Grok 4.8 Soft Rollout: xAI Finishes Pretrain and Rewrites RL in C++

Elon Musk said on November 10, 2026 that Grok 4.8 finished pretraining, entered reinforcement learning on a rewritten C/C++ stack, and will reach some Grok Bot subscribers in a soft rollout later that week. The post is the first concrete schedule since September's 2.5-trillion-parameter tease. It is not a full launch: no xAI newsroom article, no API price file, and no public benchmark table accompanied the message.

For builders, the actionable detail is where the model will land first (Grok Bot, xAI's persistent agent product) and what changed underneath (dropping JAX and SGLang for in-house C/C++ RL and inference). Everything else is roadmap language until weights hit a surface you control.

TL;DR: what is confirmed vs still missing

What Musk posted (primary source via quotation)

Multiple outlets on November 10, 2026 relay the same English text attributed to Musk on X. Representative summaries appear on finance and tech wires the same day; we treat the quoted sentences as primary only where they match across independent relays.

Reported lines include:

• xAI "dumped JAX/SGLang" and "rewritten the entire RL and inference stack from scratch in C/C++", which "took some extra time."
• "Training (aka pre-training) is done."
• "Lot of cleanup was needed on the SFT data" before RL.
• "4.8 is in RL and the evals are starting to climb."
• "We will do a soft rollout to Bot users later this week and see how it goes."
• "Grok 4.8 will bear no resemblance to 4.7 (thankfully)." It is "not yet incredible," but "this time, the foundations are built right."

We did not independently scrape Musk's account in this run. If the live post text diverges from wire quotes, trust the on-platform post over secondary summaries.

Why the JAX and SGLang line matters for infra readers

Grok 4.8 reinforcement-learning phase after xAI rewrites training and inference away from JAX toward a custom C++ stack

xAI's early Grok stack was public about outside components. The original Grok announcement on x.ai (November 2023) said training and inference were built on Kubernetes, Rust, and JAX. JAX is Google's research-oriented Python array framework; SGLang is an open inference stack popular for agentic workloads and RL rollouts.

Musk's November post describes a phase change: not just a bigger Grok, but owning the full RL + serving path in C/C++. That aligns with his September 14 claim that Grok 4.8 trained on a new C++ software stack at 2.5 trillion parameters, documented in our September Grok 4.8 explainer.

What you cannot verify from the post alone: throughput numbers, GPU utilization, or side-by-side latency versus SGLang-served Grok 4.7. Treat the rewrite as credible engineering intent, not a published benchmark.

Timeline: September promise vs November status

The 57-day gap between the September "RL next" framing and the November "now in RL" detail is calendar math between two Musk posts, not an xAI SLA. Musk explicitly blamed stack rewrite time and SFT cleanup, not a silent cancel.

Grok Bot vs the chat app: who actually gets 4.8 first

Musk scoped the rollout to Grok Bot users. That product line is xAI's persistent, tool-using agent surface — the same family we covered when Grok Bot entered early beta with always-on computers and third-party connectors.

A soft rollout usually means feature flags or cohort sampling: some paying Bot customers see the new model, failures get rolled back, and product telemetry feeds the remaining RL loop. Musk gave no cohort size, no geography, and no eligibility rule (SuperGrok tier, enterprise, etc.).

Not promised on November 10: Grok.com chat default model switch, Grok Build harness coverage coding routes, Microsoft Foundry / Bedrock SKUs, or Cursor's model picker. Until those channels list grok-4.8 (or whatever the API string will be), assume 4.7 remains production everywhere except the unnamed Bot slice.

SFT cleanup: what Musk implied without detail

Supervised fine-tuning is the stage where a base model learns curated prompt-and-answer pairs before RL rewards shape behavior. Musk said "lot of cleanup was needed on the SFT data to get that mostly right." He did not list which datasets, which safety filters, or which refusal templates moved.

