OpenAI Build Week: $100K Results, Winners, and Everything That Happened

OpenAI Build Week ran from July 13 to July 21, 2026, and it turned into the largest hackathon OpenAI has run. Nearly 47,000 builders from 186 countries took part. They submitted more than 8,000 projects across eight days, with seven digital events and 60 in-person community events running alongside the work.

The eight winners were announced in late August, after Devpost moved the original August 12 target. Devpost’s schedule lists the announcement at August 25, 2:30 PM PT.

Those eight teams shared $100,000. First place in each track took $15,000. Second place took $10,000. First-place teams also received DevDay passes, time with the Codex team, and a year of ChatGPT Pro.

Build Week in numbers

• Nearly 47,000 builders took part
• 186 countries represented
• More than 8,000 projects submitted
• Eight winners across four tracks
• $100,000 in total cash prizes

The brief was narrow on purpose. Build something real using Codex and GPT-5.6 together, in one of four tracks: Apps for Your Life, Work and Productivity, Developer Tools, or Education. Judges scored technical implementation, design and user experience, potential impact, and quality of the idea.

Apps for Your Life

First place: Second Voice, built by Ravitez Dondeti

Second Voice is a communication aid for people with dysarthria or limited motor control. It takes a person’s partial speech, combines it with their phrasebook and the current context, and suggests a small set of likely sentences. The speaker picks or edits one before the app says it out loud, so the person stays in control of every word.

Dondeti built it alone during the week. His starting point was personal: a relative lived with a speech disorder, and he remembered communicating through gestures. His hardest problem turned out not to be reconstructing the sentence. It was the interaction design. How many options to show, how fast they appear, and how much effort confirmation costs someone with limited motor control.

Second place: AirBridge for Windows, built by Adam Tarantino

AirBridge streams audio from a Windows PC to HomePods, Apple TVs, and Macs without virtual audio cables. It supports multiple speakers, calibrates delay per room, and includes a browser extension that fixes lip sync by delaying the picture instead of the audio. A GPT-5.6 assistant can control it by voice, while a local policy layer checks what the assistant is allowed to do against the real hardware.

Tarantino had tried the idea with an earlier OpenAI model in 2024 and did not get a working product. This time he did, and it was his first hackathon.

Work and Productivity

First place: veTriage, built by Erin Downes

veTriage is a phone triage app for veterinary teams. It helps receptionists collect the right history, spot urgent warning signs, and route cases for clinical review without asking non-clinical staff to make medical calls. The design principle is blunt: a full schedule does not make a patient less urgent.

Downes is a 61-year-old veterinarian and practice owner with no coding background. Her clinic went from three veterinarians to one, while worried pet owners kept calling. She turned decades of clinical judgment into a working workflow, and her team is already piloting it.

Second place: Pulse, built by Mohamed Mostafa Mohamed Labib Abu Taleb

Pulse is a research prototype for cardiac arrest response. It listens to the room during a resuscitation, tracks clinical state including code-switched Egyptian Arabic, and keeps a shared view of what has happened and what may be due next. Deterministic, auditable code tracks the workflow. GPT-5.6 interprets messy speech, and the team confirms anything unclear.

Abu Taleb is a Cairo-based cardiologist who built it between hospital shifts, with no team, no lab, and no funding.

Developer Tools

First place: Echo Canvas, built by Kevin Yang

Echo Canvas is a browser workbench for spatial audio. A designer can sketch a room, place sound sources and listeners, open a doorway, change a wall material, and hear the difference immediately. It is the acoustic equivalent of a visual wireframe, aimed at the stage before a scene is built in a game engine.

GPT-5.6 authors and explains scenes inside constrained schemas. Geometry and acoustic calculations stay deterministic, and the audio renders locally in the browser.

Second place: Sentinel, built by Malik Bashaar Javaid

Sentinel scans MCP servers for security problems before they ship. MCP servers can hand an agent access to files, APIs, databases, and shell commands, and many are shared as quick-start templates with unsanitized shell calls and embedded credentials.

Sentinel combines static analysis, a tightly constrained GPT-5.6 review of findings in their source context, and Docker-isolated probes that test real behavior. The model can corroborate or challenge a finding. It cannot invent probes or cite code that does not exist. Findings map to the OWASP Agentic Top 10 and can feed GitHub code scanning.

