.

Rao said sub-micron pixels encode high-frequency data in complex ways that traditional ISPs fail to decode. That information is not lost, just hard to recover. Glass models the point spread function, the sensor, and the noise of that exact module, then tries to undo the blur at the source.

The same constraint shows up in foldables and ultra-thin phones, where a third camera costs thickness a designer will not spend. Rao said the team is also aiming at wearables, drones, cars, and medical cameras, and that people should expect to see Glass Imaging in more places over time. A pocket companion with no camera bump is the same physics problem with a harsher size budget.

Why a Pocket Companion Cannot Hide a Camera Bump

Sam Altman told staff the first Ive device would not be a phone, glasses, or something worn on the body. He said the product will be capable of being fully aware of a user’s surroundings and life, will be unobtrusive, and will rest in a pocket or on a desk. He called a prototype the coolest piece of technology the world will have ever seen, and he told employees the io deal could add $1 trillion in value, with a goal of shipping 100 million units faster than any company has shipped 100 million of something new.

The product will be capable of being fully aware of a user’s surroundings and life, will be unobtrusive, able to rest in one’s pocket or on one’s desk.

Sam Altman, CEO, OpenAI staff call

Awareness, in that pitch, is a camera and microphones. A desk puck or pocket slab cannot hide a flagship phone’s camera hump. It also cannot spare the volume for a long telephoto. If the gadget is to read a room, a whiteboard, a face across a table, or a dim kitchen at night, the module has to be small and the software has to wring the file dry.

On May 12, 2025, Glass told manufacturers its software could boost camera performance 10x while staying true to life, with no hallucinations. That claim is the company’s, not a lab’s independent score. What the Honor 600 work does show is a narrower, concrete trick: drop the extra lens, crop a dense sensor, and still keep texture in the zoom. For an unobtrusive gadget, dropping the extra lens is the whole industrial-design problem.

OpenAI is also rumored to be exploring phones, earbuds, and other companions. A phone would use this stack too. The harder case remains the screenless object that has to see without looking like a camera.

Attar and Bishop Built Portrait Mode, Then Left

Attar is an Israeli optical engineer who studied at the Technion, then worked at Rafael. He founded OpEra Optics in 2007 and co-founded LinX in 2011 with Andrey Tovchigrechko. LinX built compact multi-sensor cameras and an early dual-camera portrait effect. Apple bought LinX in 2015. Attar joined Apple’s camera algorithm team and helped turn that work into Portrait Mode, which shipped on the iPhone 7 Plus in 2016.

He met Bishop at Apple. They left in 2019 to start Glass. The second company was a bet that phones finally had enough on-device compute to run full-resolution neural processing, and that the next gains would come from smaller, cheaper, thinner modules rather than another year of larger sensors.

FROM LINX TO OPENAI

1. 2011: Attar co-founds LinX, a compact multi-sensor camera firm.
2. 2015: Apple buys LinX and Attar joins the camera algorithm team.
3. 2016: Portrait Mode ships on the iPhone 7 Plus.
4. 2019: Attar and Bishop leave Apple and found Glass Imaging in Los Altos.
5. May 12, 2025: Insight Partners leads the $20 million Series A.
6. June 2026: GlassAI ships on the Honor 600 series for zoom.
7. September 14, 2026: OpenAI’s purchase, valued at more than $300 million, is described by people familiar with the matter.

“It’s not about taking existing photos and enhancing them, that’s something many people are already doing,” Attar said. “Our goal is to create hardware changes that are only possible thanks to the AI we are developing.” He argued the software could let makers build smaller and cheaper cameras that still look sharp, and thinner modules that steal less space inside a crowded device. That sentence is the acquisition in miniature.

A Screenless Gadget That Still Has to Look

In a promotional video last year, Ive said everything he had learned over the last 30 years had led him to this place and this moment, and that a new generation of technology could make people their better selves. The public pitch has been less phone, less screen, less of the harm that comes with staring at glass all day. On July 9, 2025, OpenAI said the io Products team has officially merged with OpenAI, while Ive’s LoveFrom studio stayed independent and took on design duties across the company.

A device that weans people off screens still has to collect the world somehow. Altman’s own words put cameras in the product: fully aware of surroundings and life. Later accounts of prototypes have described a portable, screenless speaker with a camera, a battery, and moving parts meant to feel alive. OpenAI has not confirmed those details, and it has not set a public ship date.

Ive’s objection to wearables, including a jab at pin-style gadgets, pushes the camera into a body that sits on a desk or in a pocket. That is a worse optical problem than a phone, which can at least dedicate a corner to a bump. Glass’s licensing work on Honor is the closest thing on the market to that compromise: one dense sensor, no extra telephoto, software doing the zoom.

Anything the Sensor Misses Gets Guessed Later

Peyman Milanfar, a distinguished scientist at Google who works on computational imaging, wrote that it is good to see a large investment in the field, and that the camera is the future of intelligence. The line is easy to nod at. The operational version is blunter. An always-on assistant only knows what hit the sensor. If the file is noisy, blurred, or cropped to mush, the model fills gaps. Filling gaps is slower, costlier, and less trustworthy than reading a clean scene.

That is the second effect of this deal, and it cuts two ways. Better upstream vision can make the whole stack cheaper, because the model spends less effort guessing. It also makes a quiet desk object more intimate. Night mode that actually sees, and zoom that reads a far whiteboard, are photography features on a phone. On a companion that is “fully aware,” they are surveillance features with a friendlier industrial design.

WHAT WE KNOW

• The price: People familiar with the matter put the Glass Imaging purchase at more than $300 million.
• The product: GlassAI is a neural ISP trained per camera, already shipping for zoom on the Honor 600.
• The people: Attar and Bishop left Apple after Portrait Mode and spent seven years on small-sensor software.

WHAT IS UNCONFIRMED

• The confirmation: OpenAI and Glass Imaging have not announced the sale themselves.
• The use: OpenAI has not said whether Glass goes into the Ive companion, a rumored phone, or both.
• The ship date: No public date exists for the first Ive device, and the form is still under wraps.

OpenAI already paid $6.5 billion for a hardware team that wants an unobtrusive third device on the desk. Paying more than $300 million more for a neural ISP is how that object gets a camera it can hide. The model on the other side of the lens will only be as sure as the photons that made it through.