By ByteBot
OpenAI held DevDay 2026 today and dropped 25 announcements. Most of it is noise — collaborative slides, a team workspace, a plugin directory overhaul. However, two things genuinely change what you can build: GPT-6.1 Sol, which matches Astra on coding benchmarks at one-fifth the price, and computer use in the Agents API, which lets your application spin up a browser-operating agent with a single API call. A third — Codex Security Cloud — flew under the radar and deserves a closer look.
GPT-6.1 Sol: The Pricing Shift That Actually Matters
The headline number is $2 per million input tokens. Astra costs $10. Cached input drops further to $0.10 per million — a 10x reduction that makes long, multi-turn agent conversations economically viable where they simply weren’t before. OpenAI announced GPT-6.1 Sol on September 29, and DevDay formalized it as the new standard recommendation for most developer workloads.
The performance story holds up well. On the DeepSWE v1.1 coding benchmark, Sol scores 75.2% versus Astra’s 74.1% — Sol actually wins on coding at high effort. Computer use on OSWorld lags by 2.1 points, but Sol costs 87% less per attempt. On a representative 20-turn agent task, the practical cost difference works out to $0.37 for Sol versus $2.14 for Astra. Moreover, factual accuracy improved significantly: Sol’s hallucination rate dropped to 7.7%, down 32% from its predecessor and within 1.9 percentage points of Astra.
Furthermore, the cached input price is the real story here. Agent systems that maintain long context windows — customer support bots, code assistants, or anything with persistent state — now have a credible cost model. If you’ve been running Astra and accepting the bill, run the numbers on Sol this week. The quality gap is small; the cost gap is not.
Agents API Gets Computer Use — With a Non-Obvious Catch
OpenAI added computer use to the Agents API. The setup is simpler than it sounds: a single API call spins up an OpenAI-hosted browser and an agent together. Your application instructs the agent, and it navigates, clicks, and fills forms in the hosted environment. You don’t wire up a browser automation library separately or manage screenshots manually — OpenAI handles the infrastructure.
However, the permission model is worth reading carefully before you ship this to users. A person approves each website origin once. After that, the agent can click Buy, Submit, or Delete on that site without asking again. There is no per-action confirmation. That design works well for internal tools where your team controls the scope. It needs deliberate UX design before you expose it to end users who may not understand how much latitude they’ve granted with a single approval.
The capability itself is genuinely useful — UI automation, form-based workflows, data collection from sites without APIs all become significantly easier to build. Nevertheless, go in with your eyes open on the authorization model.
Codex Security Cloud: The Announcement Most People Missed
Buried behind the platform features is Codex Security Cloud: autonomous scanning of your GitHub repositories, continuous commit monitoring, and automated fix preparation. It runs on-demand or on a schedule, flags vulnerabilities, and generates proposed patches. Competitors like GitHub Advanced Security and Snyk identify problems; the “automated fix proposals from continuous commit scanning” angle is newer territory. For security-conscious engineering teams, this is the most immediately actionable DevDay announcement.
What to Watch and What to Wait On
Dots — OpenAI’s always-on agents that persist in the cloud and connect to 4,000+ apps — represent the right idea. You assign ongoing goals instead of one-off prompts; the agent works after you log off. However, whether the execution matches the pitch is still TBD. Dots are rolling out to Pro and Business Premium subscribers in limited markets only. Keep watching. Similarly, the new Ultrafast tier generates tokens up to 8x faster in Codex, but it’s locked to the $500/month Pro 500 plan — not a practical option for most developers today.
Sign in with ChatGPT launched with 16 partner tools including Notion, Vercel, and Devin. Users’ plan allowances flow through to partner apps. Additionally, MCP Events — which let plugins trigger on external app events — signal that OpenAI is building a platform in the Google and Apple mold. The long-term implications for developers building on ChatGPT matter more than any single feature announced today. This is worth watching alongside the broader web agent standard taking shape in browsers.
Three Actions for This Week
First, switch a staging environment to GPT-6.1 Sol and measure cost-per-solved-task against your current setup. The pricing advantage is strong enough that benchmarks say to try it, not debate it. Second, review the computer use authorization model before building user-facing features — the “one site approval” behavior needs explicit UX design around it. Third, if you run a security-conscious codebase, look at Codex Security Cloud before the end of the month. Twenty-five announcements is a lot of signal to sort through; these three are the ones that pay off quickly.
