What Is Grok 4.7? The 2.1T xAI Model Explained

Inside Grok 4.7: The 2.1T-Parameter xAI Model Trading Speed for Token Efficiency

Grok 4.7 is an xAI large language model that launched on September 21, 2026, for coding and knowledge work. The release describes a larger base model than Grok 4.6, longer training on difficult multi-hour tasks, stronger self-checking, and improved long-running agent capabilities. The model launched at the same headline price and speed as Grok 4.6. The 2.1 trillion parameter figure associated with Grok 4.7 remains a founder-stated specification rather than a fully confirmed technical detail. The model was first surfaced through Musk's own X posts on July 24, 2026 and July 28, 2026, before the September 21 launch.

Updated 2026-09-21: Grok 4.7 has launched, with access reported through the xAI API, Grok Build, Cursor, OpenRouter, and other rollout channels. The launch also exposed low, medium, high, and xhigh reasoning variants.

Key Takeaways

• Grok 4.7 launched on September 21, 2026, after the earlier September target slipped.
• The model uses a larger base model than Grok 4.6, with longer reinforcement learning on difficult, multi-hour tasks and improved self-checking.
• The 2.1 trillion parameter figure remains associated with Grok 4.7, but the supplied launch material does not independently confirm the exact count.
• Grok 4.7 launched at the same headline price and speed as Grok 4.6. Launch coverage reports API pricing of $2 per million input tokens and $6 per million output tokens.
• Reported model-card results include 20.3% to 38.0% on Terminal-Bench 4.0, 31.9% to 46.0% on SWE-Marathon, 48.5% to 56.7% on HealthBench Pro, 15.8% to 19.6% on Legal Agent, and 60.0% to 66.0% on EEBench.
• Additional reported results include 40.4% to 46.3% on CursorBench 4.0 and 65.2% to 71.0% on DeepSWE v1.1.
• Grok 4.7 is available through the xAI API, Grok Build, Cursor, and OpenRouter. Its rollout into other Grok surfaces is not fully documented in the supplied material.
• Artificial Analysis gave Grok 4.7 an Intelligence Index score of 46, up from 44 for Grok 4.6 in the cited post-launch comparison.

What Is Grok 4.7?

Grok 4.7 is the launched next iteration of xAI's Grok flagship line, following Grok 4.5 and Grok 4.6. It launched on September 21, 2026, as a model aimed particularly at coding, knowledge work, and longer-running agents. Launch coverage describes a larger base model than Grok 4.6 rather than merely another post-training pass on the same base.

The release materials emphasize a longer reinforcement-learning run on harder tasks that can take hours to complete. They also highlight better self-checking, improved handling of long context, stronger document and presentation creation, and native understanding of the Grok Bot harness for general knowledge work. The model includes a rebuilt safeguard system described as more resistant to jailbreaks and less prone to unnecessary refusals.

The 2.1 trillion parameter figure remains part of the model's public description, compared with 1.5 trillion for Grok 4.6. However, the supplied launch evidence still treats that figure as unconfirmed rather than as a fully documented xAI specification. The same caution applies to claims about the exact amount and measurable effect of SpaceX or Starlink engineering data used in training.

The launch itself resolves the earlier availability question. Grok 4.7 was first reported in the xAI API on September 21 and then appeared in Grok Build, Cursor, and OpenRouter. It is a hosted, closed model rather than an open-weight release.

For readers tracking the lineage, our earlier breakdown of the Grok 4.5 canary traces and 1.5T V9 foundation covers the base model that Grok 4.6 iterates on.

Grok 4.7 at a Glance

Attribute	Value
Developer	xAI
Type	Large language model for coding, knowledge work, and reasoning
Parameters	2.1 trillion associated with the model, but not independently confirmed in the supplied launch material
Predecessor	Grok 4.6, released in August 2026 with a reported 1.5T scale
Modality	Launch coverage discusses coding, knowledge work, and multimodal work; the complete modality specification is not supplied
Context window	500K reported by a tester; the official context specification remains unclear
Pricing	Same headline price as Grok 4.6; launch coverage reports $2 / M input and $6 / M output tokens
API model ID	Available in the xAI API catalogue; grok-4.7 was also reported in third-party integrations
Serving speed	Same headline speed as Grok 4.6, according to launch coverage
Token efficiency	Improved token efficiency was a pre-launch and community claim; the supplied launch material does not provide a complete efficiency measurement
License	Proprietary and not open-weight
Availability	Launched September 21, 2026; available through the xAI API, Grok Build, Cursor, and OpenRouter
Official model card	Benchmark-card results are reported; a complete public technical specification is not supplied

How Grok 4.7 Works and What Makes It Different

The original pre-launch framing centered on a tradeoff: a larger model would be slower to serve but more token-efficient. The launch changed the practical description. SpaceXAI and launch coverage describe Grok 4.7 as running at the same headline speed and price as Grok 4.6 while using a larger base model and longer training on difficult, multi-hour tasks.

