OpenAI’s latest DevDay suggests the next phase of AI demand may be driven less by humans sending more prompts and increasingly by autonomous agents continuously consuming inference, software tools and cloud resources on their behalf.
OpenAI used DevDay 2026 to announce more than 20 products and platform updates spanning models, autonomous agents, Codex, APIs, workplace collaboration, plugins and enterprise partnerships.
The company also disclosed that ChatGPT has reached 1.2 billion weekly users, versus more than 800 million around DevDay 2025—a roughly 50% increase in the user base in about one year.
But we think the more important message for AI investors was elsewhere. The architecture of AI consumption is changing.
Until recently, most inference demand was relatively episodic:
Human → Prompt → Model → Response
What OpenAI demonstrated at DevDay increasingly looks like:
Human → Agent → Multiple models → Tools → Browser → Computer → Other agents → Monitoring → Repeated actions
And some of those loops can run continuously, without the user being present.
That distinction could have meaningful implications for the amount of inference consumed per user, the importance of latency, the value of caching and the infrastructure required to support AI applications.
The DevDay launch map
| Announcement | What OpenAI launched | Investor interpretation |
|---|---|---|
| Dots | Always-on GPT-6 Astra agents with their own cloud computers and access to 4,000+ connected apps | Potential transition from episodic inference to persistent AI workers |
| GPT-6.1 Sol | Near-Astra capabilities at roughly 1/5 of Astra’s standard input/output token prices | Lower inference cost can unlock substantially more agentic workloads |
| Ultrafast | Up to 300 tokens/sec and up to 8× faster generation in Codex | Creates a premium latency tier and suggests latency is becoming monetizable |
| Pro 500 | $500/month plan with 25× Plus usage allowance and Ultrafast access | Segments heavy AI users by compute consumption rather than just features |
| Decisions API | Luna-based low-latency API for classification, routing and finite-choice decisions | Targets the enormous volume of small decisions inside agent workflows |
| Agents API + Computer Use | Computer interaction, multi-agent capabilities, tool search/calling and context compaction | OpenAI increasingly provides the full runtime for autonomous applications |
| Codex Cloud | Persistent cloud development environments accessible from any device | Coding inference can continue even when the developer is away |
| Codex Security Cloud | Continuous repository scanning, deduplication and automated remediation | Another workload category that naturally becomes recurring rather than prompt-driven |
| ChatGPT Space / Pages / Slides | Collaborative productivity environment for humans and agents | Pushes ChatGPT toward becoming an AI-native workplace |
| Plugin extensions | Full applications can run inside ChatGPT | Turns ChatGPT distribution into a developer platform |
| Sign in with ChatGPT | Users can spend their ChatGPT allowance inside external applications | ChatGPT subscription begins to resemble an AI compute wallet |
| OpenAI Marketplace | Enterprise commitments can be spent on approved partner software | OpenAI is moving toward becoming a procurement/distribution layer for enterprise AI |
| Microsoft Agent 365 | Specialist Dots integrated with Microsoft’s enterprise governance stack | Deepens enterprise interoperability despite growing product overlap |
| AWS Bedrock Managed Agents | OpenAI-powered agents operating natively inside AWS | Expands OpenAI distribution beyond its own infrastructure and Microsoft ecosystem |
The biggest announcement may be Dots — because it changes the unit of consumption
Dots are OpenAI’s attempt to move the interaction model beyond ChatGPT. Each Dot is powered by GPT-6 Astra, receives its own cloud computer, can connect to more than 4,000 applications, can access browsers and other devices with permission, and can continue working toward objectives even while the user is doing something else.
That architecture matters! A conventional ChatGPT interaction might involve tens of seconds of compute initiated directly by a user. An autonomous agent could instead perform dozens or hundreds of intermediate operations:
research → browse → retrieve context → reason → call a tool → inspect result → reason again → execute → verify → report.
And OpenAI explicitly describes Dots as working 24/7.
For AI infrastructure investors, this potentially introduces a new consumption multiplier:
Users × agents per user × tasks per agent × actions per task × tokens per action
rather than simply:
Users × prompts per user × tokens per prompt.
The company is starting conservatively with one primary Dot, but OpenAI says it eventually expects users to add more Dots and increase either their speed or the total amount of work they can perform per month. That wording is particularly interesting because it points toward compute itself becoming an explicit product dimension.
OpenAI is effectively laying the commercial framework for selling AI labor capacity.
The monetization design also provides clues!
The first Dot is included with Pro or Business Premium. Ordinary conversations with the Dot do not consume normal ChatGPT limits, while jobs the Dot launches through Codex or ChatGPT Work do count toward usage limits. That looks like an adoption strategy:
make the agent relationship inexpensive → encourage delegation → monetize the expensive work generated by that delegation.
