OpenAI

Confidence 0.95 · 26 sources · last confirmed 2026-10-18

AI research and deployment company. Originator of ChatGPT (Nov 2022) and the GPT-3 / GPT-4 / GPT-5 series. Self-described in OpenAI HBR article as “an AI provider to over one million businesses.”

Governance trajectory

Originally founded (2015) as a nonprofit foundation. Two or three “OpenAI crises” ago (Ries 2026, 1:12:28), Dario and Daniela Amodei left OpenAI to spin out Anthropic — the moment Ries dates as his own entry into the AI-governance-advice arc. Mid-2025, OpenAI converted from the nonprofit-foundation governance to Public Benefit Corporation (PBC) structure (Ries: “although now they’ve converted to a public benefit corp structure”).

Per Ries’ framework, PBC alone is structurally weaker than the PBC + Long-Term Benefit Trust structure adopted by Anthropic — a PBC replaces the any-lawful-act-or-purpose clause with a specific declared purpose (making mission-pursuit a fiduciary obligation of the board) but lacks an outside trustee body holding the for-profit board accountable to that mission. The implication Ries does not state outright but the structural comparison implies: OpenAI’s original nonprofit-foundation guardian was stronger than the eventual PBC-only structure; the 2025 conversion is a governance downgrade by Ries’ criteria.

Ries describes the OpenAI history as “a really hard case study to learn from because it’s such a bizarre story and has involves like mega personalities like Elon and Sam like dueling to the death.” The deeper case-study narrative is deferred to Incorruptible (book, out 26 May 2026), not given on-air.

Models referenced in this wiki

• GPT-3 — substrate of the customer-support augmentation system in Brynjolfsson, Li & Raymond (2025) QJE.

• GPT-3.5 / ChatGPT — the democratization breakthrough cited by Anand-Wu and the trigger for the GenAI adoption wave.

• GPT-4 — basis of the Dell’Acqua et al. 2026 BCG RCT and the Boussioux et al. 2024 crowdless-future study.

• GPT-5 — used by Lopopolo’s Codex team (Feb 2026) to generate the initial repository scaffold via Codex CLI.

• GPT-5.6 Luna — per Bloomberg Originals (Sep 2026), launched in July 2026 as OpenAI’s most cost-efficient model. In Bloomberg’s single-task test it built a small-business website for just over $4, against about $12 on Moonshot’s Kimi K3 and about $50 on Claude Fable 5. A startup founder who had moved to Chinese open models on cost reports a price cut of about 80% and began testing Luna again. It is the wiki’s first recorded case of a frontier lab repricing in response to open-weight competition; see open-source-ai.

Products and tooling referenced in this wiki

• Codex — agent-driven software engineering platform documented in Lopopolo 2026. Five months of internal use produced ~1M LOC and ~1,500 PRs across an internal product, with 0 lines of manually-written code and throughput averaging 3.5 PRs / engineer / day across 7 engineers (increasing with team size).
• Codex CLI — command-line interface for Codex; used (with GPT-5) to generate the initial repository scaffold for Lopopolo’s project. The wiki’s first concrete operational reference to Codex CLI tooling.
• Aardvark — second agent operating on the same Codex-team codebase, named in passing in Lopopolo 2026 as a beneficiary of repository legibility investments. Not detailed.
• Sora — text-to-video model + social-network app. Surfaces in Nika 2025 as the parallel demo to Google Flow / Veo for PM-side promo videos, with the Cameo feature (custom-instruction-shipped — “always make me look good”; reportedly Mark Cuban’s cameo auto-injects a Cost Plus Drugs ad). Marily / Claire’s tentative read: “the social-network aspect of it is a distribution method for the underlying API models — power users find what’s possible, then commercial applications follow.”
• Custom GPTs — user-authored, system-prompt-+-knowledge ChatGPT instances. Nika 2025 demonstrates a domain-specific custom GPT (“AI Product GPT”) seeded with a PRD template and the author’s voice, used as a structured prompt-input-to-prototype-tool pipeline component. Operationally a domain-specific Context-layer artifact in Chatterjee’s vocabulary.

OpenAI as an AI-native organization (first-party deployment vantage)

The wiki’s first first-party OpenAI source — Beutler 2026, a talk by Joe Beutler (Head of Solutions Engineering, Strategics) at IT Revolution’s Enterprise AI Summit — reports how OpenAI is restructuring itself and what its solutions team sees succeed across its largest enterprise customers:

