OpenAI company dossier 061
OpenAI: Company, Models, Products and Platform
Governance, research, model families, ChatGPT, Codex, media, agents, developer infrastructure, distribution and compute—and how the parts fit together.· verified
01
Executive summary
OpenAI is a controlled, mission-led AI group whose research, models, consumer applications and developer platform form one vertically integrated product system. The OpenAI Foundation controls OpenAI Group PBC; investors hold economic interests, but economic ownership is not the same as governing control.[1][2]
The operating chain is: external capital and compute support model research and training; trained models are exposed through ChatGPT, Codex and specialised media products or through the API Platform; tools, agents and partner integrations turn model outputs into work; consumer, business and developer usage funds the next infrastructure and model cycle. The model is one component. The product determines context, tools, permissions, execution and distribution.
Controlling bodyFoundation[D] Controls OpenAI Group PBC [1]
Current frontier familyGPT-5.6[D] Sol, Terra and Luna [3]
Core developer interfaceResponses[D] Model, state and tool orchestration [7]
Principal direct systemsChatGPT + Codex[D] General work and software engineering
Evidence markers. [D] officially disclosed; [R] credibly reported; [C] calculated from disclosed or reported facts; [I] analytical inference; [ND] not publicly disclosed. Markers identify decision-relevant evidence boundaries; an unmarked historical statement is supported by the cited primary-source ledger.
Scope. Prices, promotional superlatives and unnormalised benchmark rankings are excluded. OpenAI does not publish enough information to reconstruct the parameter architecture, training corpus, full compute bill or unit economics of its closed frontier models.
Essential questions
| Question | Concise answer |
|---|---|
| What exactly is OpenAI? | A controlled group comprising the OpenAI Foundation and OpenAI Group PBC. It conducts AI research and operates consumer, enterprise and developer products rather than functioning only as a model laboratory. |
| Who controls it? | The OpenAI Foundation controls the public benefit corporation and appoints its governing board. Microsoft and other investors have economic and contractual interests, but those interests do not by themselves equal Foundation control. |
| What is the difference between GPT and ChatGPT? | GPT identifies model families. ChatGPT is an application system that combines selected models with conversation state, files, tools, browser and computer capabilities, memory, workspace policy and a user interface. |
| Is the ChatGPT model the same as the API flagship? | Not necessarily. The chat-latest alias follows the changing Instant model used in ChatGPT, while production API users can select documented model IDs and dated snapshots. Product routing and API model identity must be treated separately.[5] |
| Where does Codex fit? | Codex is OpenAI's software-engineering agent system. It uses models, repository context, tools, execution environments, permissions and review workflows. The companion Codex dossier will cover that architecture in depth. |
| What is the OpenAI API Platform? | The developer delivery layer for models, Responses, tools, state, agents, evaluation, administration and deployment controls. It lets customers build their own applications rather than using only OpenAI's direct products. |
| Are Computer Use and Agent Builder separate models? | No. Computer Use is a tool interface and action loop. Agent Builder is a legacy visual workflow product scheduled for retirement. Both depend on models and host-side execution rather than replacing them. |
| Does OpenAI publish model architecture? | Only selectively. The open-weight gpt-oss models publish weights and parameter information. Closed GPT-5.6 and specialist models publish interfaces and operating limits but not enough internal detail for reproduction. |
| Who pays? | Consumers and teams pay for direct applications; enterprises buy managed workspace and platform access; developers and software companies consume API models and tools; strategic partners distribute capacity and products. |
| Why does the Microsoft relationship matter? | Microsoft is simultaneously an investor, infrastructure partner and distribution channel. These roles create material dependencies, but they should not be collapsed into a claim that Microsoft operates or unilaterally governs OpenAI. |
| How fast is the model portfolio changing? | The API frontier moved from GPT-5.4 on 5 March to GPT-5.5 on 24 April and GPT-5.6 on 9 July 2026: 126 days from 5.4 to 5.6. Product aliases, specialist models and tools changed between those releases. |