This resembles how cloud platforms historically reduced friction around software creation while monetizing the compute generated by the resulting applications.
GPT-6.1 Sol may matter more for demand than another frontier-model benchmark
GPT-6.1 Sol was marketed as offering near-Astra intelligence at one-fifth of Astra’s standard token pricing.
API pricing is:
OpenAI says cached input is 95% cheaper than standard input pricing. The important part here is not merely model competition. It is demand elasticity.
If capable agent inference becomes 5× cheaper, developers will not necessarily spend five times less. Historically, falling compute costs often unlock applications that were previously uneconomic.
Agents are particularly sensitive to this dynamic because they repeatedly reload context, inspect data, call tools and reason about intermediate results. The unusually aggressive cached-context price is therefore notable.
Persistent agents will repeatedly revisit company documents, repositories, user preferences, project history, emails, Slack conversations, database schemas and application state.
Cheap cached inference could materially lower the marginal cost of keeping agents continuously context-aware.
In other words: lower token prices may reduce cost per inference while increasing total inference volume.
That is an important distinction when thinking about long-term GPU and data-center demand.
OpenAI is beginning to segment AI compute like cloud infrastructure
Another revealing announcement was Ultrafast. OpenAI says Ultrafast can reach approximately 300 tokens per second, delivering up to 8× faster generation in Codex and up to 6× in the API. GPT-6 Astra Ultrafast is available immediately, with GPT-6.1 Sol support planned.
OpenAI simultaneously introduced Pro 500, priced at $500 per month and offering 25× the ChatGPT Plus usage allowance alongside Ultrafast access.
This is interesting because the model stack is increasingly being segmented across three dimensions:
Intelligence × latency × consumption allowance.
That looks increasingly similar to traditional cloud segmentation.
Some workloads optimize for lowest cost.
Others optimize for maximum intelligence.
And highly interactive agentic applications—coding, voice, computer use, robotics—may pay substantial premiums for latency.
For infrastructure providers, latency-sensitive AI can be particularly attractive because customers cannot always shift those workloads to slow batch inference.
Decisions API could quietly generate enormous inference volume
One of the less flashy launches could ultimately prove very important.
The new Decisions API uses GPT-6 Luna to answer predefined questions with finite possible answers. OpenAI positions it for functions such as classification, routing requests or choosing what action an agent should perform next. Think of a sophisticated AI agent. It may repeatedly need to decide:
- Does this email require action?
- Which tool should I call?
- Should I escalate this issue?
- Which workflow applies?
- Has the task succeeded?
- Should I continue?
- Which agent should handle the next step?
These decisions individually require little intelligence compared with a frontier-model reasoning task—but they occur constantly.
That suggests a model architecture where expensive models such as Astra handle difficult reasoning while inexpensive Luna-class models orchestrate potentially millions or billions of micro-decisions.
From an infrastructure perspective, tiny inference can still become enormous inference when multiplied by agent loops.
This reinforces OpenAI’s increasingly visible model-routing strategy:
- Astra → hardest intelligence
- Sol → capable everyday agentic work
- Luna → high-volume inference and orchestration
Rather than every application calling the largest model, OpenAI can optimize intelligence-per-dollar across the workflow.
Codex is becoming persistent cloud compute rather than simply a coding assistant
Codex received several strategically important upgrades.
Developers can now run Codex locally, remotely from a phone or entirely in the cloud. Cloud environments persist with project repositories, dependencies and permissions and can continue executing while the user’s computer is offline.
OpenAI also added:
- Code Review,
- multi-agent workflows,
- voice-controlled CLI,
- Codex Security Cloud,
- automated repository scanning,
- and cloud-based remediation.
- The common theme is again persistence.
AI coding demand is gradually moving from:
Help me write this function.
toward:
Take responsibility for this project, monitor it, review changes and fix problems.
That dramatically increases the addressable compute workload.
The economic unit becomes closer to developer-hours delegated to AI than questions asked to a chatbot.
OpenAI is building an agent runtime, not simply selling models
The Agents API is another important strategic piece.
It now incorporates computer use, multi-agent capabilities, tool search, tool calling and context compaction, with OpenAI operating the underlying infrastructure. This moves OpenAI farther up the stack.
The company increasingly provides:
Models → reasoning → memory/context → tools → computer interaction → orchestration → cloud runtime → applications.
For developers, this reduces the infrastructure required to build an agent.
For OpenAI, it potentially captures a larger portion of every agent’s economic stack.
And for cloud infrastructure, reducing developer friction generally expands the number of applications that reach production.
The AWS announcement deserves more attention
OpenAI and Amazon expanded their relationship around Bedrock Managed Agents powered by OpenAI.