For practitioners, that line matters because SFT quality caps RL. A noisy SFT mix produces reward hacking and odd refusals that RL must burn GPU weeks to sand down. xAI has not published an SFT card for 4.8. Compare Anthropic's habit of pairing flagship releases with system cards — Fable 5.2 shipped with documentation on November 9 while Grok 4.8 remained post-only on November 10.

Grok Bot complaints the week before rollout

Musk's timing sits next to paid Grok Bot usage complaints reported on November 9 in the same news cycle: subscribers said quotas drained faster than expected, and Musk promised overnight fixes in relayed coverage. A model still in RL is a risky moment to widen a paying cohort, which may explain soft language instead of a default swap.

If you run agents on Grok Bot, treat the rollout as a canary: log task success rate, token burn, and tool-error loops per session, and keep a rollback plan to 4.7 routes in your own router config even if the product UI does not expose one.

Open stack vs in-house C++: builder takeaway

Teams choosing inference frameworks often debate SGLang/vLLM versus bespoke C++ for last-mile latency. Musk's post is a datapoint that a top lab bet the rewrite was worth months of calendar slip. That does not prove C++ wins for your API — most teams lack xAI's Colossus-scale ops — but it explains why 4.8 calendar diverged from 4.7 marketing.

If you fine-tune open models, none of this changes your week unless xAI drops 4.8 weights or API pricing that undercuts Grok 4.7's launch economics. Watch the API changelog, not the social thread.

Competitive context without invented scores

The week of November 10 is crowded at the frontier:

• Fable 5.2 landed with Anthropic pricing and safety messaging on November 9.
• Gemini 4 Carbon access chatter continued on cloud price lists, as we tracked in Vertex access claimed-vs-verified coverage.
• Grok 4.7 still holds xAI's latest documented benchmark page from September.

Musk's "evals climbing" line is an internal monitoring claim, not a replacement for how to read public leaderboards. Our frontier comparison hub will need a row for 4.8 only after reproducible evals exist — not when a soft Bot rollout starts.

What builders should watch this week

Soft Grok 4.8 rollout path from finished pretrain through RL evals to a limited Grok Bot cohort before any API-wide default switch

1. Model ID announcements on the xAI API docs and partner consoles. Soft Bot tests often precede API GA by days or weeks — see the pattern on Grok 4.7's staged launch.
2. Regression suites on coding and tool-use tasks. Musk already contrasted 4.8 with 4.7 qualitatively; your repo's CI tasks are the honest judge.
3. Token economics. If 4.8 is larger or runs on a new inference stack, watch price per million tokens and reasoning-length defaults — Grok 4.7's history included RL fixes for response length.
4. Musk's own bar. He said 4.8 is not yet incredible. That is unusually blunt pre-release framing; plan for rapid patch cadence during RL, similar to early Grok 4.6/4.7 timeline slips.

What people are asking

Will API customers get 4.8 the same day as Bot users? Unknown. The November 10 quotes mention Bot users only.

Does soft rollout mean GA? No. It is an experiment gate before wider defaults.

Is 4.8 the model Musk tied to "Fable/GPT-6 class" timing? In September he sketched 4.9 as likely Astra/Fable-class and pointed AGI jokes at Grok 5, in our two-to-three-month forecast post. November 10 does not move that ladder — it says 4.8 is still in RL.

Should I change Cursor or Copilot defaults now? Not without measurement. Partner integrations trail xAI's internal Bot channel.

Related reading

• Grok 4.8: 2.5T parameters and C++ stack (September)
• Grok 4.7 launch benchmarks and pricing
• Grok Bot early beta architecture
• Claude Fable 5.2 launch
• Frontier models compared (November table)
• Grok 4.7 vs GPT-6.1 Sol midterm study (API behavior)

Specs, rollout scope, and pricing are accurate as of November 10, 2026. Check x.ai/news and your provider's model list before routing production traffic.