| What most constrains the roadmap? | Compute, power, data-centre delivery, model reliability, inference capacity, safety controls, enterprise governance, regulation and the ability to convert capability into useful products at sustainable operating cost. |
02
The OpenAI system and its seven objects
| Canonical layer | What belongs here | Do not confuse it with | Evidence boundary |
|---|---|---|---|
| Company and governance | OpenAI Foundation, OpenAI Group PBC, boards, management, investors and contracts. | A model, product or platform service. | [D] Control structure is disclosed; [ND] the complete current cap table and private contractual rights are not. |
| Model family | A related public grouping such as GPT-5.6, GPT-Image or GPT-Realtime. | A single deployable model ID or a product label. | Family membership describes the published portfolio, not undisclosed technical lineage. |
| Model instance | A named, selectable model or snapshot such as GPT-5.6 Sol or gpt-5.6-sol. | A moving alias, ChatGPT label or complete application. | Interfaces and limits may be disclosed while closed weights, data and architecture remain [ND]. |
| Product surface | OpenAI-operated user systems such as ChatGPT, Codex and media products. | The underlying model, API identifier or a reusable tool. | Products may route among models and add state, tools, policy and user experience. |
| Platform abstraction | Responses, Realtime, SDKs, agents, evaluation and administration interfaces. | A model or an independently acting agent. | The platform coordinates services; customer systems retain responsibility for their own credentials and side effects. |
| Capability or tool | Search, file retrieval, code execution, computer use, image/audio processing, functions and MCP connections. | An inherent capability of every model or product. | Availability depends on model, API, product, account, region and host policy. |
| Infrastructure | Accelerators, cloud, data centres, power, storage and networks used to train and serve models. | Model capability or immediately usable customer capacity. | Announcements, contracts, construction and accepted production capacity are different states. |
How the seven layers fit. Research is conducted inside the company and produces or evaluates models. Distribution carries products and platform services to customers. Neither research nor distribution is itself a model layer.
Central thesis. OpenAI's strategic asset is not one model name. It is the feedback system joining research, model deployment, direct product usage, developer adoption, tool execution and infrastructure scale.
03
From research laboratory to integrated product platform
| Period | Milestone | System consequence |
|---|---|---|
| 2015 | OpenAI established as a non-profit AI research organisation. | Mission and control began outside a conventional investor-owned company. |
| 2019 | A capped-profit operating entity and Microsoft partnership added capital and cloud infrastructure. | Research became linked to a commercial deployment and compute model. |
| 2020 | GPT-3 and the API turned general-purpose models into developer infrastructure. | External software companies could build products on hosted OpenAI inference. |
| 2021 | Codex and early code-generation products specialised the model stack for software work. | Coding became a distinct product and model pathway. |
| 2022 | ChatGPT packaged conversational models into a direct mass-market application. | OpenAI gained a direct product feedback and distribution channel. |
| 2023 | GPT-4, enterprise ChatGPT and multimodal capabilities broadened professional use. | Consumer chat expanded into managed workplace and developer deployment. |
| 2024 | GPT-4o joined text, vision and voice; o-series models separated deliberate reasoning from faster interaction. | Model selection became a portfolio decision rather than a single flagship choice. |
| 2025 | Responses, Agents SDK, computer use, deep research, ChatGPT agent, GPT-5 and Codex expanded tool-using work. | The platform moved from generation endpoints towards stateful agent execution. |
| 2025 | The Foundation-controlled public benefit corporation became the operating structure. | Capital formation and mission control were separated more explicitly. |
| 2026 | GPT-5.4, GPT-5.5 and GPT-5.6 successively expanded context, tools and agent orchestration. | The flagship model became one layer inside longer-running, tool-rich systems. |
| 2026 | Realtime 2.1, GPT-Image-2, Daybreak and GPT-5.6 Cyber widened specialist modalities and governed access. | OpenAI operates multiple model families with different interfaces and safety boundaries. |
Reading the chronology. Chronological succession, product replacement and technical or training lineage are different claims. The rows show public release and product evolution; they do not imply disclosed weight, architecture, dataset or training lineage unless the cited source explicitly says so.
