Elon Musk posted a Grok 4.7 timeline that is not the one explainx.ai covered in July. After Grok 4.6 shipped, he said 4.7 is significantly better, initial training is complete, and SpaceX company data is going into supplemental training — with a ship window of 3 to 4 weeks from around August 12, 2026. That is early-to-mid September 2026, not the July 25 claim of roughly August 22. This post is the August update. The July rumor thread stays as a record of what was claimed then.
The unique angle is not another point-release date. It is the SpaceX corpus as the stated edge, plus an unusually explicit compliment to Anthropic in the same breath as a "we'll still win at real-world engineering" claim.
TL;DR
| Question | Direct answer |
|---|---|
| When does Grok 4.7 ship? | Musk: 3 to 4 weeks from ~August 12, 2026 → early-to-mid September 2026 |
| Did the July date slip? | Yes. July 25 said ~August 22. That window is gone. Call it a slip, not a rounding error. |
| Has Grok 4.6 shipped? | Yes. Use it this week if you need a coding agent. See the 4.6 launch evals. |
| What's new in 4.7 vs 4.6? | Musk: "significantly better." Confirmed detail: SpaceX company data in supplemental training after initial training finished. No model card yet. |
| Will it beat every current model? | Claim, not eval. Musk said it will, especially at real-world engineering. No SWE-Bench, Terminal-Bench, or CAD suite published for 4.7. |
| 2.1T / "surpass Fable 5"? | Unverified community posts. Do not treat as official xAI or SpaceXAI specs. |
| Should I wait? | Engineering/CAD/rockets/manufacturing → wait-and-eval 4.7. Coding agent this week → shipped 4.6, or Grok 4.5 / Fable / Sol. |
What Musk actually said
Two posts, quoted in full because the paraphrase war on X is already rewriting them.
On beating the field, with a nod to Anthropic:
Grok 4.7 will exceed all current models. That said, Anthropic is a great company and will probably release improved models soon. However, the SpaceX training corpus is so awesome & unique that I would be shocked if any model is better at real-world engineering than 4.7.
On the calendar and the training stage:
Grok 4.7 is significantly better than 4.6 and should be ready in 3 to 4 weeks. Initial training is complete and now we're adding a massive amount of SpaceX company data in supplemental training. This will be something special.
Read those as founder posts, not a SpaceXAI blog, changelog, or model card. That is the same pattern as the July 25 4.6/4.7 thread and the Grok 4.5 public launch: X first, evals later — if they land on the stated schedule at all.
Did the timeline slip? Yes. Here is the math.
| Claim | Source window | Implied ship date |
|---|---|---|
| "Grok 4.7 in 4 weeks" | July 25, 2026 | ~August 22, 2026 |
| "ready in 3 to 4 weeks" | ~August 12, 2026 | ~September 2–9, 2026 (early-to-mid September) |
That is a two-to-three-week slip on 4.7, and it is honest to say so. Grok 4.6 itself landed after the July "2 weeks / ~August 8" mark — see the 4.6 launch companion for what actually shipped versus what was promised.
Do not collapse the two announcements into one roadmap. July was a cadence claim with no training details. August is a status claim: initial training done, supplemental SpaceX data in progress, 4.6 already out, 4.7 "significantly better."
What "SpaceX company data in supplemental training" actually means
Supplemental training is not a new pretrain from scratch. It is the same move SpaceXAI already used for Grok 4.5: take a trained base, then continue on a proprietary corpus (then: Cursor IDE traces; now: SpaceX company data).
What Musk is selling:
- Domain density. Rocket, manufacturing, and operations text that no public web crawl matches. That is the "awesome & unique" clause.
- A narrow win condition. He did not say 4.7 wins SWE-Bench Pro, Terminal-Bench, or chatbot arenas. He said he would be shocked if any model is better at real-world engineering.
- A compliment with a knife. Anthropic is "a great company" and will "probably release improved models soon" — then the SpaceX corpus is the reason 4.7 still wins the engineering lane.
What he did not sell, because none of it was disclosed:
- Token count, mix, or license of the SpaceX dump
- Whether ITAR, export-control, or employee-PII filters exist
- Whether the corpus is design docs, telemetry, CAD, manufacturing travelers, or Slack
- Any engineering eval (CAD questions, FEA setups, launch-ops procedures) that a third party can rerun
Until a model card or an independent suite lands, treat "SpaceX-trained" as a data-source claim, not a benchmark. explainx.ai's Grok 4.5 vs Claude Opus 4.7/4.8 comparison is still the last place we had numbers instead of adjectives.
