How Work Gets Done
Shipping code at OpenAI means pushing to half a billion users while the org chart rewrites itself quarterly and the mission statement has changed multiple times in recent years. The company runs on a wartime clock: in a single year it ballooned from roughly 1,000 to 3,000 people, former engineer Calvin French-Owen said, and now numbers around 7,800. ChatGPT crossed 500 million active users in March 2025. That velocity snaps the usual infrastructure: how teams talk, who reports to whom, how they hire, the codebase itself. French-Owen called the central repository, known internally as the back-end monolith, "a bit of a dumping ground" where code quality swings wildly between seasoned ex-Google engineers and newly minted PhDs writing Python.
OpenAI's fast-paced, research-first environment shapes how teams ship models, reinforcing a culture of high autonomy and relentless iteration. Candidates who thrive pair deep technical expertise with comfort amid ambiguous goals and rapid change.
The org chart stays flat. Everyone carries the title Member of Technical Staff, a structure Built In documented that de-emphasizes hierarchy for mission alignment. Small teams form around a Directly Responsible Individual who owns a project end to end: design, build, launch, iterate. The DRI model removes ambiguity about who decides and who delivers. French-Owen's own team of 17 (engineers, researchers, designers, and product) built and launched Codex in seven weeks, "with almost no sleep."
"Everything breaks when you scale that quickly: how to communicate as a company, the reporting structures, how to ship product, how to manage and organize people, the hiring processes, etc."
That startup texture persists. Slack runs the company. Decisions happen in channels and DMs, not tickets or design docs. Teams duplicate work because coordinating costs more than colliding. Autonomy is real — people act on ideas with "little to no red tape" — but it creates silos. Strict confidentiality, meant to stop leaks, deepens the isolation. Internal surveys show employees feel disconnected from the broader vision despite collaborative ideals.
The research-to-product pipeline drives everything. GPT-4 became ChatGPT; DALL-E became the API; Sora moved from lab to limited release. Forward-deployed engineers embed inside enterprise customers to wire agents into live workflows, a model Amazon now copies. But the tension between shipping fast and deploying responsibly never lifts. Internally, safety focus has shifted from theoretical existential risk to practical guardrails: hate speech, prompt injection, political bias manipulation, self-harm, bio-weapon crafting.
Glassdoor shows the strain. Work-life balance rates 3.6 out of 5. Reviews cite "crunch culture," "weekend work during launches," and "60- to 70-hour weeks as the norm, not the exception." The overall rating holds at 4.5, driven by outsized ownership, compensation, and the thrill of shipping to millions overnight. The tradeoff is explicit: speed and autonomy over process and sustainability.
Top engineering managers know the monolith creaks. They're working on improvements. But OpenAI still runs like a scrappy startup in ways that help and hurt. The question for any candidate: can you build inside the breakage?
Principles Forged in Tension
OpenAI publishes five operating principles dated April 2026: democratization, empowerment, universal prosperity, resilience, and adaptability. The list reads like a mission statement, but the company's history shows each principle was forged in response to a concrete tension: between openness and safety, between research freedom and commercial pressure, between a lab culture and a consumer product reaching half a billion users.
The democratization principle declares that OpenAI will "resist the potential of this technology to consolidate power in the hands of the few" and that key decisions should be made "via democratic processes and with egalitarian principles, and not just made by AI labs." That language echoes the 2015 founding dinner at the Rosewood Hotel, where Greg Brockman said the best thing he could imagine was "moving humanity closer to building real AI in a safe way" and Ilya Sutskever left Google "to a very large extent, because of its mission." The early pitch was explicit: give away what might become the 21st century's most transformative technology. But a 2016 Wired profile caught Brockman acknowledging OpenAI "won't give special treatment to its sister companies" and admitting "OpenAI's idealistic vision has its limits. The company may not open source everything it produces." In practice, democratization has meant API access, not weight releases, a distinction Reid Hoffman later called "open because it has open access to APIs but it's not open because it's open source."