04
The pace of frontier-model development
The current API sequence is unusually compressed. GPT-5.4, GPT-5.5 and GPT-5.6 were released within 126 days. This demonstrates deployment cadence, not a measured 126-day training cycle: research, training, evaluation and infrastructure work overlap before a public release.[4]
| Frontier release | Date | Days from prior release | Principal published system change | Product implication |
|---|---|---|---|---|
| GPT-5.4 | 5 Mar 2026 | — | One-million-token context, native compaction, tool search and built-in computer use. | Longer agent runs and larger tool surfaces became first-class API concerns. |
| GPT-5.5 | 24 Apr 2026 | 50 [C] | Expanded hosted tools, Skills, MCP, shell and apply-patch support. | The model interface moved closer to a reusable agent runtime. |
| GPT-5.6 | 9 Jul 2026 | 76 [C] | Sol, Terra and Luna tiers; persisted reasoning, programmatic tool calling, max effort, pro mode and multi-agent beta. | Capability, latency and throughput can be selected within one family while orchestration grows more explicit. |
| GPT-5.6 Cyber | 7 Aug 2026 | 29 [C] | Purpose-trained cybersecurity model under separately approved Daybreak Red access. | Specialist capability is governed through programme access rather than general availability. |
Cadence boundary. Release intervals do not establish architecture inheritance, benchmark improvement, commercial adoption or economics. ChatGPT aliases and Codex models can move on a different schedule from the general-purpose API frontier.
05
The current OpenAI model portfolio
As of 24 August 2026. Read names in four layers: family → named model → API/model identifier → ChatGPT-facing label. A dated snapshot is fixed; an alias may move to a newer implementation. Product labels do not prove API-model identity.[3][5]
| Named model / family | API ID or product label | Published interface | Status / availability | What is not established |
|---|---|---|---|---|
| GPT-5.6 Sol GPT-5.6 family | gpt-5.6-sol; gpt-5.6 alias | 1.05M context; 128K maximum output; text and image input; configurable reasoning; supported tools. | Current · API | [ND] Parameter count, training compute, dataset, dense/MoE design and full post-training recipe. |
| GPT-5.6 Terra GPT-5.6 family | gpt-5.6-terra | Published long-context, text/image, reasoning and tool interface. | Current · API | [ND] Whether it is distilled, routed or independently trained. |
| GPT-5.6 Luna GPT-5.6 family | gpt-5.6-luna | Published long-context, text/image, reasoning and tool interface. | Current · API | [ND] Internal size, active parameters and exact capability trade-offs. |
| Chat Latest product alias | chat-latest; ChatGPT Instant label | Moving alias intended to track changing ChatGPT chat behaviour. | Current alias · testing, not a fixed snapshot | Equivalence to a fixed GPT-5.6 model or stable reproducibility. |
| GPT-5.6 Cyber specialist family | gpt-5.6-cyber; Daybreak access label | Cybersecurity-oriented Responses workflows and tools. | Current · restricted programme access | General availability or permission to test systems without authorisation. |
| GPT-Image-2 image family | gpt-image-2 | Text/image inputs and image generation or editing output. | Current · API and product surfaces | A video, general-reasoning or conversational model. |
| GPT-Realtime-2.1 realtime family | gpt-realtime-2.1 and documented variants | Low-latency audio/text sessions and tool-using interaction. | Current · Realtime API | Interchangeability with text-response models or every endpoint. |
| Transcription and speech audio families | Task-specific documented model IDs | Speech recognition, diarisation and speech generation. | Current · specialist APIs | A single universal voice model. |
| gpt-oss-120b / 20b open-weight family | gpt-oss-120b; gpt-oss-20b | Apache 2.0 weights; 120b publishes 117B stored and 5.1B active parameters. | Current · external hosting | The architecture or behaviour of OpenAI's closed frontier models. |
| Embeddings and moderation specialist families | Task-specific documented model IDs | Vector representation and safety classification. | Current · API | General generation or autonomous-agent capability. |
[ND] Closed-model boundary. Context length and tool support describe the service interface, not the neural architecture. A one-million-token request limit does not prove uniform recall, attention cost or accuracy across the complete window.