The most concrete design difference is therefore not a published parameter count but the training and behavior description. Grok 4.7 is said to stay on difficult problems longer, check its own work more carefully, and handle long-running agentic workflows more effectively. Those claims align with the reported gains on Terminal-Bench, SWE-Marathon, Legal Agent, CursorBench, and DeepSWE.

The reported model-card results show the largest cited changes on long-running coding and agentic evaluations. Terminal-Bench 4.0 rose from 20.3% for Grok 4.6 to 38.0% for Grok 4.7. CursorBench 4.0 rose from 40.4% to 46.3%, while DeepSWE v1.1 rose from 65.2% to 71.0%. A separate report puts Grok 4.7 at 19.6% on the Harvey Legal Agent Benchmark.

These results should still be read with care. The supplied research notes that independent validation is incomplete, and a separate discussion reported a mismatch between the headline Terminal-Bench result and an independent rerun using a different harness. The benchmark gains are therefore useful launch evidence, not a complete verdict on every workload.

The larger-base-model claim is supported by launch coverage describing Grok 4.7 as more than additional post-training on Grok 4.6. The exact architecture remains undisclosed: the supplied material does not establish whether the model is dense or mixture-of-experts, how many parameters are active during inference, or how much of the reported 2.1T figure represents active versus total parameters.

The launch also exposed multiple reasoning levels: low, medium, high, and xhigh. Community testing suggests that xhigh can use substantially more reasoning tokens than the comparable Grok 4.6 setting. That may help on difficult tasks, but it also means headline token prices do not by themselves describe the total cost of a completed task.

What You Can Do With Grok 4.7

Grok 4.7 now has concrete shipped use cases rather than only predecessor-based projections. The launch and early testing point most strongly toward:

• Coding and agentic loops — Grok 4.7 is available in Cursor and Grok Build, with reported gains on CursorBench, Terminal-Bench, and DeepSWE. It is designed to keep working through longer computer tasks rather than ending early.
• Long-running technical work — the model was trained on harder tasks that can take hours, and launch coverage emphasizes longer-running agents, better self-checking, and improved long-context handling.
• Knowledge work and document creation — the launch description highlights improved document and presentation creation, together with native understanding of the Grok Bot harness.
• Legal and engineering workflows — reported gains on the Harvey Legal Agent Benchmark and EEBench suggest particular improvements in legal and electrical-engineering evaluations, although the supplied evidence does not establish broad superiority across all professional tasks.
• Cost-sensitive API workloads — Grok 4.7 launched at the same headline price as Grok 4.6. The reported $2 / $6 input-output rates put it below the cited token prices of several competing frontier models.

Early hands-on results are mixed outside the strongest coding and agentic signals. Some testers reported better SVG and 3D work in Grok Build, while others criticized frontend and user-interface output. The supplied research also notes divided sentiment on non-coding and multimodal tasks, so teams should evaluate their own workloads rather than assume that benchmark gains transfer uniformly.

Teams building long-context chat and reasoning workloads can now test a shipped Grok generation such as Grok 4.6 or compare it directly with Grok 4.7 where their platform provides access.

How Grok 4.7 Compares

The comparison table below uses the launch and post-launch evidence available in the supplied record. Unconfirmed specifications remain labeled as such.

	Grok 4.5 (shipped)	Grok 4.6 (released Aug. 2026)	Grok 4.7 (launched Sep. 21)	Kimi K3 (shipped)
Parameters	1.5T	1.5T	2.1T associated with the model, unconfirmed	2.8T
Weights	Proprietary	Proprietary	Proprietary and not open-weight	Open
Context window	500K tokens	Not confirmed in the supplied record	500K reported by a tester; official figure unclear	Not confirmed
Input price	$2 / M tokens	$2 / M tokens	Same headline price as 4.6; launch coverage reports $2 / M tokens	See vendor
Output price	$6 / M tokens	$6 / M tokens	Same headline price as 4.6; launch coverage reports $6 / M tokens	See vendor
Artificial Analysis Intelligence Index	Not applicable to the cited comparison	44 in the cited post-launch comparison	46 in the cited post-launch comparison	Not provided in the supplied record
Availability	Shipped	Shipped	xAI API, Grok Build, Cursor, OpenRouter	Shipped

For the open-weight side of this comparison, our Kimi K3 release breakdown covers Moonshot's 2.8T model in detail, and the Kimi K3 vs Claude analysis puts it against Opus 4.8. On the closed-frontier side, the What Is Claude Opus 5 overview covers the model Musk name-checked in the earlier positioning discussion.