Will Grok 4.7 beat Fable 5? Do not launder community rumors
Musk's own words: 4.7 will exceed all current models, especially at real-world engineering. That is a founder forecast.
Separately, community posts have claimed a 2.1T model that will surpass Claude Fable 5. Those posts are unverified. They are not in the quotes above. They are not an xAI spec sheet. Do not treat them as official.
If you need a current Fable / Sol baseline while you wait, use shipped coverage: Claude Fable 5 launch, GPT-5.6 Sol preview, and the GPT-5.6 vs Fable 5 comparison. How to read launch numbers without getting rolled is in the AI benchmarks guide.
Should I wait for 4.7 or use 4.6 this week?
This is the only decision the announcement actually changes.
| If you… | Do this |
|---|---|
| Design, review, or debug CAD / rockets / manufacturing-style engineering | Wait-and-eval 4.7 in early-to-mid September. The SpaceX-corpus claim is the only new information, and it is aimed at you. |
| Need a coding agent this week | Use shipped Grok 4.6 — launch evals and Cursor notes. Do not freeze a sprint on a September maybe. |
| Already standardized on Grok 4.5 | Keep it. Public numbers still live on the 4.5 launch post. Re-bench when 4.7 has a card. |
| Live in Claude Fable 5 or GPT-5.6 Sol | Stay. Musk praised Anthropic in the same post that claimed 4.7 would still win engineering. That is not a reason to rip out a working harness. |
| Run persistent agents | Grok Bot early beta is a product, not a model bump. Don't conflate Bot access with 4.7 shipping. |
| Care about the harness, not the weights | Grok Build (open source) is still the thing you can clone today. |
Copy-paste eval prompt for when 4.7 actually appears — run it on 4.6 and 4.7, same repo, same tests:
You are evaluating a candidate model for real-world engineering assistance.
Work only from the files in this repo. Do not invent vendor-specific APIs.
1. Identify the load path and the governing constraint in README + the CAD/notes folder.
2. List assumptions that are not in the repo.
3. Propose a change that reduces mass or cycle time without violating the constraint.
4. Write a test or checklist another engineer could run without the model.
5. Stop if a required drawing, material spec, or ITAR-marked file is missing — say so.
If 4.7 cannot beat 4.6 on your constraint set, the SpaceX corpus did not matter for you. That is the whole eval.
Honest limitations
- No model card. Parameter count, context window, price, and evals for 4.7 are unpublished as of August 13, 2026.
- No independent engineering bench. "Real-world engineering" is undefined. CAD? Propulsion? Factory ops? Software for vehicles? Musk did not say.
- Timeline is a founder estimate. 3–4 weeks already replaced a 4-week July date. SpaceXAI has slipped these windows before.
- Supplemental training is a black box. "Massive amount of SpaceX company data" is not a datasheet. Contamination, leakage, and export-control questions are unanswered in public.
- 4.6 vs 4.7 quality is Musk's adjective ("significantly better"), not a delta on Terminal-Bench or SWE-Bench Pro.
- 2.1T / Fable-killer rumors are not sources. Repeat them and you are amplifying fan math.
- Anthropic will move too. Musk said so. A September 4.7 drop into a static Fable/Sol field is a fantasy schedule.
Related on explainx.ai
- Grok 4.6 launch evals in Cursor (August 2026) — what already shipped; use this week
- Grok 4.6 and 4.7: Musk's July 25 timeline — the ~August 22 claim this post supersedes
- Grok 4.5 public launch — SpaceXAI, benchmarks, pricing
- Grok 4.5 vs Claude Opus 4.7 and 4.8
- Grok 4.5 private beta at SpaceX and Tesla — the earlier supplemental-training pattern
- Grok Bot early beta — persistent agents
- Grok Build open source
- GPT-5.6 vs Claude Fable 5
- AI benchmarks complete guide 2026
Primary sources: Elon Musk on X (~August 12, 2026) — the two quotes reproduced above. No SpaceXAI model card for 4.7 at time of writing.
Grok 4.7 timing, training details, and capability claims reflect Elon Musk's public statements as of August 13, 2026 and are unconfirmed by an xAI or SpaceXAI model card. Verify against official channels before migrating production workloads. Follow @explainx_ai for updates.