06
Products people and organisations use directly
As of 24 August 2026. Product surfaces can combine multiple models, tools, retrieval systems, stored state, policy and user-interface logic.
| Product system | Primary job | What OpenAI adds around the model | Principal users |
|---|---|---|---|
| ChatGPT | General knowledge work, analysis, creation and interaction. | Conversation state, files, search, data analysis, images, voice, browser/computer tools, memory and workspace policy. | Individuals, teams, enterprises and education. |
| ChatGPT agent | Longer-running research and action across websites and connected systems. | Task planning, browser/computer interaction, tools, checkpoints and user confirmations. | End users and managed workspaces. |
| Codex | Software engineering across repositories and development environments. | Repository context, shell and file tools, sandboxes, cloud environments, worktrees, review and integrations. | Developers and engineering organisations. |
| Sora and media tools | Video, image and audio creation or transformation. | Media-specific interfaces, asset handling, provenance and safety systems. | Creators, product teams and developers. |
| Workspace administration | Managed organisational deployment of ChatGPT and Codex. | Identity, roles, model controls, plugins, connectors, analytics, compliance and audit interfaces. | Enterprise, public-sector and education administrators. |
Product boundary. ChatGPT capability ≠ base-model capability. ChatGPT agent, Codex and Computer Use overlap in tools but are not synonyms. ChatGPT agent is a general end-user workflow; Codex is a software-engineering system; Computer Use is a tool loop that a host application must execute and supervise.
07
API Platform and agent infrastructure
As of 24 August 2026. This is the public product and platform stack, not a claim about OpenAI's undisclosed internal implementation.
| Platform layer | Role | Key components | Responsibility boundary |
|---|---|---|---|
| Models | Reason over inputs and request outputs or tools. | GPT, image, realtime, speech, embedding and specialist models. | A model proposes; it does not automatically own application side effects. |
| Responses API | Unifies model output, state, reasoning items and tool calls. | Conversation state, background mode, streaming, WebSocket, compaction and multi-agent beta. | The API coordinates records and hosted tools; customer systems still own their functions, credentials and policies. |
| Built-in tools | Add retrieval, computation, media and interaction. | Web search, file search, code interpreter, shell, computer, image generation and apply patch. | Each tool has separate availability, execution and safety boundaries. |
| External tools | Connect business systems and custom actions. | Function calling, MCP, connectors and tool search. | The external server or customer function performs and authorises the effect. |
| Agents SDK | Build coded agents with tools, handoffs, state, tracing and guardrails. | Agent definitions, orchestration, sandbox agents, results and observability. | The SDK is an application framework, not an autonomous service or foundation model. |
| Agent Builder | Legacy visual workflow composition. | Canvas, typed nodes, preview, ChatKit and export. | Scheduled shutdown 30 November 2026; existing workflows require migration. |
| Evaluation and optimisation | Measure and improve model or agent behaviour. | Evals, graders, tracing, prompt optimisation and fine-tuning. | Evaluation quality depends on representative tasks and valid graders. |
| Administration | Control organisational access and operations. | Projects, service accounts, RBAC, data controls, regional processing, spend and usage controls. | Platform controls do not replace application-level authorisation or governance. |
Agent execution chain. User goal → application context → model reasoning → tool request → policy/approval → host or service execution → observation → verification. API platform capability ≠ individual model capability. Model intent is not execution authority.[7]
08
Compute, capital and distribution infrastructure
As of 24 August 2026. OpenAI is operationally dependent on infrastructure it does not manufacture itself. The company combines long-term partnerships, cloud capacity and the Stargate build-out with its own model and product systems. The relevant analytical unit is usable end-to-end capacity: accelerator, memory, network, storage, power, cooling, software and deployment readiness must all be available together.
