OpenAI after DevDay 2026: known issues for serious development work

As of 30 September 2026.

OpenAI's DevDay 2026 on 29 September brought a genuinely interesting set of announcements: always-on Dots, ChatGPT Space and Pages, Codex Cloud, a refreshed Codex CLI and GPT-6.1 Sol at a fraction of Astra's price. Some of it is excellent. For serious development work, however, what happens behind the announcements matters more than the announcements themselves. The most serious problem is mundane: long agent runs fail, and both the work and the usage are gone.

I also work with Claude Code and OpenCode (running Opus 5.5 or GPT-5.6 Sol). They are not flawless either: the occasional bug comes with the territory. What I see from ChatGPT Work and Codex right now is of a different magnitude. Notably, GPT-5.6 Sol performs well through OpenCode, which suggests the problems lie in OpenAI's products and infrastructure rather than in its models. My impression is that OpenAI is shipping too much at once without keeping quality under control. In my assessment it still has a long way to go before its offering is enterprise-ready.

One gap stands out: there is no dedicated first-line support for Plus, Pro and Business users, including $500/month Pro 500 subscribers. When a multi-hour agent run fails and takes a large share of the weekly allowance with it, the official route is a help-center chat that starts with a bot. Only Enterprise customers get 24/7 support with SLAs.

All cases below are user reports on GitHub and the OpenAI Developer Community, not issues confirmed by OpenAI. Sources are listed in SOURCES.md. Corrections are welcome via issues, and resolved items will be logged in CHANGELOG.md.

Reliability issues, most serious first

1. Failed runs consume allowance with no recovery. Codex retries a dropped stream a few times, but once those retries fail there is no checkpoint to resume the run from, in Chat, Work or Codex. Interrupted work must be redone and the consumed usage is not automatically refunded. OpenAI did announce a worldwide usage reset at DevDay, but that was a one-off gesture rather than a fix. It makes no difference whether the run stops on a connection error or on a usage limit: the task has to start over. Examples:
   • Codex desktop: stream loss during context compaction burned 20% of a weekly allowance with no result (openai/codex#48685).
   • Codex desktop: a repair run consumed about 50% of a weekly allowance and hit the usage limit unfinished after 7h 15m (openai/codex#48188).
   • Codex desktop: a single task exhausted two full usage-reset credits with no accounting of why (openai/codex#47566).
   • ChatGPT Work: a run hung in RUNNING for over 10 hours and never resumed, not even after the quota reset (OpenAI Community, 18 Sep).
   • VS Code extension: a file was missing after a restart and remaining usage dropped from 70% to 0%, following a disconnect (openai/codex#36276).
2. MCP connections die in long sessions and cannot be restored. Custom MCP tools stop working after a while and the thread never recovers, even when the server is healthy. The only workaround is a new conversation, losing all context.
   • Chat and Work: in several reports no request even reaches the server, so the fault sits on the ChatGPT side (openai-apps-sdk-examples#241, #230, OpenAI Community, 25 Aug).
   • Codex: the MCP client has no auto-reconnect, and the desktop app drops MCP servers that work fine in the CLI (openai/codex#11489, #26347).
3. Stream disconnects and capacity errors across all surfaces.
   • "Selected model is at capacity" errors since 8 September (openai/codex#43738).
   • Codex CLI on Linux: frequent "stream disconnected" errors on a Pro plan (openai/codex#45949).
   • Codex desktop: reconnect loops where the CLI on the same machine works fine (openai/codex#37992).
   • ChatGPT Work: start and reconnect loops on both Windows and Linux (openai/codex#45099, #48482).
4. Tool-enabled Chat turns cut off after about 25 to 26 minutes since 20 August (OpenAI Community, 14 Sep).
5. Codex Cloud tasks go silent or report contradictory status (openai/codex#49505, #46174).
6. Computer Use in ChatGPT Work failed repeatedly for five days, and one troubleshooting session took about 40% of the allowance (openai/codex#49358).
7. Wasted usage on bad output: incorrect or incomplete work consumes allowance and every fix costs more (openai/codex#44455).
8. No per-task budget or token cap. Consumption is only visible after the fact.
9. No dedicated first-line support for Plus, Pro and Business. Only Enterprise includes 24/7 support with SLAs; everyone else, including $500/month Pro 500 subscribers, goes through a bot-first chat widget (OpenAI Help Center). Email support is unverified.

Limitations and plan changes

1. Codex Cloud is a delegation sandbox, not a dev machine: GitHub only, HTTP/HTTPS proxy with a domain allowlist, no interactive shell, SSH or computer use. Unsuitable for runtime, network or hardware-level work (docs).
2. Still no remote web access to Codex CLI. Local sessions can only be controlled from the mobile app or another desktop app, and no timeline has been announced (docs, openai/codex discussion #9200).
3. Remote requires the desktop app on macOS or Windows as host. Headless Linux CLI pairing needs an unofficial workaround.
4. ChatGPT Work is unavailable in projects with project-only memory, and shared projects are locked into that mode (release notes).
5. Pro 200 halved from 30 October: Work and Codex usage drops from 20x to 10x Plus (The Next Web).
6. Dots are unavailable on Pro in the EEA, Switzerland and the UK. Only Business Premium works there (OpenAI Help Center).
7. Sharing is limited: Space, Pages and Teams exclude Plus, and local Codex projects cannot be shared (DevDay recap).
8. Smaller gaps: separate memory stores for ChatGPT and local Codex, no strict zero data retention and no local hooks in Work Cloud with local access, and SSH remote projects limited to a single folder.

What would make the biggest difference

1. Resume interrupted runs, and refund usage lost to platform failures.
2. MCP connections that recover within an existing thread.
3. Per-task budgets and transparent usage accounting.
4. Dedicated technical support for paying professional users.