Charge xAI's real rate for the slugs it retired on 2026-05-15

[adambalogh]

Found while syncing the SDK's TEE_LLM enum against this registry (OpenGradient/OpenGradient-SDK#309).

The bug

xAI retired grok-4, grok-4-fast, grok-4-1-fast, grok-4-1-fast-non-reasoning, grok-code-fast-1 and grok-3 on 2026-05-15 (retirement notice). It still accepts every one of those slugs — each is silently redirected to a live model (grok-4.3, or grok-build-0.1 for grok-code-fast-1) and, crucially, billed at grok-4.3's rate of $1.25 / $2.50 per MTok regardless of which name was sent.

The registry still carried their pre-retirement prices, so the gateway was charging callers less than xAI charges us:

Model	Gateway charged (in/out per MTok)	xAI bills	Effect
grok-4-fast	$0.20 / $0.50	$1.25 / $2.50	6.25x under on input, 5x on output
grok-4-1-fast	$0.20 / $0.50	$1.25 / $2.50	6.25x / 5x under
grok-4-1-fast-non-reasoning	$0.20 / $0.50	$1.25 / $2.50	6.25x / 5x under
grok-code-fast-1	$0.20 / $1.50	$1.25 / $2.50	6.25x / 1.7x under
grok-3-mini	$0.30 / $0.50	$1.25 / $2.50	under
grok-4	$3.00 / $15.00	$1.25 / $2.50	overcharged callers
grok-3	$3.00 / $15.00	$1.25 / $2.50	overcharged callers

The fix

All seven now price at _XAI_RETIRED_INPUT_USD / _XAI_RETIRED_OUTPUT_USD, i.e. grok-4.3's rate.

They stay registered rather than being removed, so older SDK versions keep working — the SDK has dropped them from TEE_LLM in the companion PR — and they're marked for removal once no client sends them. grok-3-mini isn't named in xAI's notice but is gone from their models page, so it's priced the same way rather than left at its old rate.

Also in this pass

Audited every provider's deprecation list while I was here:

Corrected the Gemini note. It claimed gemini-2.5-flash/-pro/-flash-lite were "scheduled for deprecation on June 17, 2026 (flash-lite: July 22, 2026)". Those dates have passed and Google's list doesn't carry these models — they're still served. The only Gemini retirement that matters here is gemini-3.1-flash-image-preview, and the registry already calls the GA gemini-3.1-flash-image.

Recorded the announced OpenAI shutdowns next to the affected entries (deprecations):

• gpt-4.1-nano and o4-mini stop resolving 2026-10-23. OpenAI names those exact bare slugs, not dated snapshots, so they hard-fail that day. Replacements gpt-5.6-luna / gpt-5.6-terra are already registered.
• o3, gpt-5, gpt-5-mini are undated aliases of snapshots retiring 2026-12-11. OpenAI does not document whether such an alias is repointed to a successor or retired with its snapshot. If it's repointed, these become mispriced exactly the way the xAI slugs were — worth a calendar reminder.

Fixed the model lists in CLAUDE.md and README.md, which advertised models the registry doesn't accept: claude-sonnet-4-0, claude-3-7-sonnet, claude-3-5-haiku, grok-2 and gemini-3-pro-preview are all absent from _MODEL_LOOKUP, and the README's completion example posted "claude-3.7-sonnet". Both lists are now checked against the registry, and the retired-but-registered models are documented in one place.

For the record, no Anthropic model in the registry is retired — all 11 resolve to models Anthropic lists as Active, checked at the api_name level.

Testing

• 457 passed, 6 skipped on the CI selection (tee_gateway/test/, tests/test_pricing.py, tests/test_opengradient_field.py)
• make lint clean (ruff + mypy, 61 source files)
• The five pricing tests that pinned the old Grok rates now pin the new ones, plus a new test_retired_xai_slugs_priced_at_grok_4_3_rate asserting the whole set matches grok-4.3. test_grok_4_fast_cheaper_than_grok_4 became vacuous (both retired, now identical) and is replaced by the same comparison over the live tiers, grok-4.3 vs grok-4.6.

PCR note

This touches model_registry.py, which is baked into the image, so PCRs change. measurements.txt is not updated here — matching every prior model-registry PR (e.g. #160, #153), which touch this same set of four files and leave it alone.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QmnaiLePKAQEkNfR8VPfja

---

Generated by Claude Code