These agents use OpenAI’s agent harness but operate inside AWS infrastructure. AWS states that the agents have their own identity, log their actions for auditing, run inside the customer’s environment, and execute inference through Amazon Bedrock without customer data leaving AWS.
Strategically, this is important. OpenAI appears increasingly willing to distribute its intelligence wherever enterprise workloads already live rather than requiring customers to move workloads onto a purely OpenAI-controlled stack.
That lowers enterprise adoption friction. It also suggests the AI infrastructure ecosystem may be more interconnected than the simplistic narrative of Microsoft versus AWS versus OpenAI implies.
AWS can monetize infrastructure. OpenAI can monetize intelligence and agents.
Enterprise customers retain AWS governance and existing cloud commitments.
And the combination potentially accelerates the number of production agents deployed.
Microsoft remains simultaneously partner and increasingly overlapping competitor
OpenAI also announced work with Microsoft to make specialist Dots manageable through Microsoft Agent 365’s enterprise security and governance controls.
That is classic coopetition. OpenAI is simultaneously moving directly into territory traditionally associated with Microsoft:
- documents,
- slides,
- meetings,
- team collaboration,
- enterprise search,
- workflow automation,
- software development.
Yet Microsoft remains an important distribution and governance layer. This suggests OpenAI is not necessarily attempting to replace every enterprise system.
Instead, its emerging architecture looks more like:
OpenAI intelligence + agents sitting across existing software systems.
That could eventually be more disruptive than replacing individual applications because the agent becomes the interface through which users interact with those applications.
ChatGPT itself is becoming a productivity operating system
ChatGPT Space may initially look like another collaborative workspace.
But combined with Pages, collaborative Slides, spreadsheets, Meetings, Sites, Plugins and shared team tasks, the broader direction becomes clearer. OpenAI is building a workspace where:
- humans create information,
- agents manipulate it,
- plugins connect external systems,
- automations monitor events,
- and teams collaborate around the resulting output.
This creates obvious overlap with Microsoft 365, Google Workspace and Notion.
But the competitive question may not simply be whether ChatGPT replaces Word or PowerPoint.
The more interesting question is:
Where does the user begin the workflow?
If a user begins by asking an agent to prepare an earnings deck and the agent subsequently pulls Salesforce data, builds an Excel-style model, creates charts, produces slides and distributes them through Slack, the AI orchestration layer may capture much of the user relationship even if existing SaaS applications remain underneath.
That could shift value away from individual application interfaces toward the agentic orchestration layer.
“Sign in with ChatGPT” could turn the subscription into an AI compute wallet
One of my favorite strategic announcements was surprisingly easy to overlook.
OpenAI now allows users to Sign in with ChatGPT and use their existing ChatGPT allowance across 16 participating products, including Devin, Notion, Vercel, T3, OpenClaw and Dactyl.
Conceptually, this could turn a ChatGPT subscription into something resembling an AI compute wallet.
Instead of subscribing separately to inference inside every AI application:
User buys OpenAI compute → application consumes that entitlement → developer gains distribution → OpenAI retains the compute relationship.
If broadly adopted, OpenAI could potentially establish itself as both an identity system and payment/compute layer for AI applications.
And it immediately gives developers potential access to 1.2 billion weekly ChatGPT users.
That is a powerful platform flywheel.
The Marketplace pushes OpenAI deeper into enterprise software economics
The OpenAI Marketplace is another strategically important development.
Eligible enterprise customers can apply portions of their existing OpenAI commitments toward approved third-party software.
Initial partners include companies such as:
- Adobe and Figma in creative software;
- Salesforce, ServiceNow, HubSpot, Sierra and Decagon in customer experience;
- Harvey and Legora in legal;
- Palo Alto Networks and CrowdStrike in cybersecurity;
- and Baseten in AI infrastructure.
This potentially changes what an OpenAI enterprise contract represents. Rather than purchasing only OpenAI models or ChatGPT seats, customers could increasingly purchase an AI ecosystem budget.
This is analogous to strategies successfully used by cloud marketplaces.
Once companies commit spending to a platform, allowing that commitment to fund third-party products makes the platform more valuable while making those commitments increasingly difficult to displace.
What DevDay implies for the AI consumption curve
Putting the announcements together, we think there are five structural demand multipliers emerging.
1. More users: ChatGPT has reached approximately 1.2 billion weekly users, up from more than 800 million around DevDay 2025.
2. More inference per user: Dots, Work, Codex and automations perform multiple model calls for every objective the user delegates.
