Grok 4.7 shipped 21 September 2026 — what Grok Bot builders can use today

SpaceXAI released Grok 4.7 on 21 September 2026 as API id grok-4.7. It is live on the API, in Cursor, and as the default in Grok Build. Consumer Grok chat and running Grok Bots are not confirmed on the new model.

Botski is an independent Grok Bot workshop. This note is a visitor briefing of public launch materials. It is not SpaceXAI, xAI, Grok, or Cursor, and it is not endorsed by them. Nothing on bot.ski flips to a new model because of this post. LIVE_AI stays off here.

What shipped

Official product name: Grok 4.7. API model id: `grok-4.7`. A faster serving tier is named Grok 4.7 Fast.

Launch pages brand the company as SpaceXAI. The model card states that SpaceXAI is a DBA of XAI LLC and that xAI and SpaceXAI may be used interchangeably. Product URLs remain on x.ai.

Confirmed on launch day:

• xAI / SpaceXAI API at console.x.ai, slug grok-4.7, via Responses API and Chat Completions
• Grok Build (docs call 4.7 the default)
• Cursor, on every plan tier
• Microsoft Word, PowerPoint, and Excel add-ins (model card: 4.7 is the default there)
• Named gateways: OpenRouter, Vercel, Cloudflare, Snowflake, Databricks Mosaic, and others
• US regional API endpoint https://us.api.x.ai/v1 (US inference; token use billed at a 10% premium)

@SpaceXAI (21 September 2026, 16:17 UTC): “Grok 4.7 is here. It's a notable improvement over Grok 4.6 at the same price and speed.” Follow-up in the same thread: available now in Cursor, Grok Build, and the Grok API.

Not on the official launch list:

• Consumer Grok web and mobile apps
• Grok-in-X
• A stated cutover for running Grok Bots

The model card says SpaceXAI plans to add Grok 4.7 to consumer surfaces later. The Grok Bot marketing page at x.ai/bot still lists “Grok 4.6 model” on SuperGrok / Plus / Heavy cards.

What’s new

Versus Grok 4.6, the launch post says 4.7 uses a new, larger base model and a longer reinforcement-learning run on a harder mix of tasks, weighted toward problems that take many hours. SpaceXAI says the model is better at verifying its own work, managing longer context, and creating documents and presentations.

The line that matters for this site: SpaceXAI “trained Grok 4.7 to natively understand the Grok Bot harness, making it better at conversational tasks and general knowledge work.” That is a training claim, not a statement that existing Bots now run 4.7.

Safeguards: the launch post says 4.7 was built with an entirely new safeguard stack and is the strongest Grok they have tested on refusals and jailbreak resistance, while keeping refusal rates low for legitimate security and research work. Treat those as company claims.

API specs from docs.x.ai (accessed 21 September 2026 CT):

• Context window: 500,000 tokens (same figure published for Grok 4.6)
• Input: text and images. Output: text only. No text output limit stated
• Reasoning effort: low, medium, high (default), xhigh
• Tools: function calling, web search, X search, code execution
• Knowledge cutoff on the API docs page: May 2026 (model card wording differs — see Caveats)
• Responses API: grok-4.7 always returns reasoning.encrypted_content, even when include does not list it. Pass those items back unchanged on the next turn. Chat Completions is unchanged
• Batch API: model detail page marks Batch as not supported for grok-4.7

Company-published scores vs Grok 4.6 High (launch post table):

\*DeepSWE marked high-effort in the launch post. These are SpaceXAI’s numbers, not an independent rerun.

Elon Musk the same day: “Grok 4.7 is a strong combination of intelligence, speed & low cost.” Separate posts point builders at the Build harness for best results.

Availability and cost

API list price matches Grok 4.6.

• Below 200k prompt tokens, per 1M tokens: $2.00 input / $0.50 cached input / $6.00 output
• At or above 200k prompt tokens: $4.00 / $1.00 / $12.00 for the whole request

Grok 4.7 Fast is the same model on faster infrastructure, billed at twice those standard rates (docs: $4 / $1 cached / $12 below 200k; $6 / $1.50 / $18 at or above 200k). Fast is available only in Cursor and Grok Build. It is not on the public xAI API and is not in Grok Build’s free tier.

