Author’s note: According to Musk’s exact words on 2026-11-10, Grok 4.8 will first get a small-scale soft release for Grok Bot users later this week. The public API currently has no date and no model ID, so for now, keep calling grok-4.7.
Let me cut to the conclusion first: there’s no Grok 4.8 API yet. In a reply on 2026-11-10, Musk said there would be a soft rollout to Grok Bot users later this week — meaning they’ll first test it with a small group to see how it performs. So the thing that moves first this week is the Bot, not the API. As of the 2026-11-13 check, Grok 4.8 still has no model card, so you can’t treat it as live.
This isn’t a rumor. Everything quoted below comes straight from Musk’s replies on X, and the platform-side info is from the model list and pricing endpoint opened on 2026-11-13. This article skips the parameter-count speculation and looks only at the five things he named and the three-tier timeline.
Core value: In three minutes, get clear on where Grok 4.8 is stuck, who gets it first this week, and which model ID to use when calling Grok today.
Grok 4.8 API: Key Facts at a Glance
This reply began with a question from user @cradrrat on 2026-11-10: “Can we have an update on when to expect Grok 4.8 please”. Musk replied at 17:35 Beijing time that day (09:35 UTC). The table below only covers what the original post explicitly says.
| Item | Details (based on the original wording) |
|---|---|
| Time posted | 2026-11-10 17:35 (Beijing time), replying to @cradrrat |
| Who posted it | Musk’s own X account; this is his own words, not an official blog |
| What’s happening this week | A soft rollout to Grok Bot users later this week, to see how it goes first |
| Training stage | Pretraining is complete; it’s now in the RL stage |
| Public API | Not mentioned in the original post—no date and no model ID |
| What the original post doesn’t say | The exact date and time, leaderboard name, score, price, context length, model ID |
Two things are worth separating here. “Bot users” means Grok Bot users, not hardware robots. A “soft rollout” means releasing it to a small group first and watching the feedback—it’s not a full launch, and it definitely doesn’t mean the API is open. A lot of people see “this week” and assume the API will be callable this week, but that’s mixing up two different tracks.
Where Grok 4.8 Lost Time: The 5 Items Musk Named
The progress part of the reply is just five sentences, each covering one step. Here they are in the original order; the quoted text is his wording.
| No. | Key quote | Meaning |
|---|---|---|
| 1 | dumped JAX/SGLang | Dropped the old JAX training framework and SGLang inference framework |
| 2 | rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time | Rewrote the entire RL and inference stack from scratch in C/C++; he explicitly says this is why it took extra time |
| 3 | Training (aka pre-training) is done | Pre-training is complete |
| 4 | Lot of cleanup was needed on the SFT data to get that mostly right | They did a lot of cleanup on the SFT data; in his words, they got it “mostly right” |
| 5 | 4.8 is in RL and the evals are starting to climb | It’s in RL now, and evals are climbing, but no benchmarks or scores were given |
The Main Reason Grok 4.8 Is Delayed: Rewriting the RL and Inference Stack
Of the five items, only the second was flagged by Musk himself as having “taken some extra time.” Rewriting the RL and inference stack from scratch is foundational engineering work: it affects not just training but also the inference service after launch. So the delay isn’t in pre-training; it’s in the second half of training and the inference infrastructure.
That also explains why the Bot comes first. The inference stack has just been rewritten, so observing it at a small scale inside a controlled in-house entry point like Grok Bot is safer than opening the API right away. That said, this is reading between the lines—Musk didn’t explain why he chose the Bot.
Grok 4.8’s Evals: Direction Only, No Numbers
“evals are starting to climb” only indicates a trend. The original post doesn’t name any benchmark or give any scores. Until the official model card and evaluation data are released, any claim about “how many points 4.8 scored” has no basis in this original post.
In the same reply, Musk also made two qualitative statements: “Grok 4.8 will bear no resemblance to 4.7 (thankfully)” and “It is not yet incredible, but this time, the foundations are built right.” In other words, he admits that 4.8 isn’t incredible yet and stresses that the foundations were built right. In another reply that same day, someone asked how 4.8 compares with Opus 5.5; he answered “Not as good” and put the expectation of surpassing it on Grok 5. So moving fast doesn’t mean the model is already very capable.
