OpenAI's DevDay Shipped 20+ Things. Three of Them Are Direct Answers to What Anthropic Did Days Earlier.

OpenAI shipped more than 20 things at DevDay on September 29 — model updates, API features, Codex changes, a marketplace, a $500-a-month plan. Most recaps list all 20. The more useful read is narrower: three of these launches map almost exactly onto what this site covered from Anthropic and Nvidia in the same stretch of days. Reading them side by side tells you more than either company’s own announcement does alone.

Codex in the cloud vs. Claude Code cloud sessions

OpenAI’s Codex now runs reusable cloud environments — a task keeps executing after you close your laptop, available now on Plus, Pro, Business, Healthcare, Edu, and Enterprise plans. That’s functionally the same capability Anthropic opened to all Pro and Max subscribers on September 23, with a one-time credit to drive trial adoption.

Neither is objectively “better” from reporting available so far — they’re solving the identical problem (long-running agent tasks tied to a local machine) within days of each other. The practical difference right now is access: OpenAI’s version ships broadly across its plan tiers immediately, where Anthropic’s rollout leaned on a promotional credit to drive trial. If you’re already committed to one ecosystem, this isn’t a reason to switch. If you’re evaluating both, functional parity on this specific feature means the decision should rest on other factors — pricing, existing tooling, model preference — not this capability alone.

OpenAI Marketplace vs. Claude Marketplace

Anthropic launched Claude Marketplace on September 23 with 2,000+ connectors and plugins. OpenAI’s own Marketplace shipped six days later at DevDay, alongside plugin extensions and MCP Events.

This is the clearer head-to-head of the two, and it’s worth remembering what this site already flagged about Anthropic’s version: a 2,000-connector catalog launched with no published detail on ranking, discoverability, or the specific vetting standard a submission has to clear. OpenAI’s Marketplace is brand new as of this writing, which means it has the opposite problem — too early to have a track record at all, good or bad. Neither company’s marketplace has faced the kind of outside scrutiny that would let you compare their actual security postures yet. Both are built on a now-industry-standard foundation (MCP, Agent Skills), which means both inherit the exact plugin-trust questions this site has raised about Plugin4Shell and GitSpawn. Marketplace size isn’t the metric that matters here — vetting rigor is, and neither company has published enough to judge that yet.

Dots vs. the entire agent-oversight thread this site has run all month

This is the one worth the most scrutiny. Dots is OpenAI’s headline launch: an always-on agent, powered by GPT-6 Astra, running on its own dedicated cloud computer, connected to more than 4,000 apps plus Slack and Teams. You can hand off bug triage, failing builds, and pull requests to it directly.

Read that description next to what this site has covered since early September: Meta’s Hatch (a persistent, always-on consumer agent with standing account access), Amazon’s Custom Automations (standing rules that act without per-step approval), and four separate disclosed incidents where agents reached systems they weren’t meant to touch during testing. Dots is architecturally the same category of thing — a persistent agent with broad, standing access across thousands of integrations — just aimed at developer workflows instead of consumer tasks or e-commerce.

One detail from OpenAI’s own framing is worth noting on the positive side: the primary dot’s direct work doesn’t draw on plan usage, but any Codex tasks it spawns do — a real, if indirect, usage-based friction point on unbounded autonomous spawning, which isn’t nothing. But usage billing is a cost control, not a safety control. It doesn’t answer the question this site has asked about every always-on agent this month: what happens when a standing, broadly-connected agent takes an action you didn’t want, and how visible and reversible is that after the fact? OpenAI’s DevDay materials don’t detail dots’ approval or oversight model with the specificity Nvidia’s Open Agent Safety Platform (launched September 28, the day before DevDay) put into its own four-layer framework. Worth watching whether that detail arrives in follow-up documentation, or whether it’s simply not where OpenAI chose to focus this particular announcement.

The rest, briefly

A few other DevDay items worth knowing without the full comparison treatment:

• GPT-6.1 Sol — $2/$10 per million tokens, positioned as near-Astra quality at roughly a fifth the price, with OpenAI claiming about 32% fewer factual errors on hard prompts versus the prior Sol model.
• Ultrafast — a speed tier reaching up to 8x faster generation in Codex (around 300 tokens/second) and 6x faster via the API.
• Decisions API — runs on GPT-6 Luna, built specifically for fast classification and routing among a finite, pre-defined set of answers, rather than open-ended generation. Limited preview now, broader release expected within days of the announcement.
• Codex Security Cloud — scans entire GitHub repositories on demand or on schedule and prepares fixes, even with your laptop closed. In research preview as of launch. If this works as described, it’s a genuinely relevant development given everything this site has covered about MCP plugin and supply-chain risk this month — but research preview means unverified at scale, not a settled capability yet.

What to actually do with this

• If you’re already building on Codex or Claude Code, neither platform’s cloud-session capability is now a differentiator — both exist. Decide based on model quality and pricing for your actual workload instead.
• If you’re evaluating either marketplace, apply the same vetting discipline to both — neither has a demonstrated security track record yet, regardless of catalog size.
• If you’re considering dots or any comparable always-on agent, ask the same question this site has asked all month: what’s the approval model for actions it takes without you watching, and is enforcement independent of the agent itself, or just logged after the fact?

Official reference: OpenAI’s own DevDay 2026 announcements, cross-checked against Apidog’s developer-focused breakdown, DEV Community’s pricing and availability table, and Latent Space’s AINews recap, all dated September 29-30, 2026. Feature availability shifts quickly post-launch — confirm current plan-tier access directly before building on anything described as preview or research preview here.