Empowerment and universal prosperity sit side by side in the 2026 principles. Empowerment promises broad latitude for users while "erring on the side of caution in the face of uncertainty, and relaxing constraints with more evidence." Universal prosperity calls for driving infrastructure costs down and building "huge amounts of AI infrastructure." Sam Altman put it bluntly in 2026: "Our mission is to make powerful intelligence extremely cheap, abundant, and decentralized to empower humanity. I worry deeply about AI authoritarianism, where a small group of people or a single company controls a machine god to dominate the world." The prosperity principle also acknowledges that governments may need new economic models so "everyone can participate in the value creation in front of us."
Resilience makes the principles operational. The 2026 document says OpenAI will "work with other companies, ecosystems, governments, and society to solve" new risks, citing pathogen-agnostic countermeasures and using models to secure open-source software. It frames iterative deployment as a societal necessity: "society needs to contend with each successive level of AI capability, understand it, integrate it, and figure out the best path forward together. This cannot be done in a vacuum; society and technology co-evolve, and that requires time." Brockman confirmed this in 2026: "One thing people don't understand as much about OpenAI is our extreme focus on AI safety and how seriously we've taken that over the years." He pointed to chain-of-thought monitoring as a "clear line in the sand", refusing to optimize reasoning traces for appearance because it destroys faithfulness. "Every time people start to drift in that direction, we course-correct it. We make these decisions constantly. It's not something that we broadcast much, but I think it is something we are leading the industry on."
Adaptability may be the most candid principle. It admits OpenAI is "a much larger force in the world than it was a few years ago" and commits to transparency "when, how, and why our operating principles change," giving a concrete example: "we can imagine periods in the future where we have to trade off some empowerment for more resilience." Altman described the same mindset in 2025: "The price of being on the forefront of innovation is you make a lot of dumb mistakes because you're so deep in the fog of war." He traced iterative deployment to the GPT-2 release: "in retrospect that was a misplaced worry, but it led to us discovering the strategy of iterative deployment, which has been one of the most important things we've figured out."
The principles also reveal what the company refuses to do. Brockman said OpenAI shut down robotics when GPT-3 succeeded, and recently "made the painful decision to pause Sora and the browser to redirect resources" toward coding agents showing "massive, compounding growth." Altman called this "continuously focus on our critical path." The 2026 principles note that "we expect there will be periods where we need to collaborate with governments, international agencies, and other AGI efforts to ensure that we have sufficiently solved serious alignment, safety, or societal problems before proceeding further with our work." That language — "before proceeding further" — is a gate, not a guideline.
Employee accounts and leadership interviews show these principles functioning as a decision framework, not wall art. When DALL-E 2 was ready for release in 2022, the team held it back four months for safety training to prevent child sexual abuse material, deepfakes, and assault content. Mira Murati described the deliberation: "we sort of hit a point where we could really benefit from having more feedback from how people are using it... and learn more about this technology that we have created and start bringing it in the public consciousness." The same iterative logic drove the ChatGPT launch: Altman overruled internal skepticism and pushed a "research preview" of GPT-3.5 in November 2022, betting that society needed to experience the capability incrementally. "If we had launched ChatGPT with 4, it would have been a mega, mega viral moment," he said later. "The fact that there was a little bit less demand because the model was much worse than 4 turned out to be, that was really important."
The tension between research-lab origin and consumer-company reality runs through every principle. Altman admitted in 2025: "What I wanted was to get to run an AGI research lab and figure out how to make AGI. I did not think I was signing up to have to run a big consumer Internet company." Brockman's 2026 cycling quote — "It never gets easier; you just go faster" — captures the tempo the principles sustain. The principles carry a decade of course corrections: from a nonprofit research lab with "no idea, no even sketch of an idea" to a company managing half a billion users, a Microsoft partnership, and infrastructure demands Altman summarized as "transistors are the primary bottleneck, followed closely by electrons, meaning the physical energy grid."
The principles don't resolve the contradictions. They document them: the same breakage French-Owen described in the monolith, now written into policy.
Inside the Interview Loop
The interview loop runs two weeks to two months, typically five to seven rounds from recruiter screen to final decision. Glassdoor data shows a 3.3 difficulty rating out of 5, with 39 percent of candidates rating the experience positive, numbers reflecting a process designed, in the company's words, to "stretch you beyond your comfort zone." Most applicants never clear the resume screen.
