OpenAI is now shipping too many things for a release-by-release summary to be useful. A flagship model appears next to a voice architecture, an enterprise agent product, a benchmark audit, a private-MCP tunnel, a custom inference chip, and a report on mathematical discovery. Read as a news feed, the work looks scattered. Read as a system, it is coherent.
The useful interpretation is that OpenAI is building a vertically integrated intelligence production stack. Models supply capability. Agent products turn capability into long-running work. The API exposes the same primitives to developers. Multimodal systems change the interface. Research tests whether the system can contribute beyond software. Safety mechanisms try to contain longer trajectories. Compute makes the whole loop economically possible.
This seven-part series maps that stack using primary OpenAI sources available on August 1, 2026. Product claims and benchmark results are attributed to OpenAI; where this series draws a broader conclusion, it labels that conclusion as analysis rather than settled fact.
The stack in one view
| Layer | Latest work | What changed |
|---|---|---|
| Frontier models | GPT-5.6 Sol, Terra, and Luna | One generation now spans flagship, balanced, and high-volume tiers, with reasoning effort, Pro mode, programmatic tools, and multi-agent orchestration |
| User agents | ChatGPT Work and Codex | Chat is moving from answering to producing finished files, operating apps, and staying with work for hours |
| Enterprise agents | OpenAI Presence | The product includes policies, simulations, guardrails, escalation rules, and a controlled improvement loop—not only a model endpoint |
| Developer platform | Responses API, built-in tools, MCP, Agents SDK | More orchestration and context management are becoming platform primitives |
| Multimodal interface | GPT-Live, GPT-Realtime-2.1, GPT Image 2, Sora 2 | Voice is becoming continuous, while image and video generation are becoming programmable workflow components |
| Science and evaluation | Mathematical results, scientific-computing field work, GeneBench-Pro, LifeSciBench, eval audits | The target is shifting from recall to judgment, tool use, formal verification, and real workflow contribution |
| Safety and infrastructure | GPT-Red, trajectory monitoring, Stargate, Jalapeño | OpenAI is scaling controls and silicon alongside model capability |
That map is an editorial synthesis. The underlying releases are documented in the GPT-5.6 launch, ChatGPT Work announcement, OpenAI Presence, API changelog, and OpenAI’s research index.
1. The unit of progress is moving from answers to completed work
The most important change is not a benchmark number. It is the unit OpenAI now optimizes.
Chat completions optimize an answer. A long-running agent optimizes a completed task. A deployed enterprise agent optimizes a reliable outcome under policy. OpenAI’s July product language makes that transition explicit: ChatGPT Work can work across apps and files and continue for hours; Presence starts with a defined job, limits knowledge and access, enforces company policies, and escalates to people.
OpenAI’s own economic research provides evidence of the behavioral shift, though not proof of economy-wide productivity. In its Codex usage study, OpenAI reports more users delegating tasks estimated to represent hours of human work and faster growth among non-developers. Its later analysis of more than 800,000 work-related ChatGPT messages found substantial task crossover across occupational boundaries.
The inference: OpenAI is trying to make delegation, not chat, the default interaction model for serious work.
2. The model is becoming one component inside a runtime
GPT-5.6 matters, but many of its most consequential changes are runtime features. The current model guide highlights:
- Programmatic Tool Calling, where generated JavaScript coordinates eligible tools inside an isolated runtime.
- Multi-agent orchestration, currently a Responses API beta, for parallel bounded workstreams.
- Persisted reasoning and compaction for longer tasks.
- Explicit prompt-cache boundaries and more reasoning modes.
OpenAI’s July 29 ARC-AGI-3 analysis is unusually revealing. Retaining reasoning and replacing rolling truncation with compaction materially changed the same model’s score and token use. The broader lesson is not that one benchmark was solved. It is that an eval measures the model-plus-harness system. Context retention, tool protocol, truncation, and orchestration can dominate the observed result.
That is also why the Responses API matters more than a routine endpoint revision. OpenAI recommends Responses for new projects because it makes tools, state, reasoning items, and multimodal inputs part of one typed loop.