Grok 4.8’s Three Layers of Timing: Bot, API, and Platform Listing
Put the original post and the platform pages side by side, and the timeline breaks into three layers. Certainty drops with each layer, so they shouldn’t be conflated.
| Layer | Status | Basis |
|---|---|---|
| Grok Bot user soft launch | Musk said later this week; as of 2026-11-13, no model card has appeared | Original post on 2026-11-10 |
| Public API | No date, no model ID | Not mentioned in the original post |
| APIYI platform | Neither the model list nor the pricing endpoint has grok-4.8 | Page checked on 2026-11-13 |
Grok 4.8’s Bot Soft-Launch Window
2026-11-13 is a Friday, and “later this week” refers to the remaining days of the week that 11-10 falls in, so the window is still open. But as of the check, no model card had been seen, so we can only say the window hasn’t closed—we can’t say it’s already been released. Even if the Bot side releases it, that’s only a small-scale observation; whether it expands later still depends on what Musk called “see how it goes.”
Grok 4.8 API and Platform Status
The public API layer wasn’t mentioned in the original post. On the platform layer, as of 2026-11-13, APIYI’s model list has grok-4.7, not grok-4.8; the pricing endpoint’s model_name also only has grok-4.7. The real-time updates page’s November 2026 archive has 26 entries total, from 11-01 through 11-13, and contains no record of Grok 4.8 being listed.
Once grok-4.8 is listed, we’ll get this model onto the platform as soon as possible. Before that, the platform doesn’t have this ID; even if you enter it, the call won’t go through.
Grok 4.8 Impact Analysis: Which ID Should You Call Now
Impact on Developers
- Don’t change your code and wait for 4.8: The Bot soft launch and the API are two separate tracks. Just because the Bot makes moves this week doesn’t mean the API will follow.
- What you can call now is grok-4.7: This is the model xAI released on 2026-09-21. According to xAI’s own official blog, pricing starts at $2 per 1M input tokens and $6 per 1M output tokens.
- When 4.8 arrives, you’ll need to retest: Musk said 4.8 and 4.7 have “nothing in common.” It’s best to rerun your existing prompts and evaluation baselines at that point—don’t assume they’re compatible by default.
- Don’t invent a fast model ID yourself: Officially, 4.7 has a separate fast variant that’s 2x the speed and 2x the price. APIYI’s master list doesn’t include a separate fast model ID, so don’t make one up and try to call it.
Parameters for invoking grok-4.7 now
You can just follow the OpenAI-compatible interface pattern. Here are the key parameters:
- Endpoint base_url:
https://api.apiyi.com/v1 - Model ID:
grok-4.7 - Endpoint type:
chat(Chat Completions) - Available groups:
CodexResponses,Default,SVIP - Billing: pay-as-you-go, billed per 1M tokens
grok-4.7 doesn’t have a separate model detail page (the corresponding path returns 404), so parameters and pricing follow the master model list. Also note that the pricing, context window, and metrics on the neighboring grok-4.6 page belong only to 4.6—they can’t be applied to 4.7 or 4.8.
🚀 Quick start: If you already have OpenAI SDK code, just swap base_url for APIYI’s endpoint and set model to
grok-4.7, and you’re good to go. New users can claim free trial credits at APIYI apiyi.com, measure a baseline for 4.7 first, and then compare once 4.8 is available.
grok-4.7 pricing tiers at 200K tokens
grok-4.7 has two tiers based on the context length of a single request, split at 200,000 tokens. Data date is 2026-11-13, and prices are in USD per 1M tokens.
| Tier | Input | Output | Cache read | Source note |
|---|---|---|---|---|
| ≤ 200K | $2 | $6 | $0.5 | Master list and pricing expression agree |
| 200,001–500K | $4 | $12 | $1 | Input $4 agrees in both places; output $12 and cache read $1 appear only in the pricing expression |
There’s one inconsistency worth calling out here. The tiered description in the master model list only covers input: $2 for 0–200,000 tokens and $4 for 200,001–500,000 tokens. It doesn’t separately list output or cache read for the higher tier. The pricing API’s billing expression, however, specifies that after 200K, input, output, and cache read are $4, $12, and $1, respectively. Both are listed here. Actual charges follow the expression, and model and pricing details follow the model list on the APIYI apiyi.com platform.
For cost planning, you can split it like this: keep everyday conversations and short-to-medium documents under 200K to enjoy the lower-tier $2 / $6 / $0.5 pricing. For tasks that must max out the context window, like large code repositories or bulk document processing, budget with the higher-tier input and output unit prices, which are doubled. For long prefixes you can reuse, route them through cache reads whenever possible—that’s the cheapest component in both tiers.
Impact on the industry
- xAI’s cadence has become “Bot first, API later”: At least this time, the timeline Musk gave publicly only covers Bot users — API developers will have to wait for the next statement.
- Rewriting the underlying stack is a long-term investment: Dropping JAX/SGLang in favor of a homegrown C/C++ stack slowed things down in the short term; he’s betting on faster iteration down the line, and that’s the weight behind the line “foundations are built right.”