| Partner / route | Capital or financing | Governance | Compute / infrastructure | Distribution | Strategic dependency |
|---|---|---|---|---|---|
| Microsoft | [D] Material economic investment and continuing commercial relationship. | [D] No replacement of Foundation control. | [D] Azure capacity and infrastructure partnership. | [D] Enterprise and cloud channel. | High across capital, capacity and route to market; this is not unilateral legal control. |
| SoftBank | [D] Initial Stargate equity funder; financial responsibility assigned in the project announcement. | [ND] No OpenAI governance role disclosed in the cited announcement. | [D] Stargate financing and build-out coordination. | [ND] No OpenAI product-distribution role disclosed there. | High for the announced Stargate financing and delivery route; project-specific, not company control. |
| Oracle | [D] Initial Stargate equity funder. | [ND] No OpenAI governance role disclosed there. | [D] Initial technology partner and infrastructure route. | [ND] No general OpenAI distribution role established by Stargate. | Material to one capacity path; announcement does not prove accepted production capacity. |
| MGX | [D] Initial Stargate equity funder. | [ND] No OpenAI governance role disclosed there. | Exposure through Stargate financing rather than a disclosed operating role. | [ND] None disclosed in the cited project announcement. | Capital relationship; operational dependence is not established. |
| Arm and NVIDIA | [ND] No Stargate equity role disclosed in the cited announcement. | [ND] No OpenAI governance role disclosed. | [D] Named initial Stargate technology partners. | [ND] No OpenAI product-distribution role disclosed there. | Technology ecosystem dependence; a partner label does not disclose volumes, delivery dates or workload allocation. |
| Direct OpenAI products | Customer receipts rather than external financing. | OpenAI-operated product route. | Consumes internal and partner-supplied serving capacity. | Direct ChatGPT, Codex and media user relationships. | High integration and product telemetry; model-level economics remain [ND]. |
Capacity states. Announcement ≠ contract ≠ construction ≠ installed hardware ≠ energised cluster ≠ accepted production capacity ≠ capacity available to a particular model or customer.
Integration versus dependency
| Layer | OpenAI integration / control | External dependency | Analytical reading |
|---|---|---|---|
| Model research and training | High: research direction, training systems and post-training. | Compute, power, hardware, data ecosystem and specialist suppliers. | [I] Strong intellectual integration with substantial physical dependence. |
| Serving | High to moderate: model service and scheduling layers. | Accepted accelerator, cloud, network, power and facility capacity. | [I] Customer availability depends on the complete capacity chain, not chip access alone. |
| Developer platform | High: APIs, state, hosted tools, SDKs, evaluation and administration. | Customer functions, MCP services, connectors and external data systems. | [I] OpenAI controls the runtime interface but not every tool or business-system side effect. |
| Consumer products | High: ChatGPT, Codex and media product surfaces. | App, browser, device, payment and content ecosystems. | [I] Direct distribution creates feedback and customer access, but still crosses external platforms. |
| Enterprise deployment | High: workspace controls and direct contracts. | Identity, security, cloud, connectors, procurement and customer governance. | [I] Product control does not remove integration and adoption dependence. |
| Capital and compute expansion | No complete self-financing or manufacturing chain is publicly established. | Investors, project finance, hyperscalers, data-centre builders, energy and hardware partners. | [I] The broadest strategic dependency sits below the software stack. |
09
Who uses OpenAI, who pays and who distributes it
| Customer group | What they buy or use | Commercial route | Decision driver |
|---|---|---|---|
| Individuals | ChatGPT, agents, voice, images and personal workflows. | Direct consumer plans and free access tiers. | Usefulness, reliability, latency, limits, privacy and switching cost. |
| Developers and start-ups | Models, Responses, tools, realtime media and Agents SDK. | Direct API and platform accounts. | Task quality, controllability, integration effort, throughput and operating cost. |
| Software companies | Embedded model and agent capabilities inside their own products. | API agreements and strategic integrations. | Reliability, scale, data policy, model lifecycle and margin. |
| Enterprises | Managed ChatGPT/Codex workspaces and application APIs. | Direct enterprise contracts and Microsoft distribution. | Identity, compliance, data control, audit, change management and measurable productivity. |
| Public sector and regulated organisations | Governed models, workspaces and authorised specialist capability. | Approved direct or partner channels. | Security, locality, authorisation, procurement and mission assurance. |
| Creators and media teams | Image, video, voice and editing systems. | Direct products and media APIs. | Control, consistency, rights, provenance, latency and workflow integration. |
| Security teams | General defensive models or separately approved Daybreak Red capability. | API access plus programme approval. | Authorisation, reproducibility, audit and safe disclosure. |
Customer and channel are different roles. Microsoft may invest, supply infrastructure and distribute OpenAI services. An enterprise reached through Microsoft remains the end customer; the partner relationship does not reveal the exact revenue allocation.
