Grok 4.8: What We Know, What We Don’t

Elon Musk gave Grok 4.8 its first substantial status update in weeks on November 10. The short version: pre-training is finished, the model is in reinforcement learning, and a limited rollout to Grok Bot users is planned for later this week. The longer version is a story about a rewritten software stack, cleaned-up training data, and a promise that 4.8 will bear little resemblance to 4.7.

There is enough public information to explain the plan. Public data remains too thin to judge the model. No xAI model card, API listing, price, context window, or independent benchmark result for Grok 4.8 was public when this article was written. Almost everything below comes from Musk or from xAI’s existing products, and the difference between a claim and a confirmed fact matters.

The update, in plain English

What is public, what is missing

For Grok 4.8 itself, Musk’s posts are the primary public record. xAI has not published a product page or an API model ID for it, so the project is real enough for a status update but not yet a product the public can use.

Publicly stated by Musk:

A decoder for the jargon

Pre-training is the long, expensive phase in which a model learns statistical patterns from a large body of text, code, and other data. It produces the base model. Instruction tuning is a separate stage, which is why the next stages exist.

SFT stands for supervised fine-tuning. Humans or synthetic pipelines provide example prompts and ideal answers, and the model learns to imitate them. Data quality matters enormously here: a clean set of demonstrations can improve behavior, while noisy or contradictory examples teach the model to be inconsistent. Musk’s “cleanup” comment points at that work.

RL means reinforcement learning. The model generates answers, receives a reward signal, and adjusts toward responses that score better. In modern language models, RL is often used for reasoning, tool use, and instruction following. “Evals are starting to climb” means internal test scores are improving during this process. It does not prove that the gains will hold on unseen tasks or that users will prefer the model.

Inference is the act of running the trained model to answer a request. The inference stack is the software that batches requests, manages memory, schedules work across chips, and returns tokens. It determines latency, throughput, and cost per request.

Evals are evaluations: standardized tests for coding, math, reasoning, tool use, safety, or long-horizon tasks. A benchmark score is only as useful as its methodology. Internal evals can be selected to show progress; independent evals are harder to tune.

Parameters are the learned weights in a model. A 2.5-trillion-parameter model is enormous by current standards, but the number does not translate directly into quality. Architecture, data, training recipe, and post-training matter at least as much. If the model uses a mixture-of-experts design, only a subset of parameters may be active for each token, and xAI has not published that detail.

JAX is a Python library from Google for high-performance numerical computing and machine learning. It is flexible, supports automatic differentiation, and is often used with TPUs. SGLang is an open-source serving framework for large language and multimodal models, built for fast inference. Rewriting both the RL and inference layers in C/C++ is a low-level bet: more control over memory, scheduling, and kernels, at the cost of more code and harder debugging.

Why the rewrite matters, and why it is risky

C/C++ does not guarantee speed over JAX or SGLang. Performance depends on kernel quality, memory movement, networking, batching, and the hardware underneath. What a low-level rewrite can buy is control: the ability to remove abstraction layers, tune scheduling for xAI’s own clusters, and avoid overhead that a general-purpose framework carries.

The risk is equally concrete. A rewrite of the RL and inference stack touches the system that trains model behavior and the system that serves every user request. It can improve efficiency, and it can also introduce regressions that are harder to debug than a Python stack. It is the kind of project that explains a delay without proving that the delay will pay off.

The public record supports the scope of the project. It falls short of supporting a performance claim. There is no published tokens-per-second comparison, no cost-per-task comparison, and no independent audit of the new stack. For now, treat “we rewrote it in C/C++” as an engineering decision rather than a result.

What Grok Bot is, and why it is the first test

Grok Bot is xAI’s always-on agent product, launched in August 2026. A Bot is a persistent, named agent with its own computer that can work inside tools such as Slack, Notion, Drive, and Teams, call MCP or API tools, and keep working across sessions. xAI describes Bots as “AI teammates” that can be given real work rather than single chat prompts.

That makes Bot users a sensible first audience for a new model. Agent tasks are long, multi-step, and tool-heavy: they expose planning failures, latency, and cost in ways a single chatbot answer does not. A soft rollout also limits the blast radius. It falls short of opening the model to every Grok or API user, and xAI has not said how many Bot users will get access, which plan they need, or what limits will apply.

What can we judge, and what can we not?

We can judge the direction. Musk describes a new base model, a new post-training run, and a rebuilt RL and inference stack, which amounts to a generation change rather than a point update. The engineer’s public post confirms that the infrastructure work was real. If the claims hold, Grok 4.8 is a new generation rather than a version bump.

We cannot judge the capability. There is no model card, no API model ID, no price, no context window, no benchmark table, and no independent evaluation. “Evals are starting to climb” is a statement about internal progress rather than a comparable score. “No resemblance to 4.7” is a claim about the model’s character. “Not yet incredible” is expectation management. All three can be true at once, and none of them answers the question a user actually has: is it better at the work I do?

For context, Grok 4.7 is xAI’s current public flagship. It was released on September 21, 2026, and xAI says it scored 46.3% on CursorBench 4.0 at $2 per million input tokens and $6 per million output tokens. Those are the numbers 4.8 will eventually be measured against. A new generation that costs more, runs slower, or fails more often on agent tasks would not be an improvement, even if its headline eval score is higher.

What to watch next

A model card and API listing. The first hard artifacts should confirm the parameter count, architecture, context window, pricing, and supported modalities. Until they exist, 4.8 is a development model.
Independent benchmarks. CursorBench, LiveBench, and other public evaluations matter more than a screenshot of internal progress. Look for methodology, sample sizes, and whether the tests were run after the model shipped.
Bot rollout details. Which users, which tasks, what limits, and what safety controls. A limited rollout can hide failure rates that appear at scale.
Cost and latency. The C++ rewrite is partly an efficiency bet. Ask for cost per completed task, tokens per second, and utilization, not just eval scores.
Reliability. Agent models fail in boring ways: loops, tool misuse, dropped context, and confident wrong answers. A model that is slightly less clever but far more reliable can be the better product.

Bottom line

The public record is sufficient to explain what xAI is trying to do. It remains insufficient to rank Grok 4.8. Musk has confirmed a new base model, a new RL run, a C++ rewrite, and a limited Bot rollout. He has not published the evidence that would let an outsider verify the claims or compare the result with Grok 4.7, GPT-6, Gemini, or Claude.

The next meaningful milestone is a model card, an API endpoint, and independent tests on real work, which matter more than another post about evals climbing. Until then, the honest summary is simple: Grok 4.8 is a serious engineering bet in progress, and the public can see the bet but not the payoff.