Grok 4.8 FAQ
Q1: Can I call the Grok 4.8 API right now?
No. As of 2026-11-13, the original post doesn’t mention an API, there’s no model ID, and grok-4.8 doesn’t appear in the platform model table or the pricing endpoint. What you can call right now is grok-4.7.
Q2: What does “giving it to Bot users first this week” mean?
It means a soft rollout to Grok Bot users later this week — a small-scale release to see how it performs. “Bot users” here doesn’t mean hardware robots; this isn’t a full release either, and definitely not an open API. As of the 11-13 check, there’s still no model card.
Q3: Why is Grok 4.8 later than expected?
In Musk’s own words, xAI dropped JAX and SGLang and rewrote the entire RL and inference stack from scratch in C/C++, a step that “took some extra time.” Pretraining is finished, the SFT data has been heavily cleaned up, and the model is currently in the RL stage.
Q4: What’s Grok 4.8’s actual benchmark score?
The original post only says “evals are starting to climb” — no leaderboard names, no scores. Until the official model card is out, there are no specific numbers to cite.
Q5: How do I call Grok now, and what will I need to change once 4.8 is live?
Right now, set base_url to `https://api.apiyi.com/v1` and model to `grok-4.7`. Once grok-4.8 is available, APIYI will get it listed as soon as possible; at that point, go by the official ID in the platform’s model list — usually you just change the model field and re-run your evals. On APIYI apiyi.com, you can use the same Key to switch between the two versions and compare them.
Q6: How is grok-4.7 billed above 200K tokens?
Per the billing expression on the pricing endpoint, the 200,001–500K tier is $4 input, $12 output, and $1 cached read (per 1M tokens, data dated 2026-11-13). The main table’s tier notes only list the high-tier input at $4, without breaking out output and cached read separately — that’s a spot where the two sources disagree, so keep it in mind when budgeting.
Grok 4.8 Key Takeaways
- No API yet: Grok 4.8 doesn’t have a public API or a model ID at this point.
- Bot first this week: According to the exact words from 11-10, there’ll be a limited soft rollout to Grok Bot users later this week.
- The delay comes from rewriting the stack: Rewriting the RL and inference stacks from scratch in C/C++ is the reason it’s taking longer.
- Progress has reached RL: Pretraining is done, SFT data has been cleaned up, and evals are trending up—but there are no scores yet.
- Fast doesn’t mean strong: Musk himself says 4.8 isn’t all that impressive yet, and doesn’t match Opus 5.5.
- Use 4.7 today: On APIYI apiyi.com, use
grok-4.7, with 200K tokens as the cutoff for controlling tier costs.
Summary
The Grok 4.8 situation boils down to three points:
- What’s coming fast is the Bot, not the API: The soft rollout later this week is only for Grok Bot users—there’s no date for a public API.
- The delay is in the inference and RL stack: Dropping JAX/SGLang and rewriting in C/C++ is the reason Musk himself acknowledged for the delay.
- Keep using grok-4.7 for now: The low tier is $2 / $6 / $0.5; budget for the high tier above 200K. Watch out for the discrepancy between the summary table and the pricing expression.
Until 4.8 gets an official model ID, the most practical approach is to establish a business baseline with 4.7 first, then run an A/B comparison once the new model is available. We recommend using APIYI apiyi.com to do this with a unified interface—we’ll follow up as soon as grok-4.8 goes live.
📚 References
- Musk’s reply to the original post on 2026-11-10: his exact words on Grok 4.8 progress and the Bot soft launch
- Link:
x.com/elonmusk/status/2120088070443655190 -
Note: the sole basis for the five points in this article and the time window
-
The original question from user @cradrrat: asking when Grok 4.8 is expected
- Link:
x.com/cradrrat/status/2119949070462800185 -
Note: the context for the reply above
-
Musk’s reply the same day about Opus 5.5: a qualitative comparison of 4.8 and Opus 5.5
- Link:
x.com/elonmusk/status/2120099798078656909 -
Note: used to show that moving fast doesn’t mean being more capable
-
xAI’s official blog post on Grok 4.7: release details for 4.7 and the official self-reported pricing
- Link:
x.ai/news/grok-4-7 -
Note: the explanation of the fast variant also comes from here
-
APIYI model master list: currently available models and tiered pricing
- Link:
docs.apiyi.com/models - Note: checked on 2026-11-13 — grok-4.7 is listed, grok-4.8 is not
Author: APIYI Tech Team
Technical support: If you run into integration issues, feel free to reach us through the APIYI official website homepage. Model updates are available in the docs center at docs.apiyi.com
