1. Home
2. News
3. Research

Musk says Grok 4.8 trails Claude Opus 5.5, which he calls a masterpiece

Grok 4.8 was supposed to finish training in September; SpaceXSI rewrote its training stack in C instead.

By Think Facility · November 11, 2026 · SpaceXSI AI models Save

On November 10, 2026, Elon Musk said on X that Grok 4.8 is behind Anthropic's Claude Opus 5.5. Asked how the two would compare, he answered "Not as good," called Opus 5.5 a masterpiece, and said Grok 5 should pass it in two or three months.

Earlier that morning he'd explained why Grok 4.8 is late. SpaceXSI threw out the software it used to train and run the model and wrote a new one from scratch in C and C++. On September 13 he had posted that Grok 4.8 would finish training that week.

1. Sep 13, 9:24 pmMusk says training ends this week
2. Nov 10, 4:35 amMusk says it's in RL
3. Nov 10, 4:46 amMusk names the language
4. Nov 10, 5:22 amMusk ranks it below Opus 5.5

Eastern time, read off each post on X.

What took two months

Musk's update came as a reply to a user asking when Grok 4.8 would arrive. He says pre-training is done and the model is now in RL, with its evals starting to climb. The stack for RL and for serving the model is gone too: SpaceXSI "dumped JAX/SGLang," in his words, and wrote the replacement itself.

JAX is Google's open-source library for numerical computing and SGLang is an open-source engine for serving language models, so he's describing a move off widely shared tools. Asked what language, he said C with a little C++. In another post he said the new stack comes from work SpaceX engineers did on Starlink.

He also admits the September timeline slipped. The rewrite "took some extra time," and his post says the supervised training data needed a lot of cleanup too. He gives no date beyond a soft rollout to Bot users later this week. His own verdict on the result is that it's not yet incredible, but that the foundations are right this time.

Ranking his own model second

The Opus line was a reply to two other users, and it's short enough to show whole.

It isn't the first time Musk has leaned on Anthropic's model. On October 7 he said Grok Bot would route tasks to whichever back end gives the best result, and Claude Opus 5.5 was first on his list.

Grok 5 is the model he's promising to beat it with. In September he'd also said Grok 5 is the one he expects to reach AGI. The "2 or 3 months" comes with a "should" and a "maybe" in front of it, and his record on this exact model is a one-week estimate that turned into two months, so I'd treat it as a hope.

None of these posts gives a benchmark score for Grok 4.8. Anthropic hasn't replied to Musk in any post I found.

Who's paying for Grok in the meantime

The federal government is one customer that has to pick between these models. In September 2025 the General Services Administration signed a OneGov deal that lets federal agencies use Grok for $0.42 per organization, for 18 months, valid until March 2027.

Musk was quoted in that release calling xAI's models the most capable anywhere. A month earlier GSA had signed a similar deal with Anthropic, offering Claude to all three branches of government for $1 per agency for up to a year. So agencies can already get either one for close to nothing, and the price isn't what separates them.

For a procurement officer, the company's own chief now says its current model trails the rival on the same price list. The Grok deal runs out in March 2027, a few weeks after Musk's two-or-three-month window for Grok 5 would close. Grok 4.8's soft rollout to Bot users is due later this week.