1. Musk says Grok 4.8 will trail Claude Opus 5.5
Elon Musk said on Tuesday, November 10, that xAI's upcoming Grok 4.8 model will be "not as good" as Anthropic's Claude Opus 5.5, which he described as a "masterpiece." Replying to a user on X, Musk added that Grok 5 should exceed the Anthropic model in roughly two or three months, and that Grok 4.8 could begin a limited rollout later this week.
The remarks came in response to a user named Dimdv, who had asked how the forthcoming Grok release would compare with Anthropic's model. Musk's answer was unusually direct for the head of a company competing in the same market: the update xAI is about to ship will not match the product a rival has already had on sale for weeks.
xAI is the artificial intelligence company Musk leads, and Grok is its family of models, offered to users through a chatbot that is closely tied to the X platform. Anthropic is a rival developer whose models are sold under the Claude name, with Opus among its more capable tiers. The two companies compete for the same developers, businesses and consumers, alongside OpenAI and others.
Praise of this kind for a competitor is uncommon in the artificial intelligence industry, where developers typically present each release as a step ahead of the field. Musk, who leads xAI and owns X, instead placed the Anthropic model above his own company's next release and shifted the claim of leadership to a model that has not yet arrived.
2. Grok 5 is the model Musk expects to pull ahead
Two separate releases are in play in Musk's comments, and they carry very different expectations. Grok 4.8 is an update to the current generation, due to reach some users this week. Grok 5 is the next-generation model that follows it, and it is the one Musk believes could put xAI in front.
Musk said Grok 5 "should exceed" Opus 5.5 in about two or three months. He offered no supporting detail for that forecast: no comparison scores, no description of what would make the difference and no firm date.
The timeline should be read with some caution. Musk has a record of missing targets he has set for himself, and the estimate he gave is a prediction, not a scheduled launch. Even if Grok 5 arrives on the timetable he described and performs as he expects, any lead would last only until competitors release new models of their own, which is the normal pattern in a field where the leading developers ship major updates several times a year.
The distinction between the two models also matters for how users should read this week's release. An incremental version number such as 4.8 signals a refinement of an existing generation, not a new one. Musk's comments make clear that he does not expect that refinement to change the competitive order, and that he regards the generational step to Grok 5 as the one that counts.
For the moment, then, the position Musk has set out is that xAI is behind, expects to remain behind after its next update, and is counting on the release after that to change the ranking.
3. Why is Grok 4.8 taking longer?
Grok 4.8 has been delayed by unfinished training work and by a decision to rebuild part of xAI's software, according to a technical update from Musk. Pretraining of the model is complete, he said, but reinforcement learning is still under way, and the team also spent considerable time cleaning the data used for supervised fine-tuning.
Those three terms describe successive stages in building a large language model. Pretraining is the initial and most compute-intensive phase, in which a model learns general patterns of language from a very large body of text. Supervised fine-tuning follows, using curated examples to teach the model how it is expected to respond. Reinforcement learning then refines behavior further by rewarding outputs that meet a desired standard.
Musk's update therefore indicates that the foundation of Grok 4.8 is in place while the later, behavior-shaping stages are still being worked through. The time spent cleaning fine-tuning data points to an effort to improve the quality of the examples the model learns from, since errors or inconsistencies in that material tend to carry over into how a model responds.
He summed up the state of the model in measured terms, calling Grok 4.8 "not yet incredible" while saying that its foundations had finally been built correctly.
4. A rewritten stack in C and C++
Alongside the training work, xAI replaced core pieces of its engineering toolkit. Musk said the company dropped JAX, a machine-learning framework, and SGLang, a tool for serving models, in favor of an inference and reinforcement learning system rewritten in C and C++.
Inference is the stage at which a trained model generates an answer after it receives a prompt. It is the part of the system that users interact with directly, and its efficiency affects both how quickly responses arrive and how much each one costs to produce. C and C++ are lower-level programming languages that give engineers close control over memory and hardware, at the price of more demanding development work than higher-level frameworks require.