Availability: How to Access Grok 4.7

Grok 4.7 launched on September 21, 2026. The initial rollout exposed the model in the xAI API, Grok Build, and Cursor. OpenRouter also announced that Grok 4.7 was live, providing another API access path.

The confirmed access paths in the supplied launch material are:

• The xAI API, where Grok 4.7 appeared in the authenticated model catalogue.
• Grok Build, where both Grok 4.7 and Grok 4.7 Fast were reported in the model selector.
• Cursor, where the model was released for coding and agentic work.
• OpenRouter, which announced the model as live for coding and knowledge work.
• Vercel AI Gateway, where a 40% promotion was offered through September 27, 2026.

Initial rollout reports said Grok 4.7 was not yet available in Grok Chat at the time it appeared in the API and Grok Build. The supplied material does not establish the final rollout status for every consumer-facing Grok surface, so access should be checked at the product level rather than assumed from the API launch.

The model is not open-weight and cannot be run locally according to launch coverage. The available evidence instead points to hosted access through xAI and partner platforms.

What We Don't Know Yet

The launch resolved the biggest status, availability, and headline-pricing questions, but several technical questions remain open:

• Exact parameter count. The 2.1T figure is repeatedly associated with Grok 4.7 and attributed to Musk, but the supplied launch material still does not independently confirm it.
• Full context specification. A tester reported a 500K context window, while the official context details are not included in the supplied launch evidence.
• Architecture details. Dense versus mixture-of-experts, active versus total parameters, and the precise relationship between the larger base model and the 2.1T figure remain undisclosed.
• Fast-variant pricing and speed. Grok 4.7 Fast is present in Grok Build, but the supplied reports disagree or remain incomplete about its exact speed and price relationship to the standard variant.
• Independent benchmark reproduction. The supplied record contains model-card figures, Artificial Analysis results, and community tests, but independent methodology and reruns are not consistent across every benchmark.
• Broad multimodal performance. Earlier roadmap commentary said multimodal performance still needed work, and post-launch reports continue to criticize some frontend and visual tasks. The evidence does not establish a consistent performance profile across non-coding and multimodal workloads.
• Grok Bot routing. The supplied material does not definitively identify whether every Grok Bot interaction uses Grok 4.7.

The launch has replaced the earlier model-card and release-date uncertainty with a testable product. The remaining questions are mostly about the exact technical specification, variant economics, and how consistently the model's benchmark gains translate to real-world work.

Frequently Asked Questions

What is Grok 4.7?

Grok 4.7 is an xAI large language model launched on September 21, 2026, for coding and knowledge work. It uses a larger base model than Grok 4.6, was trained longer on difficult multi-hour tasks, and is described as more capable at self-checking and long-running agentic work. The 2.1 trillion parameter figure remains a founder-stated, unconfirmed specification.

When will Grok 4.7 be released?

Grok 4.7 launched on September 21, 2026. It is available through the xAI API, Grok Build, Cursor, and OpenRouter. Initial rollout reports also said it was not yet available in Grok Chat.

How many parameters does Grok 4.7 have?

Grok 4.7 is associated with a 2.1 trillion parameter count, compared with 1.5 trillion for Grok 4.6, but the supplied launch material does not independently confirm the exact parameter count. Treat 2.1 trillion as a founder-stated specification rather than a fully documented technical figure.

How is Grok 4.7 different from Grok 4.6?

Grok 4.7 uses a larger base model, longer reinforcement learning on difficult multi-hour tasks, and improved self-checking and long-context handling. It brings stronger coding, longer-running agents, and better document and presentation creation while launching at the same headline price and speed as Grok 4.6.

How much will Grok 4.7 cost?

Grok 4.7 launched at the same headline price as Grok 4.6. Launch coverage reports API pricing of $2 per million input tokens and $6 per million output tokens, although the supplied material does not establish a complete pricing and speed schedule for every variant.

Is Grok 4.7 open source?

No. The supplied launch coverage describes Grok 4.7 as a closed, non-open-weight model that cannot be run locally. It is accessed through hosted products and APIs.

How does Grok 4.7 compare to Kimi K3?

Grok 4.7 has a post-launch Artificial Analysis Intelligence Index score of 46, while the supplied record does not provide a directly comparable current score for Kimi K3. Kimi K3 is described as a 2.8 trillion parameter open-weight model from Moonshot AI, while Grok 4.7's 2.1 trillion figure remains unconfirmed.

