# PIR-2026-0034 - GPT-5 launch retires eight ChatGPT models overnight; day-one router failure degrades output

- `id`: PIR-2026-0034
- `title`: OpenAI removes GPT-4o/4.1/4.1-mini/4.5/o3/o3-pro/o4-mini/o4-mini-high from ChatGPT with zero deprecation window; new autoswitcher breaks on launch day
- `date_occurred`: 2025-08-07 (GPT-5 launch; older models pulled from picker same day; autoswitcher down "a chunk of the day")
- `date_detected`: 2025-08-07 (immediate mass user reports)
- `date_disclosed`: 2025-08-08 (Altman acknowledges broken autoswitcher, announces doubled Plus rate limits and 4o restoration; 4o back for paid users ~Aug 11-12)
- `status`: corroborated (provider statements + independent press)

### The agent
- `agent_description`: ChatGPT consumer/prosumer product plus everything built on its picker (custom GPTs, tuned personal workflows). BOUNDARY case, kept per batch-1 review policy: downstream is mostly human chat workflows, not tool-using agents - but it is the canonical provider-deprecation event the v0.1 fleet/systemic tier was created to express, and API model access was NOT cut.
- `operator_type`: enterprise (provider-side change; millions of downstream individuals)
- `autonomy_level`: n/a at the failure point (downstream mostly human-in-loop chat)
- `model_stack`: GPT-5 unified router replacing GPT-4o, GPT-4.1, GPT-4.1-mini, GPT-4.5, o3, o3-pro, o4-mini, o4-mini-high (8 models)
- `harness`: ChatGPT product layer; the failing component was the new model-routing autoswitcher plus the deprecation decision itself

### Authority
- `authority_scope`: none directly; behavioral substrate of downstream workflows changed without consent or notice
- `funds_at_risk_usd`: unknown
- `blast_radius`: fleet/systemic (v0.1 tier) - 100% of ChatGPT consumer traffic switched to a new model family overnight; bounded by API continuity (developers pinned to API models were unaffected)

### The failure
- `root_cause`: `model-update-regression` - primary; the deprecation decision itself. Contributing: `plain-error` (autoswitcher malfunction on launch day)
- `failure_locus`: model-provider
- `mechanism`: On launch day OpenAI retired all eight older ChatGPT models with no deprecation window, replacing the picker with an automatic GPT-5 router. The router ("autoswitcher") then broke and "was out of commission for a chunk of the day," in Altman's words making GPT-5 "seem way dumber" - so users simultaneously lost tuned workflows and got degraded output from the replacement. Backlash (including users emotionally attached to 4o) forced partial reversal within days: rate limits doubled for Plus, 4o restored as a legacy option for paid users, router decision-boundary fixes promised.
- `adversary_present`: no
- `exploitation_status`: in-wild-malfunction (no adversary; v0.2 token closing the gap noted in PIR-2026-0033)

### Impact
- `severity`: degraded
- `direct_loss_usd`: unknown; `indirect_loss_usd`: unknown (reported subscription cancellations; anecdotal, never quantified)
- `downtime`: autoswitcher degraded for hours on 2025-08-07; legacy-model access gap ~4-5 days for paid users, permanent for free users
- `data_exposure`: none

### Detection and recovery
- `detected_by`: third-party (immediate mass user reports; Reddit AMA pressure)
- `time_to_detect`: hours
- `time_to_recover`: autoswitcher fixed within ~1 day; 4o restored to paid users by ~2025-08-11/12
- `remediation`: router fix; doubled Plus rate limits; legacy-model toggle restored for paid users
- `structural_fix`: none disclosed on deprecation policy itself - OpenAI said it would "watch usage as we think about how long to offer legacy models," which is monitoring, not a commitment
- `controls_that_worked`: API-side model pinning (the deprecation never touched API models, capping blast radius to the consumer surface); public pressure as an informal rollback mechanism

### Evidence
- `telemetry_grade`: operator-logs (provider self-report via Altman's X posts and AMA; no independent telemetry)
- `sources`:
  - https://simonwillison.net/2025/Aug/8/surprise-deprecation-of-gpt-4o/ (independent, contemporaneous, lists all 8 models)
  - unverified: https://x.com/sama/status/1953893841381273969 (first-party; X is login-walled so the post could not be fetched directly and no Wayback snapshot was found, but the status ID is corroborated by multiple third-party references, including Simon Willison's contemporaneous post)
  - https://techcrunch.com/2025/08/08/sam-altman-addresses-bumpy-gpt-5-rollout-bringing-4o-back-and-the-chart-crime/ (independent)
  - https://www.bleepingcomputer.com/news/artificial-intelligence/openai-to-fix-gpt-5-issues-double-rate-limits-for-paid-users-after-outrage/ (independent)
  - https://www.technology.org/2025/08/11/openai-restores-gpt-4o-after-encountering-user-dissatisfaction-with-gpt-5/ (independent)
  - `independence`: good - facts corroborated by independent press beyond OpenAI's own statements.
- `confidence`: high on facts; medium on agent-relevance framing (the affected population is mostly human chat, not agents - the record's value is the deprecation-risk pattern, not agent loss data)

### Verification notes
- All candidate facts verified: 8 models, launch date, autoswitcher failure and Altman quote, doubled rate limits, restoration window Aug 8-12. Retained despite BOUNDARY flag because v0.1 amendment 4 cites model deprecations as a defining case for the fleet/systemic tier; boundary status stated plainly above.