10
Capital, ownership and governance
As of 24 August 2026. Formal control, economic ownership, financing, commercial relationships, compute dependence and strategic influence are separate dimensions. None should be used as shorthand for another.
| Stakeholder | Governance / control | Economic ownership | Financing | Commercial / compute relationship | Strategic influence boundary |
|---|---|---|---|---|---|
| OpenAI Foundation | [D] Controls the PBC and appoints its board. | [D] Holds an economic interest; exact current percentage is not treated as perpetual. | Not analysed as an external financier. | Mission and governance relationship. | Formal control does not mean day-to-day product management or sole economic ownership. |
| OpenAI Group PBC board and management | [D] Corporate oversight and operations under Foundation control. | Operates the commercial research and product company. | Raises and deploys capital through the PBC structure. | Research, products, contracts and infrastructure commitments. | Operational authority remains bounded by Foundation purpose, law and contracts. |
| Microsoft | [D] Economic and contractual importance does not replace Foundation control. | [D] Material economic exposure. | [D] Major investor. | [D] Azure infrastructure and enterprise distribution relationship. | [I] High practical influence across several layers is not identical to legal control. |
| Employees | [ND] No single employee voting bloc is disclosed. | [D] Employees collectively hold economic interests; individual holdings are [ND]. | Human capital rather than an external financing category. | Research, build and operate the system. | Operational importance does not disclose collective governance power. |
| Other investors | [ND] No equal governance rights should be inferred. | Negotiated economic interests; complete current terms are [ND]. | Provide external capital. | May also hold separate commercial relationships. | Financing, information rights and influence vary; none establishes Foundation control. |
| Infrastructure partners | [ND] Supply relationships do not establish governance authority. | Project or contractual economics vary. | May finance facilities or capacity. | Compute, data centres, energy, hardware and systems. | [I] Operational leverage can be material without equity or board control. |
Investor boundary. The Foundation-controlled PBC structure is public. The complete current fully diluted ownership, investor rights, liquidation preferences and employee-by-employee holdings are not. Any percentage from a restructuring announcement is a dated snapshot, not a perpetual cap table.
11
Economics: can product demand absorb infrastructure scale?
OpenAI's economic system joins a high fixed-cost research and infrastructure base to multiple distribution routes. Direct applications can monetise users and organisations; APIs monetise consumption by external products; enterprise and partner channels add governed distribution. None of these routes discloses model-level gross margin or the cost of an individual capability.
| Economic layer | Cash or value mechanism | Scaling advantage | Principal constraint |
|---|---|---|---|
| Consumer applications | Recurring subscriptions and engagement. | Direct global distribution and product feedback. | Serving cost, retention, abuse and product substitution. |
| Business workspaces | Managed seats and enterprise agreements. | Higher-value workflows and organisational deployment. | Security review, procurement, integration and measurable adoption. |
| API Platform | Model and tool consumption by external applications. | Many companies create use cases without OpenAI building each interface. | Developer switching, lifecycle changes, reliability and customer margins. |
| Agent execution | Longer tasks consume models, tools, containers and external services. | Greater share of a completed workflow. | Latency, verification, authorisation and variable compute intensity. |
| Media and realtime | Specialised generation, editing and communication workloads. | Extends beyond text into high-value modalities. | Compute, rights, provenance, safety and experience consistency. |
| Infrastructure commitments | Long-term access to training and inference capacity. | Supports frontier research and product availability. | Capital intensity, power delivery, utilisation and technology turnover. |
Core economic question. Can increasing capability and agentic task completion produce durable customer value faster than compute requirements, competition and product complexity increase? Public product adoption does not answer this without audited segment economics.