The funnel starts with a recruiter review taking about a week. If a fit exists, a coordinator schedules a 30-minute call with the hiring manager or recruiter. Mid-to-senior engineers then face a technical phone screen and a system design screen, each an hour, sometimes run back-to-back as a mini-onsite before the real one. Candidates hear back within a week after each stage.
Skills assessments vary by team: pair coding, take-home projects, technical tests, sometimes more than one. Questions skew practical over algorithmic: traversing file systems, implementing KL divergence for continuous distributions with the formula provided, calculating expected iterations for probabilistic functions, finding minimum error using cross entropy. Interviewing.io reports niche topics specific to OpenAI: time-based data structures, versioned data stores, coroutines and concurrency primitives, object-oriented design patterns. Questions run multi-part, typically four segments, with tight timeboxes. Most candidates use every second. A weak coding score can sink an offer even when other rounds go well; OpenAI wants engineers who ship, not just design.
The onsite (virtual by default, with an option to visit San Francisco headquarters) spans four to six hours across one to two days. Candidates who have already passed coding and system design screens skip those. The loop shifts to: an agentic coding round in beta, a behavioral interview with a senior manager, a technical presentation, a coding session, a system design session, and a team-focused behavioral. The beta agentic round hands candidates an existing codebase and a problem too large to solve from scratch; they work through it using an AI coding agent. AI assistance stays prohibited in every other round.
Senior candidates prepare a four- or five-slide presentation on a significant project they led: problem context, approach, trade-offs, impact, lessons. Interviewers evaluate clarity, depth of thinking, and the ability to tell a compelling technical story. System design discussions demand thoughtful trade-offs, API and schema design, scalability and reliability reasoning, and occasional feature suggestions. Interviewers drill into any specific technology a candidate names, so naming tools without deep readiness backfires.
Behavioral rounds probe cross-disciplinary collaboration — researchers, product managers, safety specialists — not generic teamwork. Mission alignment carries weight: candidates should articulate a genuine perspective on AGI safety and beneficial AI development without needing to be alignment researchers. Interviewers described as conversational may hold veto power.
After the onsite, interviewers debrief. The hiring team reviews the complete profile: application, skills assessments, final loop performance. Three outcomes exist: an offer, a rejection (often with minimal feedback), or a request for an additional interview to clarify a specific competency. OpenAI has a reputation for downleveling incoming candidates; anchoring expectations to your current title is a mistake.
The process selects for independent problem-solving under time pressure, technical overcommunication — narrating reasoning, trade-offs, and adjustments in real time — and comfort with ambiguous goals across research, product, and safety boundaries. Candidates who treat the loop as a performance rather than a collaboration tend to miss the signal OpenAI is actually measuring. That signal is the same autonomy the monolith demands.
What the Offer Letter Doesn't Tell You
OpenAI pays like a late-stage private company competing for the same researchers Google DeepMind and Anthropic chase. Levels.fyi reports median total compensation of about $612,000, but that figure flattens a wide spread. Base salaries are competitive with peer labs — not exorbitant — and typically represent one-third to two-thirds of total pay. The equity side pushes packages toward the top of the market. Details by level appear in the table below.
| Level / Role | Base Salary Range | Equity (Annualized) | Total Comp (Avg) |
|---|---|---|---|
| L2–L6 Software Engineer | $170,000–$395,000 | Variable (PPU/RSU) | ~$612,000 (L4) |
| L6 Hardware Engineer | $325,000 | $875,000 | $1,200,000 |
| Research Engineer (board postings) | $293,000–$585,000 | Included in range | $293,000–$585,000 |
| Research Scientist tracks | $295,000–$555,000 | Included in range | $295,000–$555,000 |
Equity has undergone three regimes. Until 2019 the organization was a nonprofit; the capped for-profit pivot introduced Profit Participation Units (profit-sharing instruments, not ownership) with early investor returns capped at 100x. Employee caps dropped from 100x to 10x to 4x, Levels.fyi says, with a two-year lockup. Vesting now runs quarterly over four years. As of 2026, new hires receive double-trigger RSUs, and existing PPUs have converted to uncapped traditional equity alongside the transition to a public benefit corporation. That conversion matters: early employees who held through the caps may now hold uncapped shares they can sell in tender offers or, eventually, on a public market. The most recent tender, early 2026, let staff sell up to $30 million each, moving $6.6 billion total, about $11 million per seller on average, the Wall Street Journal reported. Bonuses are irregular; the lone exception came in August 2025, when OpenAI paid retention bonuses up to $1.5 million for senior technical staff amid a talent crunch, Levels.fyi says.
