OpenAI DevDay 2026: what to expect
DevDay lands September 29 at Fort Mason in San Francisco. The efficiency tier already shipped — GPT-6 Sol and Luna went live September 22 — and pre-event reporting points to a security-focused GPT-6 Cyber preview plus a dozen or more other products. Here's what's confirmed, what's expected, and what it means for developers.
Confirmed vs. expected — read this first
OpenAI DevDay 2026 is scheduled for Tuesday, September 29, 2026 at Fort Mason, San Francisco, with the opening keynote livestreamed. Also confirmed: GPT-6 Sol and GPT-6 Luna, released September 22, 2026 — Sol at $2/$10 per million input/output tokens, Luna at $0.10/$0.50 — both live in the OpenAI API.
The GPT-6 Cyber security model preview — and reportedly a dozen or more other products — is expected per Fortune reporting (late September 2026), not confirmed by OpenAI. A limited group of customers is reportedly alpha-testing Cyber through OpenAI's Daybreak Red program. Names, pricing, availability, and capabilities can all change — or not materialize — at the keynote. Treat this page as a planning lens, not a product announcement.
A security model, in preview
The most interesting expected item is GPT-6 Cyber: a security-focused GPT-6 model expected to be shown as a preview. The expectation comes from Fortune reporting in late September 2026, which says the model could be unveiled at DevDay alongside a new product for deploying it securely and automatically — and that a limited group of customers in OpenAI's application-only Daybreak Red program is already alpha-testing it. If the reporting holds, this would be OpenAI's fourth cybersecurity-focused model of 2026, following GPT-5.4, GPT-5.5, and GPT-5.6 Cyber releases earlier in the year.
For developers, a dedicated security model matters in two places. First, review stages: a model tuned for security analysis is a natural fit for the verification step of an agent pipeline, where a general model writes the code and the security model audits it. Second, governed missions: MasterEngine's execution chain already separates authorization (deterministic policy, never a model) from execution — a security-tuned model slots cleanly into the evidence-and-verification stage without touching the authorization path. Preview means limited access and shifting behavior, so plan evaluations, not migrations, until it stabilizes.
Efficiency models priced for volume
This tier already shipped: OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026, sitting below the GPT-6 Astra flagship and live in the OpenAI API today.
| Model | Input / M | Output / M | Positioning |
|---|---|---|---|
| GPT-6 Sol | $2 | $10 | Efficiency tier |
| GPT-6 Luna | $0.10 | $0.50 | Ultra-efficiency tier |
The story here is cost per task at scale. Sol at $2/$10 undercuts GPT-6 Astra's $10/$50 by 5× on both tiers; Luna at $0.10/$0.50 is two orders of magnitude below Astra's input price. For high-volume agent loops — triage, classification, first-pass drafting, monitoring — where the task is simple and failures are cheap, efficiency models are where the economics of agents get interesting. The open question is the completion-rate side of the cost-per-task equation: a $0.10/M model that needs three retries isn't cheaper than a $2/M model that nails it first try. Benchmark the efficiency tiers on your simple tasks before routing volume through them.
What to do before and after the keynote
Before Sept 29
Build your evaluation queue, not your architecture. List the tasks you'd route to a security model (review, audit) and to efficiency models (triage, drafting, monitoring). Instrument cost-per-task measurement now, so new models get a fair comparison the day they land.
After the announcements
Evaluate on your tasks with real pricing, real rate limits, and real failure rates. Efficiency models get tested on volume tasks; the security preview gets tested on review stages. Commit budget only after your own numbers confirm the expectations.
The broader pattern to watch: the frontier is splitting into specialized tiers — security models, efficiency models, reasoning flagships — rather than one model for everything. That split rewards architectures that route by stage instead of standardizing on one model. It's the same posture our Astra vs. Opus comparison recommends, and it's how governed missions are designed to work: the contract bounds cost, and model choice follows the evidence per stage. We'll update this page after the keynote with what was actually announced.
Frequently asked questions
When is OpenAI DevDay 2026?
Tuesday, September 29, 2026, at Fort Mason in San Francisco. The opening keynote will be livestreamed. This page was published September 26 — three days before the event — so everything except the date and the already-released Sol/Luna pricing is pre-event expectation.
What is GPT-6 Cyber?
Per Fortune reporting in late September 2026, a security-focused GPT-6 model expected as a DevDay preview — aimed at security review, vulnerability analysis, and automated defensive workflows, with a companion product for secure deployment. Not officially announced; a limited group of customers is reportedly alpha-testing it through OpenAI's Daybreak Red program.
What are GPT-6 Sol and Luna?
Efficiency-tier GPT-6 models released September 22, 2026 and live in the OpenAI API: Sol at $2/$10 per million input/output tokens, Luna at $0.10/$0.50. Positioned for high-volume, low-complexity agent work.
Should I make decisions based on these expectations?
Plan, don't commit. Use expectations to build your evaluation queue and instrument cost-per-task measurement. Make purchasing and architecture decisions only after the official announcements and your own evaluations.
Will this page be updated after DevDay?
Yes. We'll revise it after September 29 with what was actually announced — confirmed details replace expectations, and anything that didn't materialize gets marked as such. The change will be recorded in the changelog.
Evaluate, don't assume
💰 Cost per task
The math you'll need when the efficiency tiers land: break-even, cache pricing, long sessions.
📖 Docs
The execution chain: how governed missions specify, authorize, and verify work.