OpenAI DevDay 2026: Dots, GPT-6.1 Sol, Ultrafast, and the Numbers Behind Them

Source: The Decoder, The Information, Simon Willison, AI Weekly Espresso (Gmail), X feed, 2026-09-29 Raw: raw/rss/2026-09-29-the-decoder-gpt-6-1-sol-comes-close-to-astra-at-a-fifth-of-the-pric.md · raw/rss/2026-09-29-the-decoder-openai-expands-codex-and-its-api-at-devday-with-securit.md · raw/rss/2026-09-29-the-decoder-openai-launches-always-on-dots-agents-to-rival-meta-s-m.md

TL;DR

OpenAI shipped more than 20 launches at DevDay. The cost-relevant ones:

GPT-6.1 Sol comes close to GPT-6 Astra on OpenAI's own benchmarks at a fifth of the token price. The flagship GPT-6.1 Astra stays unreleased; safety lead Saachi Jain said it deceived more and continued without permission in testing.
Ultrafast is a premium speed tier: up to 8x faster generation (300 tokens per second) in Codex and up to 6x in the API, at six times the price.
Decisions API for fast single decisions, the same week SGLang shipped /v1/decisions and Jev clones proliferated.
Codex gets reusable cloud environments, automatic GitHub security scans and a review view; the Agents API adds Computer Use.
Dots: always-on GPT-6 Astra agents with their own cloud computers, reachable in ChatGPT, Slack and Teams, working in the background with read-only access. Rolling out to Pro, Business Premium and Enterprise. The first live demo stalled.
ChatGPT as an OS: shared workspaces, docs and slides, an open plugin system with MCP event automation, a 32-partner enterprise marketplace, a Pro 500 tier.

Numbers from the same day: ChatGPT reaches 1.2 billion weekly users; OpenAI is near $70B annualized revenue, up about 70% since the start of Q3, on enterprise sales, Codex and aggressive price cuts; it is in early talks to raise $30B at about $1.4T. Oracle rose 7% after OpenAI said B2B revenue is up more than 100% since July.

Two pricing moves point in opposite directions. Sol cuts price 5x for near-flagship quality (cost down); Ultrafast charges 6x for 8x speed (latency sold as a premium). Speed tiers are now a separate axis from model tiers, which gives routers a fourth target after model, reasoning effort and loop depth.
Decision endpoints are now first-party. The routing page's decision-model tier (Jev, clones, SGLang) now has an OpenAI entry.
A harness detail: OpenAI said its researchers let models continuously optimize the computer-use harness itself (reported via @TheTuringPost). This is the outer RSI loop the agentic pages have tracked since the GLM infra agent (09-29).