12
Safety and governance are product layers
| Control layer | Purpose | Where it acts | Residual boundary |
|---|---|---|---|
| Model behaviour | Shape instruction following, refusals and uncertainty. | Training, post-training and runtime safeguards. | No model behaviour is perfectly reliable across all contexts. |
| System policy | Define intended behaviour, instruction authority and restricted uses. | Product and API system layers. | Written policy requires implementation, monitoring and enforcement. |
| Tool permissions | Limit which side effects an agent may request. | Application, tool host, sandbox and approval interface. | A correct permission does not prove the action is useful or safe. |
| Access programmes | Restrict higher-risk specialist capability. | Daybreak approval and account provisioning. | Approval does not authorise activity outside the customer's legal scope. |
| Workspace governance | Control identity, sharing, plugins, models and audit. | Enterprise administration and compliance systems. | Organisation settings do not replace process or human accountability. |
| Evaluation and monitoring | Measure failures, regressions and operational behaviour. | Evals, tracing, safety checks, logs and incident response. | Tests cover sampled properties, not universal correctness. |
| Human authorisation | Approve consequential, destructive or externally visible actions. | Product confirmations and customer workflows. | Approval quality depends on clear information and manageable frequency. |
Authority rule. Model reasoning can recommend an action; policy and users authorise it; a tool or external system executes it; evidence and accountable review determine whether it succeeded.
13
Roadmap and dependencies
| Workstream | Current state | Published direction | Required gate | System consequence |
|---|---|---|---|---|
| General models | GPT-5.6 Sol, Terra and Luna with long context, configurable reasoning and tools. | Higher quality and efficiency across professional and agentic work. | Reliable evaluation, serving capacity and controlled lifecycle changes. | One family can cover more capability and throughput tiers. |
| Agent orchestration | Responses state, hosted tools, programmatic tool calling and multi-agent beta. | Longer, more composable task execution. | Authorisation, observability, recovery and end-to-end success measurement. | OpenAI becomes a runtime layer, not only a model endpoint. |
| Direct work product | ChatGPT combines conversation, files, browser, computer and connected data. | More repeatable and long-running workplace workflows. | Workspace governance, predictable behaviour and user trust. | Greater share of completed knowledge work. |
| Software engineering | Codex spans local, IDE, cloud and remote execution. | More parallel, delegated and automated engineering work. | Repository security, verification, review and dependable environments. | Distinct coding-agent platform; detailed in the companion article. |
| Realtime and media | Separate voice, transcription, image and video families. | Lower-latency multimodal creation and interaction. | Consistency, rights, provenance, latency and inference capacity. | OpenAI addresses communication and content workflows beyond text. |
| Open models | gpt-oss provides externally hostable reasoning weights. | Broader deployment and adaptation outside OpenAI hosting. | Ecosystem support, safety and competitive differentiation. | Complements rather than replaces the closed frontier service. |
| Infrastructure | Microsoft and Stargate-linked capacity routes support expansion. | Larger and more geographically distributed compute base. | Financing, construction, power, hardware, networking and utilisation. | Determines training pace and customer-serving headroom. |
| Safety and governance | Layered model, product, access and enterprise controls. | Capability-specific safeguards and governed deployment. | Evidence that controls remain effective as autonomy and tool access grow. | Controls become part of product architecture and market access. |
Strategic synthesis. OpenAI is evolving from a frontier-model laboratory into a vertically integrated AI platform spanning foundation and specialist models, consumer applications, enterprise software, developer infrastructure, agent runtimes and increasingly direct compute coordination. The durable-differentiation question is which integrated layers create compounding product, developer and research advantages. The structural-dependency question is which layers remain constrained by external capital, infrastructure, hardware, energy, distribution and customer systems. Success depends on the coherence of the complete chain, not the release pace of model names alone.