Education

First place: Mechanica, built by Weiying Zhu, Yukun Li, and Shan Wei

Mechanica is an interactive digital museum of ancient Chinese machines, from an astronomical clock tower to a programmable loom built centuries before modern computing. Visitors can operate the mechanisms, take them apart, and trace each dimension back to a classical text, a measured artifact, or a labeled scholarly inference.

Where historians disagree, the museum shows competing reconstructions instead of one confident answer. The three builders had decades of software experience and almost none in 3D modeling, physics simulation, or animation. Codex carried them across that gap. GPT-5.6 powers a docent that cites the museum’s evidence or declines to answer when the evidence is missing.

Second place: Dấu, built by Robert Huynh

Dấu is a pronunciation coach for Vietnamese tones. The same syllable can carry six meanings depending on pitch and vocal quality. Learners record a word, see their pitch curve next to a validated native-speaker reference, learn which meaning their attempt may have produced, and get a physical correction.

Huynh had left his only previous hackathon after half a day, worried he was not technical enough. He built Dấu solo in Hanoi. His most important decision was what the AI should not do. Deterministic signal processing judges the tone. GPT-5.6 coaches the learner. When the evidence is unclear, the app asks the learner to try again rather than guess.

What the eight winners have in common

Every winning project kept the model on a short leash. Sentinel’s builder described his work as building schema checks and rules against citing code that does not exist, not prompt tuning. Echo Canvas keeps geometry out of the model. Dấu keeps tone scoring in deterministic code. Pulse keeps clinical state in auditable logic.

The other pattern is domain expertise. A veterinarian, a cardiologist, a museum team, and a language learner all built tools for problems they had lived with. None of them started as professional developers of that product.

Prize structure and eligibility

The $100,000 was split across four tracks, with two cash awards in each. Beyond cash, first-place teams received DevDay passes, meetings with the Codex team, promotion from OpenAI Developers, and a year of ChatGPT Pro.

The event was open globally to anyone at the age of majority in a country where the OpenAI API is supported. Solo and team entries were both allowed. Residents of sanctioned regions including Russia, Iran, North Korea, Cuba, and Syria were excluded.

What a valid entry required

Every submission needed a working project built with both Codex and GPT-5.6, a written description, a narrated demo video under three minutes, a public repository with a README and setup instructions, and the Codex Session ID showing where core functionality was built. Projects had to stay free to use during judging.

Registered entrants could request $100 in Codex credits. The request deadline was July 17 at 12:00 PM PT, and the credits expired July 31, 2026.

Where to see the projects

OpenAI published all eight winners with demos at openai.com/build-week. The full submission gallery, including finalists in each track, is on the Devpost project gallery at openai.devpost.com.

Frequently asked questions

Who won OpenAI Build Week 2026?

Eight projects won. Second Voice and AirBridge for Windows took Apps for Your Life. veTriage and Pulse took Work and Productivity. Echo Canvas and Sentinel took Developer Tools. Mechanica and Dấu took Education.

How many people entered OpenAI Build Week?

Nearly 47,000 builders from 186 countries took part. Devpost lists 46,685 participants. More than 8,000 projects were submitted.

How much prize money was awarded?

$100,000 in total. Each of the four tracks paid $15,000 for first place and $10,000 for second place, plus DevDay passes, Codex team access, and a year of ChatGPT Pro for first-place teams.

When were the winners announced?

In late August 2026. The original rules targeted August 12, and Devpost later listed the announcement at August 25, 2:30 PM PT.

Could beginners enter?

Yes. Erin Downes won first place in Work and Productivity with no coding background at all, and two of the winners were entering their first hackathon.

What did judges score?

Technical implementation, design and user experience, potential impact, and quality of the idea. Technological implementation carried the most weight.

How did the $100 in Codex credits work?

Registered entrants could request $100 in Codex credits through the Build Week form. The request deadline was July 17 at 12:00 PM PT, and the credits expired July 31, 2026. These credits were separate from the $100,000 prize pool.

Where can I see the Build Week winners and projects?

OpenAI published the eight winners at openai.com/build-week. The full submission gallery, including finalists in each track, is at openai.devpost.com.

When was the Build Week submission deadline?

The submission deadline was July 21, 2026. Winners were announced in late August, after Devpost moved the original August 12 target.