US regional endpoint: same model ids (grok-4.7 and grok-4.6 only, as of this check), 10% token premium. Image, video, and voice APIs are not on that endpoint. Regions outside that US guarantee: not stated. The global endpoint https://api.x.ai may route between regions.

Consumer-plan pricing for Grok chat is not restated in the 4.7 launch materials. Grok Bot access remains tied to paid Cursor plans or a linked SuperGrok / Plus / Heavy subscription; Bot docs still say usage resets weekly. No new Bot rate-limit table shipped with 4.7.

For Grok Bot builders

Fact. SpaceXAI says 4.7 was trained to natively understand the Grok Bot harness — their stated reason for better conversational and knowledge-work behavior, not only code.

Fact. Grok Bot still has no customer-facing model picker. Docs (overview and settings pages last updated 18 September on this check): “Cursor manages model selection.” The security FAQ adds that the serving mix can change, a team model allowlist is Enterprise-only and not guaranteed, and usage analytics show the model that served each request, including failovers. Billing follows the serving model.

Fact. As of 21 September 2026 CT, x.ai/bot plan copy still says “Grok 4.6 model.” Overview, skills/routines, approvals, and computer docs were not rewritten as a 4.7 runtime changelog.

Fact (unchanged Bot machinery). Each user gets a persistent Cursor-hosted cloud computer (browser, filesystem, terminal) running in the United States today. Skills can be saved from a demonstrated path and rerun on a schedule. Drafts and outbound actions still go through approval. Computers are not on-prem. There is no published “4.7 changed your weekly cap” note.

Speculation — labeled. Existing Bots may already be routed to 4.7 with no picker and no marketing-page update. They may still be on 4.6. Failover mixes are possible. Do not assume a flip. Check the usage analytics model name on a live Bot request before you rewrite prompts around 4.7 behavior.

What to try today without guessing the Bot runtime:

1. Call grok-4.7 on the API. For Responses API multi-turn loops, pass reasoning.encrypted_content back unchanged. 2. In Cursor, pick Grok 4.7 (and Fast only if you accept 2× rates). Grok 4.6 remains available there. 3. In Grok Build, treat 4.7 as the documented default. 4. For a live Grok Bot: trust the model recorded on the request, not the launch blog alone. 5. If a Bot starts refusing work it used to take, treat the new safeguard stack as a plausible cause and log the prompt.

Botski is not the installer, not the marketplace, and not a live Grok endpoint.

Caveats

• Launch benchmarks and safety scores are company-published. Independent labs (including Artificial Analysis) posted their own numbers the same day — useful context, not a substitute for the official table above.
• Two official cutoff wordings disagree. API docs: knowledge cutoff May 2026. Model card: pretraining cutoff June 2026, with supplemental training data as late as August 2026. Do not collapse those into one date.
• EEBench also disagrees across official artifacts: 64.0% on the launch-post table (used above), 66.0% at xhigh in the model card.
• Parameter counts circulating today (including “2.1 trillion”) are not in the launch post, the API docs, or the model card. Left out until SpaceXAI publishes a number.
• Secondary writeups that say 4.7 is already in the consumer Grok app conflict with the model card. Trust the card.
• Token use may be higher even when list price is unchanged — a third-party observation from Artificial Analysis, not an official pricing change.
• Rate limits are described as team-specific. Do not publish a global RPM/TPM figure from one console view.

Links

• Launch: Introducing Grok 4.7
• Docs: Grok 4.7 · Models · Release notes
• Bots: x.ai/bot · Overview · Settings · Security FAQ
• On Botski: Plan your bot · FAQ hub

Disclaimers and sources

• Not official xAI. Botski is an independent project. These notes help visitors read public launch pages. They are not official SpaceXAI / xAI / Grok / Cursor documentation.
• No runtime claim for your Bot. Check usage analytics before you assume grok-4.7 is serving a live Grok Bot.
• Sources dated. Launch materials and docs accessed 21 September 2026 (America/Chicago). Claims above track the linked pages in the sources list.