14
Key risks and unresolved questions
- Governance complexity. Foundation control, PBC duties, investor economics, board oversight and management authority can be oversimplified into a single ownership claim.
- Infrastructure intensity. Frontier training and high-volume inference depend on capital, power, data-centre delivery, networking, accelerators and utilisation.
- Partner concentration. Microsoft is an important investor, infrastructure partner and distributor; changes in contractual or technical alignment could affect several layers simultaneously.
- Model opacity. Closed frontier architecture, datasets, training compute and unit economics are not sufficiently disclosed for independent reconstruction.
- Lifecycle velocity. Rapid model, alias, tool and API changes create migration, evaluation and reproducibility costs for customers.
- Agent reliability. Longer tasks compound errors in context, planning, tool selection, execution and verification.
- Authority and security. Computer, shell, MCP and external tools expand the effect of mistakes, prompt injection or compromised credentials.
- Product overlap. ChatGPT, Codex, direct agents and the API can overlap in capability while retaining different controls, economics and release schedules.
- Competition and switching. Enterprises can combine direct OpenAI products, Azure distribution, competing models and open weights; adoption does not guarantee permanent platform dependence.
- Regulation and rights. Training data, generated media, safety, privacy, labour impact, market power and sector-specific rules create multiple legal boundaries.
- Economic disclosure. Public usage and financing announcements do not reveal audited product margins, infrastructure utilisation or cash-generation by segment.
- Roadmap uncertainty. Published direction is not a delivery guarantee; technical capability must pass safety, capacity, product and commercial gates.
15
Evidence ledger and primary sources
- OpenAI structureFoundation control, OpenAI Group PBC and governance design; accessed 24 August 2026.
- The next chapter of the Microsoft–OpenAI partnershipCorporate restructuring, economic interests and continuing strategic relationship.
- OpenAI model catalogueCurrent frontier, specialised, image, realtime, audio and transcription families; accessed 24 August 2026.
- OpenAI API changelogRelease dates and lifecycle events for GPT-5.4, GPT-5.5, GPT-5.6, realtime, tools and platform changes.
- Chat Latest modelRelationship between the changing ChatGPT Instant model and recommended production API models.
- ChatGPT documentationCurrent direct-product surfaces, workflows, capabilities and workspace administration.
- Responses APIModel response, state, tools and long-running workflow architecture.
- Agents SDKAgent definitions, orchestration, guardrails, state, tracing and sandbox agents.
- Computer useScreenshot/action loop, host execution and safety boundaries.
- Agent BuilderLegacy visual workflow system, export route and retirement status.
- GPT-5.6 Sol model cardModel ID, context, output, modalities, reasoning and tool support.
- Using GPT-5.6Family roles, persisted reasoning, programmatic tool calling, pro mode and multi-agent beta.
- GPT-5.6 CyberSpecialist model role, access and tool surface.
- GPT-Image-2Current image-generation and editing interface.
- GPT-Realtime-2.1Current realtime model interface and modalities.
- gpt-oss-120bOpen-weight licence, stored and active parameter counts and deployment boundary.
- Codex documentationProduct surfaces, environments, tools, extensions, permissions and integrations.
- OpenAI API deprecationsCurrent retirement and migration schedule for legacy products and interfaces.
- OpenAI toolsBuilt-in, custom, MCP and hosted execution categories.
- Safety best practicesApplication safety, human review, input constraints and monitoring.
- The Stargate ProjectInfrastructure partnership and intended US AI data-centre build-out.
- How OpenAI uses API dataRetention and data-control boundaries for developer and enterprise deployment.
Research cut. Facts were re-verified to 24 August 2026, 19:08 ICT. Sources are first-party OpenAI documentation and company disclosures. Company statements establish what OpenAI publishes; they are not independent verification of performance, economics or future delivery.