If your code still sends requests to grok-4-1-fast-non-reasoning, grok-3, or six other Grok model IDs, those requests still work. They worked before 15 May 2026 and they have kept working since. Nothing errored, nothing warned, nothing in your logs changed. What changed that day was the bill.
They are being served by grok-4.3 and billed at grok-4.3 prices: $1.25 per 1M input tokens and $2.50 per 1M output.
Last verified: 7 August 2026 against the migration notice and current model docs. Re-checked monthly.
(The company is now SpaceXAI — xAI was absorbed into SpaceX. The docs still say xAI. More on that below.)
The eight slugs
From xAI's model retirement notice, effective 15 May 2026 at 12:00 PM PT:
| Retired slug | Now served by |
|---|---|
grok-4-1-fast-reasoning |
grok-4.3, low reasoning effort |
grok-4-1-fast-non-reasoning |
grok-4.3, none reasoning effort |
grok-4-fast-reasoning |
grok-4.3, low reasoning effort |
grok-4-fast-non-reasoning |
grok-4.3, none reasoning effort |
grok-4-0709 |
grok-4.3, low reasoning effort |
grok-code-fast-1 |
grok-build-0.1 |
grok-3 |
grok-4.3, none reasoning effort |
grok-imagine-image-pro |
grok-imagine-image-quality |
A note on the name
The company that retired these models was xAI. It isn't called that any more.
SpaceX acquired xAI in a merger that closed 2 February 2026, Musk said in May that xAI would be
dissolved as a separate company, and on 6 July 2026 the rebrand to SpaceXAI was made official
with a new logo and a renamed X account.
The developer surfaces have not caught up. As of 7 August 2026 the API docs still live at
docs.x.ai, still say "the xAI API", and still point support at support@x.ai — no mention of
SpaceX or SpaceXAI anywhere on the retirement notice. The models are still Grok.
So: the retirement was announced by xAI, the entity is now SpaceXAI, and the thing you actually
type into your config didn't change at all. This page says "xAI" where it's quoting xAI's own docs,
because that's what they say.
Why this one is easy to miss
Most deprecations announce themselves. You get a 404, a deprecation header, something in a changelog you eventually read. This one is the opposite, and xAI says so plainly:
After May 15, 2026 at 12:00 PM PT, requests to the retired model slugs above will automatically redirect to
grok-4.3. The slugs themselves continue to resolve, so you do not need to change your code to avoid breakage.
That's genuinely considerate engineering — nobody's production traffic fell over. But the same doc adds:
If you continue sending requests to a deprecated slug after May 15, please be aware that you will be billed at
grok-4.3pricing
So the failure mode isn't an outage. It's a bill that changed while every string in your codebase stayed the same. If you set grok-4-1-fast-non-reasoning in an env var eight months ago and haven't looked since, you are on grok-4.3 pricing and your config still says otherwise.
What you actually pay now
Current xAI model pricing, checked 7 August 2026:
| Model | Input / 1M | Output / 1M |
|---|---|---|
grok-4.5 (<200k ctx) |
$2.00 | $6.00 |
grok-4.5 (≥200k ctx) |
$4.00 | $12.00 |
grok-4.3 (<200k ctx) |
$1.25 | $2.50 |
grok-4.3 (≥200k ctx) |
$2.50 | $5.00 |
grok-4.20 reasoning / non-reasoning |
$1.25 | $2.50 |
grok-4.20-multi-agent |
$1.25 | $2.50 |
grok-build-0.1 (<200k ctx) |
$1.00 | $2.00 |
A note on the old prices. You'll see figures circulating for what the retired fast models used to cost — grok-4-1-fast at $0.20 / $0.50 is the one most often repeated. We're not stating it as fact: xAI no longer publishes pricing for retired models and we couldn't reach an archived copy to confirm it. For your own delta, the only record of what you were actually charged is your xAI billing history either side of 15 May.
You don't need that number to see the problem, though. Everything above is published, and it already shows you're on the mid tier: grok-4.3 at $1.25/$2.50 when grok-build-0.1 sits right there at $1.00/$2.00. You didn't choose that — a redirect did.
What isn't in doubt is the direction. The retired slugs were the cheap, fast tier. grok-4.3 is not the cheap, fast tier.
Check in thirty seconds
grep -rEn 'grok-4-1-fast|grok-4-fast|grok-4-0709|grok-code-fast-1|grok-3\b|grok-imagine-image-pro' \
--include='*.js' --include='*.ts' --include='*.py' --include='*.go' \
--include='*.rb' --include='*.rs' --include='*.java' --include='*.json' \
--include='*.yaml' --include='*.yml' --include='*.env' --include='*.toml' .
(One --include per extension on purpose. The tidier-looking --include='*.{js,ts,py}' matches nothing at all: the braces are inside quotes so the shell never expands them, and --include doesn't do brace alternation — so it fails silently and looks like a clean bill of health.)
Also check the places code search misses: environment variables in your hosting dashboard, CI secrets, serverless config, and any model name stored in a database or feature flag.
If you get hits, you have a decision rather than an emergency. Requests still succeed. But you should pick the model deliberately instead of inheriting one, and grok-build-0.1 at $1.00/$2.00 is cheaper than grok-4.3 if the work suits it.
The wider lesson
This is the same class of problem as free tiers that quietly shrink — the number you built against stops being the number you're on. Silent redirect is becoming the normal way model providers retire things, because breaking production is worse press than a quiet price change. It's the right call for uptime. It does mean the old assumption — if it still runs, nothing changed — no longer holds for AI APIs.
The practical habit: treat model IDs like dependency versions. Pin them somewhere visible, and re-read the provider's model page on a schedule rather than when something breaks. Nothing will break.
Top comments (0)