3. More hours of consumption: Agents can continue running after users stop actively interacting with them.
4. More agents per user: OpenAI currently starts users with one Dot but explicitly anticipates multiple Dots over time.
5. AI spreading into external software: Plugins, Sign in with ChatGPT, Bedrock Managed Agents and the OpenAI Marketplace extend OpenAI inference beyond the ChatGPT interface.
This is why declining token prices should not automatically be interpreted negatively for infrastructure demand.
The relevant equation increasingly becomes:
Compute demand = price efficiency × application expansion × agent intensity × user growth.
If falling inference prices unlock substantially more agentic activity, aggregate compute spending can continue increasing even while cost per token falls sharply.
Read-through for the AI investment ecosystem
NVIDIA / AMD / custom silicon
- DevDay reinforces the argument that inference—not only training—can remain a powerful long-term accelerator demand driver.
- Persistent agents, computer use, coding agents, realtime models and Ultrafast inference all increase the importance of high-throughput, low-latency inference infrastructure.
Hyperscalers
- The AWS partnership suggests OpenAI intends to distribute workloads across existing enterprise cloud environments rather than exclusively through a vertically controlled OpenAI cloud.
- That supports a broader cloud infrastructure opportunity.
Microsoft
- The relationship remains strategically deep, but product overlap continues to increase.
- The key question is whether Microsoft ultimately captures economics primarily through infrastructure and enterprise governance while OpenAI captures increasing ownership of the user interface and agent layer.
Enterprise software
- Salesforce, ServiceNow, Adobe, Figma and others simultaneously appear as OpenAI partners and potential areas of product overlap.
- That reinforces our view that the near-term outcome may be augmentation and integration before outright application displacement.
- Over the longer term, however, agents could weaken the value of application-specific interfaces if users increasingly interact with enterprise systems through AI.
Cybersecurity
- Codex Security Cloud, Daybreak integration, Palo Alto and CrowdStrike participation in the Marketplace suggest cybersecurity is becoming another major autonomous-agent workload.
- Security monitoring is inherently continuous—which makes it particularly compute-intensive if AI agents increasingly participate in the loop.
The most important KPI is no longer simply “tokens”
One notable difference from previous AI events is how difficult it is becoming to measure demand using simple metrics.
At DevDay 2025, OpenAI highlighted approximately 6 billion API tokens per minute.
At DevDay 2026, the product story is increasingly about autonomous work.
That means investors may eventually need a different set of indicators:
| KPI to watch | Why it matters |
|---|---|
| Active Dots | Measures penetration of persistent agents |
| Agent tasks per user | Better proxy for autonomous AI adoption than chats |
| Average agent runtime | Directly connected to compute intensity |
| Tool calls per task | Measures complexity of agent workflows |
| Work/Codex usage per subscriber | Helps estimate inference intensity |
| Pro 500 penetration | Indicates willingness to pay for additional compute |
| Ultrafast utilization | Measures value of latency-sensitive AI |
| Number of Dots per customer | Potential future compute multiplier |
| Specialist Dots deployed | Indicates enterprise AI-labor adoption |
| Marketplace spend | Measures OpenAI ecosystem leverage |
| External ChatGPT allowance consumption | Shows how much inference migrates outside ChatGPT |
| Cached-token share | Important indicator for agent unit economics |
Our Take
OpenAI DevDay 2026 was ostensibly a developer conference.
From an investment perspective, however, it looked increasingly like the unveiling of an AI economic stack.
OpenAI now has:
- frontier intelligence,
- low-cost models,
- premium low-latency inference,
- persistent agents,
- computer-use infrastructure,
- cloud coding agents,
- enterprise collaboration software,
- developer APIs,
- identity,
- plugins,
- a marketplace,
- and distribution to 1.2 billion weekly users.
The company appears to be converging toward four strategic layers:
- Intelligence — GPT models
- Labor — Dots, Codex and agents
- Workspace — ChatGPT, Space, Pages and Sites
- Distribution — Plugins, Marketplace, Microsoft, AWS and Sign in with ChatGPT
For AI infrastructure investors, however, the most important development may be even simpler:
AI usage is evolving from something humans actively initiate into something software performs continuously.
A chatbot waits for a question. An agent watches, reasons, acts, checks and starts again.
If that model scales from one agent per user to several agents per worker—and eventually specialist agents across entire enterprises—the inference opportunity could become substantially larger than one would estimate by extrapolating today’s chatbot token consumption.
At the same time, GPT-6.1 Sol demonstrates the other side of the equation: inference prices continue falling extremely quickly.
So the fundamental debate is no longer simply “Will AI become cheaper?”
It clearly will.
The more important question for the AI infrastructure cycle is:
Will consumption grow faster than the cost of intelligence falls?
DevDay 2026 provided several reasons to believe OpenAI is designing its entire product roadmap around making that happen!
Thank you!
CoreValue Research Team