• Internal wins: the finance team uses the OpenAI API to extract and structure contract data for “millions of dollars of impact,” and built a GPT that answered “thousands” of investor due-diligence questions during the recent fundraise. The go-to-market team automated top-of-funnel work (lead qualification, quote preparation) for the SMB self-service segment.
• The PwC benchmark: PwC benchmarked OpenAI’s finance team at only 20% of the size it should be for a firm of its revenue complexity / transaction volume / regulatory burden — “not just against legacy finance orgs … also against comps that are more digitally native.”
• Embedded engineering throughout the company: per Beutler (and Gene Kim’s introduction), the CFO and CRO each have an engineering manager working on AI integration; OpenAI has embedded engineering managers across business functions. The pattern emerged organically (domain expert builds tooling → full-time role → paired with an engineer interviewed by and on the same comp ladder as engineering → head of innovation), and OpenAI frames it as “moving centers of excellence to embedded innovation.”
• Vendor strategy — sell the builder, not the outcome: Beutler explicitly contrasts OpenAI’s approach with “a massive professional services business … lucrative contracts that will last 10 years,” betting instead on no-code agent builders (continuous evals + governance + connectors + shareable skills) so individuals build their own agents. This is the deliberate counter-position to the outcome-as-a-service thesis in YC and the buyer-side Buy Outcomes model in Nishar-Nohria.
• Ask → Assist → Automate: the deployment-maturity framework OpenAI’s solutions team uses with customers (read-only → human-in-the-loop → full automation). See automation-vs-augmentation and enterprise-ai-adoption.

Note OpenAI’s recurring appearance on the deployment side of wiki sources (the Bain/OpenAI HBR article; the ChatGPT access-democratization anchor): Beutler’s talk is the first where OpenAI describes its own organizational redesign rather than a customer’s.

Research initiatives appearing in this wiki

• Economic Research team (led by Chief Economist Aaron Chatterji). Counterpart to Anthropic’s Anthropic Economic Index initiative. Members in this wiki: Gawesha Weeratunga, Harrison Satcher.
• GDPval (Patwardhan et al. 2025) — OpenAI’s benchmark measuring frontier-model performance on real-world economically valuable tasks (1,320 tasks, 44 occupations, top-9 US-GDP sectors), graded by head-to-head win rate vs human experts. Best model at publication: Claude Opus 4.1 (47.6% wins-or-ties). Public automated grader at evals.openai.com. The capability-measurement counterpart to OpenAI’s adoption-focused Economic Research; complements Anthropic’s Economic Index (usage) and the AI Index (academic benchmarks).
• Lowe’s partnership (2026-05-02-dutt-chatterji-ai-experimentation-to-transformation): launched Mylow (online customer-facing) and Mylow Companion (in-store associate-facing) AI interfaces in March 2025; Mylow Companion deployed across 1,700+ Lowe’s stores.

Other references

• Supporting partner of the AI Index (annual report from Stanford HAI).

• Anchor case in the Anand-Wu GenAI Playbook (ChatGPT as the access-democratization breakthrough).

• Subject of the open-source-AI strategic-analogy example (GPT-4 = Walmart, open source = fruit stand) discussed in Carroll & Sørensen 2024.

• Investment-figure context: U.S. private AI investment $285.9B in 2025 (2026-04-30-ai-index-report-2026).

• Named by US Under Secretary of War Emil Michael (“OpenAI’s leadership”) among the signatories of a White House compact “for the safe development of superintelligence”, late September 2026, which in his account commits the labs to “auditing and independent evaluations” (2026-09-30-michael-miller-department-of-war-ai-adoption). His description only; the compact’s text is not in the wiki.

A critic’s sustainability doubt (added 2026-09-21)

• 2026-09-18-swisher-lafrance-atlantic-festival-tech-economy-is-ai-reshaping-america — Kara Swisher, predicting consolidation of the frontier labs to two, says OpenAI’s economics are “too expensive… just not sustainable” and that its failure would matter because AI spending “props up the stock market.” Opinion.

Open questions

• The wiki has many references to GPT-4 / ChatGPT / OpenAI as substrate but no first-party OpenAI source has been ingested yet. The OpenAI Economic Research team’s writing (this article, plus their independent research papers) is a candidate for a deeper deep-read once a primary OpenAI Economic Research publication is added.

Founding pitch as a YC case study (added 2026-09-28)

Dalton + Michael (Aug 2026) cite OpenAI as a company lifted by many “secret hands” because of its ambition, and its founding pitch as grounded in named researchers and published papers, not a promise of “magic AI”.

Working beside a defense prime’s engineers (added 2026-10-04)

Hiza (Lockheed Martin) on Bloomberg Tech, October 2026 announced on air that OpenAI’s team is working with the F-35 programme on “math and physics challenges” raised by a new sensor: “Their team is working right beside us.” Lockheed approached OpenAI after its reported solution of a long-standing math problem. It is the wiki’s first case of the lab’s people, rather than its models or products, working inside a customer’s engineering programme. No scope, terms or result were given. Lockheed otherwise stays model-agnostic, with 55 models in use.

Its consumer agent, blocked (added 2026-10-18)

Roth (The Verge, Oct 2026): OpenAI’s agent Dots is among the consumer agents users report blocked on major sites, Amazon among them. OpenAI did not reply to The Verge in time. An analyst quoted in the piece asks whether such refusals affect “the perceived value of the Anthropics and the OpenAIs”, and answers “absolutely.”