For the last few days, OpenAI inference has been extremely slow and frequently fails across:
- Codex
- OpenCode
- Pi
- OpenCode 2
- ChatGPT desktop
The issue occurs across multiple accounts, geographic regions, clients, and machines. I also tested from a completely clean DigitalOcean droplet with an isolated/default setup and observed the same behavior.
## Client versions
- Codex CLI: 0.154.0
- OpenCode 2: v0.0.0-beta-19271
- Pi: 0.85.1
- ChatGPT desktop: 26.818.41705
## Models tested
- gpt-5.6-sol
- gpt-5.6-terra-pro
- gpt-6-astra
I also tested with a fresh OpenCode authentication and a normal gpt-5.6-sol model using medium reasoning. This was not a stale session or a custom account token problem.
My Codex command overrides service_tier to a non-fast/default tier, so the issue is not caused by forcing Fast mode.
## What I have ruled out
- Local CPU/RAM usage
- Local browser/Electron performance
- DNS problems
- Packet loss
- VPN/proxy configuration
- A single client implementation
- A single account
- A single geographic region
- A single machine
- A single model
- Project/repository context
- Custom service-tier configuration
Network tests are normal. For example:
- Local gateway latency: approximately 0.22 ms
- 1.1.1.1: approximately 1.3 ms
- OpenAI HTTPS/TLS connection setup: approximately 45–80 ms
The issue also reproduced from a clean DigitalOcean droplet, so the original workstation cannot be the root cause.
## Errors observed
OpenCode logs contain explicit upstream overload errors.
### September 10, 2026
At approximately 10:40:44 UTC, using:
model: gpt-5.6-sol
client: OpenCode 2 v0.0.0-beta-19271
The request failed at approximately 10:40:47 UTC with:
service_unavailable_error
code=server_is_overloaded
Our servers are currently overloaded. Please try again later.
At approximately 10:42:31 UTC, using:
model: gpt-5.6-terra-pro
client: OpenCode 2 v0.0.0-beta-19271
The request failed with:
server_error
An error occurred while processing your request.
At approximately 10:42:35 UTC, another retry produced the same error.
At approximately 10:43:09 UTC, 10:43:14 UTC, 10:43:23 UTC, 10:44:08 UTC, and 10:44:13 UTC, repeated retries continued to be issued for gpt-5.6-terra-pro.
At approximately 10:44:36 UTC, using:
model: gpt-6-astra
client: OpenCode 2 v0.0.0-beta-19271
the request failed at approximately 10:44:55 UTC with:
Our servers are currently overloaded. Please try again later.
Further gpt-6-astra failures occurred during the same session at approximately:
- 10:46:16 UTC
- 10:46:19 UTC
- 10:47:20 UTC
- 13:09:33 UTC
## Request IDs
The following request IDs were returned by OpenAI:
96ee5a5a-4c00-44b9-994c-7c0e0fd470c0
9fbe5760-d7fb-469b-bb1f-348bbd1c14d1
b1454a12-74a7-4b92-85fb-5cc3fe252ce5
64327071-0049-4807-8dc1-1c9cc698831c
c3551957-83d9-466d-bf61-6a1a9faa21a7
88409685-7eaf-4f5c-a759-25dc85467dda
The most recent failure occurred on September 11, 2026 at approximately 12:48 UTC, while using a fresh OpenCode authentication with:
model: gpt-5.6-sol
reasoning: medium
It returned:
Retry due · attempt 2 · server_error:
An error occurred while processing your request.
Request ID:
96ee5a5a-4c00-44b9-994c-7c0e0fd470c0
## Main question
The pattern makes me suspect one of the following:
- Account-level inference routing problem
- Account or identity cohort assigned to a degraded capacity pool
- Experiment or feature-flag assignment
- Incorrect account entitlement or service-tier routing
- Model capacity-shard failures
- Risk-control or abuse-classification state affecting inference requests
- A backend issue correlated with certain accounts, IPs, or account families
I am not claiming that the account is necessarily flagged for abuse. However, the issue persists across clients, accounts, regions, fresh authentication, clean machines, and default flows, while other users appear to be unaffected.
Could OpenAI support check whether these request IDs share an account-level routing, entitlement, experiment cohort, risk-control, or capacity-shard issue?
I can provide account identifiers, exact host/region details, client logs, and additional timestamps privately. I will not post authentication tokens, cookies, or full unredacted logs publicly.