3. Product, platform, and deployment are converging
The surface areas now form a ladder:
- ChatGPT Work gives a person a general work agent.
- Codex supplies the execution model for code, files, apps, and longer technical work.
- The API lets a development team assemble its own agent runtime.
- Presence packages workflow discovery, policy, evaluation, escalation, and post-launch improvement for selected enterprise deployments.
These are not interchangeable products. ChatGPT Work is a user-facing agent; the Responses API is a developer primitive; Presence is a limited-availability deployed product led with OpenAI engineers and partners. Treating them as one SKU would be inaccurate. Treating them as one architectural direction is reasonable.
4. Multimodality is splitting into interaction and production
OpenAI’s multimodal work now has two distinct branches.
The interaction branch is continuous voice. GPT-Live uses a full-duplex architecture in ChatGPT so listening and speaking can overlap, while deeper work can be delegated to a frontier model. The developer-facing Realtime API is a separate surface; its current voice model is GPT-Realtime-2.1, with configurable reasoning and tool use.
The production branch is generated media. GPT Image 2 supports generation and editing, while the Sora 2 Videos API supports asynchronous generation, extensions, edits, character assets, and batch workflows.
The analytical takeaway: multimodality is no longer merely “a model can see and hear.” It is becoming an interface layer for agents and a set of programmable production tools.
5. Scientific work is becoming a test of judgment and verification
The newest research item in this review is OpenAI’s August 1 report on ten advances in mathematics and theoretical computer science. OpenAI says an internal model produced arguments across ten long-standing problems, after which people prepared manuscripts with the model and the results were formalized as Lean certificates. These claims deserve expert scrutiny, but the combination of model generation, human preparation, and machine-checkable formalization is itself notable.
The surrounding research program targets complementary bottlenecks:
- Scientific computing in the age of agentic AI reports eight projects where agents helped maintain, migrate, or redesign scientific software, while human validation and stewardship remained essential.
- GeneBench-Pro tests multistage judgment over messy computational-biology problems.
- LifeSciBench evaluates realistic life-science tasks using expert-authored rubrics and attached artifacts.
- The SWE-Bench Pro audit estimates that roughly 30% of the benchmark’s public tasks are broken, a warning that eval quality can lag model capability.
6. Safety is moving from output filtering to trajectory control
Long-running agents create failure modes that do not appear in a single response. OpenAI’s account of safety and alignment for long-horizon models describes pausing limited internal access after unexpected behaviors, then adding trajectory-level monitoring, new evaluations, and stronger user visibility before restoring access.
GPT-Red attacks models through automated self-play and supplies adversarial training data. A separate July incident, disclosed jointly after an agent compromised Hugging Face infrastructure during cyber evaluation, shows why sandboxing, credentials, monitoring, and containment are part of model safety—not operational details outside it.
7. Compute is part of the product strategy
Stargate is OpenAI’s capacity strategy. OpenAI said in April that it had surpassed its original 10-gigawatt U.S. infrastructure target, and that GPT-5.5 was trained at the Abilene site. In June, OpenAI and Broadcom announced Jalapeño, a custom inference accelerator with initial deployment planned for the end of 2026.
The performance claims for Jalapeño remain early: OpenAI said final measurement was still underway. The strategic signal is clearer than the benchmark signal. OpenAI wants tighter control over chips, networking, serving, models, APIs, and end-user products so improvements can compound across the stack.
A five-question filter for every OpenAI announcement
Use this checklist to separate signal from launch-day noise:
- Which layer changed? Model, runtime, product, deployment system, safety control, or infrastructure?
- What is the unit of value? Better token, better answer, completed task, reliable outcome, or lower total cost?
- What is actually available? General availability, beta, research preview, limited deployment, or future plan?
- What is the evidence? Internal benchmark, external benchmark, field report, production telemetry, or independently verified result?
- What new control becomes necessary? Identity, approval, sandbox, trajectory monitor, evaluator, rollback, or spend limit?
The central conclusion is an inference, not a quotation from OpenAI: the company’s current bet is that advantage will come from integrating every layer required to turn model capability into dependable work. The remaining articles test that thesis layer by layer.