OpenAI DevDay 2026 for API developers: what shipped, what it costs, and what's still coming soon

OpenAI DevDay 2026 took place on September 29, 2026, at Fort Mason in San Francisco, with more than 20 announcements. For API developers, the ones that matter are GPT-6.1 Sol (gpt-6.1-sol) at $2 input and $10 output per million tokens, Ultrafast mode at up to 6x the speed for 6x the price (live on GPT-6 Astra only), computer use in the Agents API, a Decisions API in limited preview, MCP Events for webhook-driven plugins, and Sign in with ChatGPT.

This hub separates what you can call today from what’s in beta, preview, or “coming soon,” then links each theme to a deeper guide, starting with what GPT-6.1 Sol is and the Agents API walkthrough. Apidog keeps your key and model ID as environment variables, so rerunning requests against the new IDs is a one-field change.

DevDay 2026 status table

Status as of September 30, 2026, per OpenAI’s DevDay 2026 recap and the linked docs.

Models and speed: GPT-6.1 Sol and Ultrafast

GPT-6.1 Sol keeps GPT-6 Sol’s $2/$10 list price and halves cached input from $0.20 to $0.10 per million tokens. OpenAI pitches it as “near-Astra intelligence for a fifth of the price.” The model page lists a 1,050,000-token window, 922,000 max input, 128,000 max output, and an April 30, 2026 knowledge cutoff.

The benchmarks are OpenAI-run. On Terminal-Bench Science at max effort, 6.1 Sol costs $5.47 per task against $23.80 for Astra, though Astra still scores highest (68.1%) and OpenAI keeps it for the hardest science work. The launch post has the rest.

The migration catch is effort: 6.1 Sol accepts low, medium (default), high, xhigh, and max, with no none or minimal. The GPT-6.1 Sol API guide walks the swap. There’s no free route (Free and Go plans don’t get it, and there’s no free API tier); the free question covers the cheapest options. Claude Sonnet 5.5 shipped a day earlier at the same $2/$10, and the Sonnet 5.5 comparison explains why no shared benchmark exists.

Ultrafast is OpenAI’s premium speed tier: up to 8x faster token generation (300 tokens per second) in Codex and up to 6x in the API. Set service_tier: "ultrafast" on a Responses request; today it’s broadly available on gpt-6-astra only. Astra’s Ultrafast TPM runs from 500,000 (Tiers 1-3) to 5,000,000 (Tier 5). Processing is US data residency and global only, and OpenAI recommends WebSockets because HTTP overhead can eat the gain. Ultrafast mode, explained does the break-even math.

Agents: computer use, Bedrock, and the Decisions API

The Agents API entered public beta on September 10 to “build and run cloud agents with the Codex harness, fully managed by OpenAI.” DevDay added computer use. You create a session with POST /v1/agents/sessions and the OpenAI-Beta: agents=v1 header, then give the agent the computer_use tool and a hosted browser. The docs’ examples use gpt-6-astra.

The agent pauses with a computer_use_approval_request of two kinds: browser_origin_access for each new website origin, and browser_authentication for email, password, and verification codes (no passkeys or QR sign-in). Put this line from the docs in your threat model: “Origin approval does not enforce confirmation before individual actions.”

OpenAI charges “no additional fees” beyond tokens and tools. OpenAI-hosted sandboxes cost $0.03 (1 GB) to $0.48 (16 GB) per 20-minute session. Two hard limits: US data residency only, and no Zero Data Retention. Agents API vs Responses API vs Agents SDK helps you pick a layer, and computer use vs structured APIs covers when a browser is the wrong tool.

Bedrock Managed Agents, powered by OpenAI, runs the Agents API’s core capabilities natively in AWS. The Decisions API, in limited preview, focuses Luna on “a specific set of user-defined questions with finite pre-defined answers” to classify content, route requests, or pick an agent’s next action. A broad release is planned “in the coming days”; docs, pricing, and limits aren’t published yet.

Plugins and MCP Events

ChatGPT now supports the proposed MCP Events spec on all plans. Your MCP server advertises "events": {} in server/discover and handles events/list, events/subscribe, and events/unsubscribe under protocol version 2026-07-28 (MCP 2.0). Delivery is webhook-only: no polling or streaming.

Your server signs each delivery with Standard Webhooks HMAC (webhook-id, webhook-timestamp, webhook-signature), caps it at 256 KiB, and first sends a signed HTTPS verification challenge that ChatGPT echoes with a 2xx. Your server must also block private and local callback addresses. MCP Events explained builds and tests the sending side.