Replacing a framework and a serving tool with custom code is a substantial undertaking. Off-the-shelf components let a team move quickly because much of the engineering has already been done by others, while a purpose-built system has to be written, tested and maintained in-house. Companies that take the second path generally do so because they want control over every layer of the system and are willing to spend engineering time to get it. Musk's update suggests that this rebuilding effort is part of the reason the release has taken longer than a routine update would.
Musk did not provide any performance figures for the new system, so the gains from the rewrite cannot be assessed from his statements. He presented the change as groundwork, in keeping with his description of the model's foundations being built correctly.
The disclosure drew a response from developers. One, posting under the name tetsuo, celebrated with the phrase "C wins again." Musk replied that the code was "C with a little C++."
He went on to credit Starlink engineers with helping to build the training and inference stack. Musk cited practices drawn from SpaceX, including writing its own headers and handling time precisely, down to accounting for leap seconds. The comment indicates that xAI drew on engineering talent from another part of Musk's group of companies for the rebuild.
5. When will Grok 4.8 be released?
A limited release of Grok 4.8 is expected later this week, according to Musk, who said the model would go first to what he called "Bot" users in a soft rollout. He described that step as an early test ahead of wider availability.
No date has been given for the broader release. A staged introduction of this kind lets a developer observe how a model behaves with a smaller group before exposing it to the full user base, and the timing of the wider launch will presumably depend on what that early test shows and on the reinforcement learning work that Musk said was still in progress.
Musk has not published any benchmark scores for Grok 4.8. Without them, his statement that the model is "not as good" as Opus 5.5 is the only public indication of where it stands.
6. Anthropic's head start with Opus 5.5
The model Musk singled out has been available since September 22, when Anthropic released Claude Opus 5.5. It reportedly performs comparably to the company's higher-end Fable 5.1 on most work, while costing 40% less to run than the previous Opus model on typical workloads.
Opus 5.5 is aimed at difficult programming work and at multi-step agentic tasks, in which a model carries out a sequence of actions with limited human help. That category has become a central battleground for AI developers, because it determines how much real work a model can be trusted to complete on its own.
Anthropic has reported a score of 66.4% for Opus 5.5 on Terminal-Bench 4.0, a benchmark that measures how well a model completes tasks on the command line. On the same test, the company put OpenAI's GPT-6 Astra at 57.9%. Those figures come from Anthropic.
Benchmarks of this type are the main yardstick the industry uses to compare models. Each one sets a fixed collection of tasks and reports the share a model completes correctly, which allows results from different developers to be placed side by side. They are imperfect guides to everyday performance, but in the absence of other data they are the closest thing to a common scoreboard, and they are the reason an unpublished score leaves a model's standing open to question.
Grok, by contrast, sits very far behind on leaderboards. That gap gives context to Musk's admission: his comments did not reveal a weakness that outside observers were unaware of, but they did confirm it from the top of the company and acknowledged that the next update will not close it.
7. What remains unknown
Because xAI has released no scores for Grok 4.8, the size of the gap between it and Opus 5.5 is unclear. Musk's own characterization is qualitative, and neither the technical update nor his replies on X put a number on the difference.
The effect of the rewritten software stack is similarly unquantified. Musk said the foundations were now right, but he gave no data on speed, cost or reliability.
The identity of the first recipients is also thin on detail. Musk referred to them only as "Bot" users and did not say how many people that covers or how long the early test would run before a wider release.
Two markers lie ahead. The first is the soft rollout of Grok 4.8, targeted for later this week, which should produce the first outside impressions of the model. The second is Grok 5, which by Musk's estimate is two or three months away and which he has said should exceed Opus 5.5. Whether that estimate holds, and what Anthropic and other rivals release in the meantime, will decide whether xAI's position changes.