Date: 2026-09-29 · Status: sanitized (see the Sanitization note) Method: direct probing of the private chatgpt.com/backend-api/codex/* endpoints, corroborated by the official openai/codex source (commit 6a39914e) and public issue reports.

Three independently established findings together form a complete picture of "new model, old substance":

1. Mislabeling (measured). Models that shipped under the names gpt-6-sol and gpt-6-luna — the "GPT-6 family" — carry a GPT-5 generation fingerprint in their instruction templates. The upstream then hard-renamed them to gpt-5.6-sol / gpt-5.6-luna; the catalog descriptions now literally read "Older coding model (formerly gpt-6-sol)", which is an implicit admission that the "6" branding never matched the generation.
2. Silent substitution (measured mechanism + community corroboration). A request for the flagship gpt-6-astra comes back with response headers unilaterally declaring a safety-buffering fallback chain whose terminal node is gpt-5.6-luna — a 5-series "older fast model". The substitution event surfaces only as response.model being rewritten inside the stream; no client-side knob can disable it, and the official source confirms the consequence is a once-per-turn warning — not a refusal, not a retry, not user-facing disclosure.
3. A hidden budget tier (measured). The catalog contains gpt-reserve (visibility: hide, "Fast and affordable agentic coding") — also 5-series at the base. A model whose name carries no generation number is in fact the 5-series cheap tier, invisible in the client list by default.

In one sentence: of the seven slugs on sale, only gpt-6-astra is a real GPT-6; every other "6-family" name is a 5-series rebrand — and even the one genuine 6 can be silently swapped for a 5-series luna entirely at the upstream's discretion.

Source: GET /backend-api/codex/models?client_version=0.154.0 (measured 2026-09-29; requires Authorization + Chatgpt-Account-Id).

Generation fingerprints come from each slug's instructions_template and internal fields:

• comp_hash: 3000 across the new family vs 2911 for gpt-5.5 — an instruction/compatibility generation marker; the 5.6 series shares astra's hash but sits on a 5-series instruction base.
• multi_agent_version: astra/sol/terra = v2; reserve/luna/auto-review = v1.
• use_responses_lite: true across the new family, false for gpt-5.5; tool_mode likewise — code_mode_only for the new family, null for gpt-5.5.
• supported_reasoning_levels: astra and sol/terra reach ultra; the rest stop at max.

Measured minimal_client_version gating: 0.153.0+ sees all seven; 0.144–0.152.x sees six (no astra); older versions narrow stepwise to none. No client version can ever see gpt-6-sol / gpt-6-luna — old clients are not shown the old names, they are locked out of the new catalog entirely. Dead names return a fatal 400 with no alias rewriting.

gpt-5.5 carries retirement_at=2026-10-14, and its official migration target is gpt-5.6-sol — i.e., the model once sold under a "6" name is the designated successor for the genuinely old 5.5. The naming system internally admits the 5.6 series belongs to the 5-series generation.

Sources: direct measurement 2026-09-22/27 + official codex-rs source (commit 6a39914e).

Responses carry x-codex-safety-buffering-enabled: true and x-codex-safety-buffering-faster-model: <model>; the measured map:

Every chain ends at a luna: whether the request asks for flagship or balanced, degradation converges on the "older fast model". The headers only announce eligibility; the event itself is response.model in the SSE payload declaring the fallback (the official client separately reads the openai-model / x-openai-model headers as the served-model signal).

Roughly fifty probes on a healthy account covered: 7 models × 6 gateway nodes (unified-125/95/88/157/97/179, steered via replayed routing-cookie pairs) × varied payloads × tools × high/xhigh effort × serial and parallel bursts × security-review prompts. Every request-side lever was tested:

• x-codex-safety-buffering-enabled: false as a request header: no effect, the chain is still declared;
• service_tier: default/priority accepted but the headers persist; auto/flex rejected outright with 400;
• store:true: rejected (Store must be set to false);
• x-codex-routing-hint: model=X: a pure hint, does not alter routing;
• previous_response_id: rejected (Unsupported parameter) — continuations ride on the turn-state header alone, and there is no mid-turn model escape.

• response_model() reads only the openai-model/x-openai-model headers; on mismatch the consequence is a once-per-turn ModelReroute warning — output is delivered as usual.
• Model comparison is to_ascii_lowercase only, with no snapshot normalization: gpt-6-astra-2026-09-15 ≠ gpt-6-astra.
• retry_model/safety_buffering is a candidate retry model semantic, not evidence a swap already happened; but turn/steer input inherits whatever model the original request negotiated — a turn that opened on a buffered node keeps feeding user input into the degraded service, with no mid-turn escape.
• No strict model-selection contract exists anywhere in the official catalog or protocol.

• openai/codex#47015 (2026-09-21): on the OAuth subscription route, a gpt-6-astra request returned gpt-5.6-luna metadata; the community formally demanded a strict model-selection contract — none exists.
• openai/codex#47765 (2026-09-24): the "Codex severely dumbed down" thread is still active.
• icoretech/codex-pooler#425: cites a Reddit case where two of five Pro accounts were persistently degraded — degradation can stick per account.
• wangyunjeff/sub2api-state-kit#16: independent measurement — without a pinned routing-cookie pair the request comes back served by Luna, while the same request with the pair returns Astra. Default routing lands a flagship request on luna; which generation the user actually gets is decided by routing credentials, not by the model name.

• available_access_programs carries only cyber: ["standard"] per model — "cyber" is an access-program codename, not a model name; gpt-cyber / gpt-6-cyber requests all return 400. Generation numbers, tier labels, and program codenames are interleaved in one namespace, so the real generation cannot be inferred from a name.
• An extra priority (Fast) tier is advertised: 2x billing for astra, 1.5x for the rest — the premium tier rides on the same silently substitutable service.

# Catalog enumeration (any valid ChatGPT-login credential)
curl -sS "https://chatgpt.com/backend-api/codex/models?client_version=0.154.0" \
  -H "Authorization: Bearer $ACCESS_TOKEN" \
  -H "Chatgpt-Account-Id: $ACCOUNT_ID" \
  -H "User-Agent: codex-tui/0.154.0" | jq -r '.models[].slug'

# Fallback declaration: send any legal streaming request to
# /backend-api/codex/responses and observe
# x-codex-safety-buffering-{enabled,faster-model} and whether
# response.created.response.model is rewritten mid-stream.

• Test accounts are referred to by plan tier only (prolite / self_serve_business_prolite / pro / team); emails, credential IDs, phone numbers, and Apple relay addresses are removed;
• no access_token / id_token / refresh_token / account_id / session UUID appears anywhere — $ACCESS_TOKEN / $ACCOUNT_ID are placeholders;
• gateway node IDs (unified-N) and edge colo codes are upstream infrastructure identifiers that do not identify any user; they are kept for reproducibility.