GPT-6.1 Sol arrives: near-Astra brains at one-fifth the price
OpenAI's DevDay opener is a mid-tier model it calls the most cost-efficient at comparable performance, with cached input discounted 95 percent off standard pricing.
Early this morning in San Francisco, OpenAI opened DevDay 2026 with GPT-6.1 Sol, a mid-tier model that sits just below its Astra flagship. The pitch is blunt: near-Astra intelligence for one-fifth the price, making it, in the company's words, the most cost-efficient model for its performance available today. Cached input is discounted 95 percent off standard input pricing, and Sol is aimed squarely at complex refactors, deep codebase investigations, and long-running agents that span apps. Sam Altman called it an extremely capable model, and early third-party tests on Higgsfield put the same 3D prototyping workflow roughly four times faster than Astra at a fifth of the cost.
Ultrafast: Codex runs up to 8x faster
OpenAI's new premium speed tier hits up to 300 tokens per second in Codex — an 8x speed-up — and up to 6x on the API side. It is bundled into the new Pro 500 plan, which drops the old five-hour usage cap.
Codex Security Cloud gets a major upgrade
The platform now integrates Daybreak Blue, a model with cyberattack and defense capabilities, by default. It scans entire GitHub repos, continuously reviews new commits, deduplicates findings, and prepares fixes for review — shifting security left to every commit.
Ultrafast is so fast. I do not ever want to go back. Sam Altman · CEO, OpenAI
dots: an always-on agent that never clocks out
dots is OpenAI's answer to what AI looks like when it stops waiting for a prompt. Each dot keeps its own cloud computer, remembers your preferences and standards, and keeps projects moving while you are away — then checks in by text, phone, or email. It ships first in Pro 100 with broader access to follow, and developers can hand off recurring triage, bug reports, and feature requests directly.
OpenAI shares safety practices for frontier RL training
A new post lays out practical guidelines for securing frontier reinforcement-learning training runs — how to contain an agent's reach before it earns it — reflecting what the lab says it has learned so far.
A roundup of every DevDay developer update
OpenAI's developer arm collected the day's announcements into a single thread: Codex, cloud environments, new APIs, and the tooling that surrounds them, all in one place for builders.
Decisions API enters preview
Define a question and its possible answers; the model picks one in about 150 ms to classify content, route requests, or choose an agent's next action.
Codex CLI gets a major refresh
A new full-screen interface, better readability, and ways to manage parallel work — all inside the terminal.
Codex cloud environments arrive
Reusable environments come pre-loaded with repo, dependencies, and settings, so agents keep working after you close the laptop.
ChatGPT opens native apps
Developers can now build full native apps with plugin extensions and ship them directly inside ChatGPT.
tldraw canvas lands in ChatGPT
A realtime multiplayer canvas is now available as a ChatGPT plugin.
Amp runs 6x GPT-6 Astra
Amp's Plaid speed mode now runs the high-speed build of GPT-6 Astra for when you're in a hurry.
GPT-6.1 Sol, now in Figma Make
Dev Mode also comes to Codex through the Figma plugin, tightening the design-to-code loop.
Runway joins the OpenAI Marketplace
Access top image, video, audio, and editing models, and apply part of your OpenAI commitment toward Runway purchases.
Cursor builds charts in chat
Type /visualize to analyze data and see the answer inline, now in the Agents window.
GPT-6.1 Sol is live in v0
Vercel's v0 brings Sol to full-stack web app generation with one-click deploy.
Replit lands on Meta Muse
Describe your app in conversation; the Agent builds it for Quest and Ray-Ban Display glasses.
Automations in Perplexity Computer
Ongoing work on event triggers or schedules, working with memory, skills, Slack, Gmail, and Linear.
AI Safety is an engineering problem. Aravind Srinivas · CEO, Perplexity
AMD to acquire World Labs for ~$8.2B
Fei-Fei Li joins AMD as executive vice president and chief scientist in the deal.
Hugging Face joins NVIDIA
Clement Delangue says the deal lets HF hire talent it couldn't as a startup, with a decade to make open-source AI win.
Anthropic tests Zhipu GLM-5.3
A report finds the model can write usable attack programs and evade simple safeguards.
Sonnet 5.5 beats Opus 5.5 in agentic coding
Scores four points higher at $2 per million tokens in and out.
Liquid AI announces d1
The first model to outscore Jev on Hugging Face's Decision Index.
IQuest-Q1 lands day-0 support
A 320B sparse MoE with 15B active per token, live in both SGLang and vLLM.
H/ Company releases Holo4
A generalist computer-use model family that can click, code, and call tools.
Agent governance is an engineering problem
Perplexity builds safeguards into its infrastructure, harnesses, and tools, then shares how.
Sol at 1/5 the price of Astra
Sam Altman flags the 95% cache-read discount on an extremely capable model.
Securing frontier RL training
Greg Brockman points to practical guidelines reflecting OpenAI's current learnings.
Profiles, skills, and connectors
New additions to the Perplexity Agent API: a saved agent under one versioned ID.
AI SDK hits 30M weekly downloads
Guillermo Rauch's milestone video was written in Rust and rendered on GPU.
Runway Agent adds Agent Tagging
Tag the Agent anywhere your assets live to request edits, options, and fixes.
Use your ChatGPT subscription in v0
Connect your account and build with Plus or Pro tokens.
What do you want from AI?
A new study with Anthropic Interviewer probes how people use AI and what role they want it to play.
Workera is acquired by Pearson
Andrew Ng calls it a privilege to support the team through the journey.
Ten build challenges
Finish five in 48 hours to enter the draw for a Codex Micro.
An open-world game for ~$2,000
Perplexity Computer rented its own GPUs and built a browser game end-to-end.