Benefits are comprehensive where they exist: fully employer-paid health, dental, and vision for employees and families; generous fertility coverage; parental leave of about six months for the birthing parent and five for the other, plus a month of work-from-home flexibility; a 50% 401(k) match with mega backdoor Roth access; daily meals in-office; and an annual learning stipend plus conference budget.
The trade-off is structural. Equity-heavy pay with periodic liquidity events replaces steady raises and predictable bonus cycles. BuiltIn's 2026 employee feedback notes stagnant progression-linked increases and ad-hoc top-ups rather than routine refreshers. Unlimited PTO exists on paper; the pace described earlier makes it theoretical. Candidates weighing offer letters should model the equity as a series of illiquid bets with occasional tender windows, not a liquid salary substitute. The compensation structure mirrors the culture: high upside, no safety net, and the same breakage French-Owen navigated daily.
Who Thrives Here and Who Doesn't
The mission is not decorative. Employees who stay and advance describe the AGI goal as the filter that makes the intensity legible: without it, the pace reads as burnout; with it, the same pace reads as necessary. Glassdoor reviews consistently cite "impact" as the primary reason people accept the trade-offs. ChatGPT reaches over 200 million weekly users. The APIs power thousands of applications. When you ship something, you see it in the world almost immediately, at massive scale. That feedback loop is the company's strongest retention mechanism.
People who thrive share a cluster of traits. First, the mission has to resonate viscerally: whether framed as "building beneficial AGI" or "building the most important technology in human history." Second, they want global-scale impact now, not eventually. Very few companies let you push code that millions use next week. Third, they treat ambiguity as fuel. Priorities shift, org structures change, and the ship-fast mentality means direction can reverse before a consensus forms. Engineers and researchers who find that energizing rather than exhausting tend to stay. Fourth, the compensation package ($350k–$550k total for engineers, with senior and staff roles pushing well above that) attracts people who want elite pay without the equity-only gamble common at earlier-stage labs. Base salaries are above-market, competitive with top-of-market big tech, unusual for a company at this stage. Fifth, they want to work alongside exceptional colleagues. The technical bar is extraordinarily high; the caliber of talent is the second most-cited pro in reviews.
The mismatch profile is clear. The 3.6 work-life-balance score (below peers like Anthropic (3.7) and well behind Linear (4.4) or Notion (4.2)) reflects a reality where nights and weekends during launches are common. If firm boundaries are non-negotiable, OpenAI will frustrate you. Organizational stability is another fault line. The company has undergone significant org changes, leadership shifts, and the nonprofit-to-capped-profit-to-for-profit transition that created cultural scars still referenced in reviews. Some employees who joined for the nonprofit mission feel priorities have shifted; others argue commercial success accelerated the research. That tension is a live wire. Career ladders can feel undefined, especially outside core research and engineering, as the workforce grew roughly 55% in 2025 to about 7,800 employees. Promotion criteria shift with the org. Finally, consensus-driven decision makers struggle. OpenAI moves fast and makes decisions quickly. Not everyone gets input on every decision. If you need buy-in before action, the pace feels disorienting.
French-Owen's team shipped Codex in seven weeks. The monolith still creaks. The principles still document contradictions rather than resolve them. The interview loop still tests for the signal: can you build inside the breakage?
Working in frontier tech? Zero G Talent tracks the openings: see every open OpenAI role, browse frontier tech jobs, the companies hiring, and the people building the field.