Plugin extensions (sidebar apps, panels, file viewers) reach all plans, alongside a Plugin Creator and a submission flow.

Identity, privacy, and billing

Sign in with ChatGPT is OAuth/OIDC identity, available globally. Your app receives the user’s name, email address, and profile picture, never their conversations, memories, or an API key. Plan usage is a separate layer: Plus and Pro users can spend their ChatGPT Work and Codex allowance in participating apps, including open-source and locally hosted tools, under a per-app weekly cap, with credits past the cap off by default. Launch partners include Cognition’s Devin, Notion, Vercel, T3, OpenClaw, and Dactyl. The Sign in with ChatGPT guide covers the flow.

ZDR with Private Safety Processing runs safety review without OpenAI retaining prompts or responses: encrypted records land in your own S3, Azure Blob, or GCS bucket on a 30-day lifecycle, reviewed by a hardware-attested runtime with no human access. It’s enabled per project. Private Inference follows as a preview “this fall.”

Codex: cloud, CLI, Code Review, Security Cloud

Codex cloud adds reusable environments for Plus, Pro, Business, Healthcare, Edu, and Enterprise. The refreshed CLI brings voice, an /agents view, prompt editing, resume, and worktrees to all plans, and the Codex changelog shows CLI 0.159.1 made GPT-6.1 Sol the default model on September 29. Code Review in the ChatGPT desktop app covers GitHub PRs and GitLab MRs with automatic cloud reviews on all plans. Codex Security Cloud runs scheduled repo scans, dedupes findings, and prepares fixes for Pro, Business, Enterprise, and Edu. Codex after DevDay 2026 has the details.

ChatGPT launches (not API features)

Several headline launches are ChatGPT features with no API surface. Dots are always-on agents powered by GPT-6 Astra, each with its own cloud computer, reachable in ChatGPT, Slack, and Teams, and rolling out to Pro and Business Premium. ChatGPT Space brings files and Pages together for collaborators. Team tasks (recurring or event-triggered work) and @ChatGPT in Slack and Teams are for Business and Enterprise. Your users will work there; your code can’t call them.

What to change in your code this week

1. Swap the model ID and fix effort. Replace gpt-6-sol with gpt-6.1-sol, move any "effort": "none" to "low", and rerun your evals.
2. Redo the cache math. Cached input is now $0.10 per million, half of GPT-6 Sol’s $0.20. Prompts over 272K input tokens bill the whole request at 2x input and cache rates and 1.5x output ($4 / $0.20 / $15). Prompt caching on GPT-6 explains how to keep hit rates high.
3. Try Ultrafast on Astra only. Send service_tier: "ultrafast" with gpt-6-astra over WebSockets, and confirm US or global processing fits your data rules.
4. Check Agents API limits first. US residency, no ZDR. Disable automatic retries on credential submissions; an approval 202 means accepted, not that sign-in succeeded.
5. Test your MCP Events sender before ChatGPT subscribes. Point it at a test receiver that checks the challenge flow and the HMAC signature (signature verification shows how), and confirm it refuses payloads over 256 KiB.

Here’s the migrated request:

curl https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6.1-sol",
    "reasoning": {"effort": "low"},
    "input": "Summarize this changelog entry in one sentence."
  }'

Run the before and after in Apidog. Save the request with Bearer {{OPENAI_API_KEY}} in the Authorization header and the model ID as an environment variable, then clone the environment so one run uses gpt-6-sol and the other gpt-6.1-sol. Assertions on the status code and the usage object flag any call that breaks, and the same scenario runs in CI through the Apidog CLI.

FAQ

When was OpenAI DevDay 2026? September 29, 2026, at Fort Mason in San Francisco. OpenAI counted more than 20 major announcements.

Is GPT-6.1 Sol free? No. It’s $2/$10 per million tokens on the API and ships only on paid ChatGPT plans. Is GPT-6.1 Sol free? lists the cheapest ways to run it.

Does GPT-6.1 Sol cost more than GPT-6 Sol? No. The list price is the same $2/$10. Cached input drops from $0.20 to $0.10 per million tokens.

Does Sign in with ChatGPT give my app the user’s API key? No. Your app gets name, email, and profile picture. Plus and Pro users can also spend plan usage in participating apps under a weekly cap.

Start with the model swap

The cheapest change on this list is the model ID: same list price, half-price cache, one effort level gone. Download Apidog, point one environment at gpt-6-sol and another at gpt-6.1-sol, and let saved assertions catch breakage before you ship.