What to Watch Next

Four signals will determine how the launch settles into the model landscape. First, independent reproductions of the model-card and Artificial Analysis results will show whether the large Terminal-Bench, CursorBench, and DeepSWE gains hold across different harnesses. Second, broader access information will clarify the rollout into Grok Chat and other consumer-facing products. Third, xAI or its distribution partners may publish the exact context, fast-variant pricing, and parameter specifications. Fourth, Grok 4.8 is described as a noticeable improvement, but the supplied material gives no confirmed release date or technical specification for it.

The immediate practical question is less whether Grok 4.7 exists and more where it performs best. The strongest current signals are coding, long-running agents, legal evaluation, and cost-sensitive developer workflows. Frontend, multimodal, and general-purpose quality remain more contested in early testing.

Update — 2026-08-13

A August 12, 2026 post puts Grok 4.7 "3 to 4 weeks" out, which would push the arrival into early-to-mid September rather than the late-August floor implied by the original "few weeks after Grok 4.6" framing. The same post now treats Grok 4.6 as a shipped, "basically Sol level" model and reiterates the previously reported drivers — more SpaceX data, more parameters, and additional training — without adding any vendor documentation. As before, none of this originates from an xAI model card, so the release-window number remains a founder-adjacent estimate, not a committed date.

Alongside the timeline chatter, a circulating YouTube short frames Musk as claiming Grok 4.7 will beat OpenAI and Anthropic. That is a competitive-positioning assertion, not a benchmark: no SWE-Bench, GPQA, or independent per-task cost figures accompany it, so it sits in the same unverified tier as the Pareto-frontier claim the article already flags. The signals to watch remain unchanged — a real 4.6 ship date, a first xAI model card or pricing page, and third-party scores that can test the token-efficiency pitch.

Update — 2026-09-02

On September 2, Elon Musk said directly that Grok 4.7 would be released “in 10 days,” a signal that points to roughly September 12 but still does not establish a guaranteed launch date, rollout scope, or access method. Commentary from Mark K and Nima Owji adopts the same September 12 interpretation, while Musk's post is the primary timing signal.

Additional commentary claims that initial training is complete and that xAI is performing supplemental training with SpaceX data, but no official xAI specification confirms the corpus, its status, or its effect. Mark K also attributes improved coding and writing to the expected release; those remain pre-release claims rather than measured results.

Update — 2026-09-11

On September 7, a reported error from the official Grok Bot named grok-4-7-0907 as an invalid model, providing the first possible model-ID trace for Grok 4.7. LuminaBench noted that xAI used a similarly dated identifier for Grok 4, while Mark K interpreted the trace as stealth testing.

The identifier is a stronger pre-release signal than timing speculation, but it does not prove that the build is final or publicly available. No release notes, pricing, reproducible benchmark results, or hands-on evaluation had surfaced alongside it.

A September 11 post from @Mr_Salio claims Grok 4.7 could be up to 10× cheaper, adding a prospective cost-competition angle to the release discussion. The post gives no baseline, token rates, or access terms, so this is not a price announcement and pricing remains unknown.

Update — 2026-09-14

The September 12 target passed without a release. Elon Musk reportedly said Grok 4.7 needs “a few more days to cook,” with Ananth7e and Rohan Paul documenting the delay. A later community report claimed another week, but xAI has not provided a replacement date or confirmed rollout details.

Several posts attribute the delay to a possible reinforcement-learning problem: the model may end difficult tasks too early and fail to check its work adequately. 0xLogicrw presented this as Musk's tentative explanation, possibly linked to excessive penalties on response length during training; the cause remains unconfirmed.

Update — 2026-09-16

Elon Musk has for the first time given Grok 4.7 a concrete capability anchor, and it is a more modest one than earlier "beats everyone" framing implied. In a September 14 roadmap reply, Musk said Grok 4.7 should roughly match Claude Opus 5 — "better in some ways, worse in others" — with multimodal performance still needing work, as relayed by @mark_k, @AGTPinsights, and @notjazii. This is still a verbal comparison, not a measured benchmark, and it sits below the earlier Pareto-frontier and "exceeds Fable and Astra combined" hype the article already flags — a tempering that some commenters read as disappointing.

The same roadmap extends past 4.7: Musk positions Grok 4.8 as a "noticeable improvement," Grok 4.9 as potentially Astra/Fable-class, and Grok 5 as possibly ahead of everything else. None of these carry dates, specs, or documentation, so they remain founder-stated trajectory rather than committed releases.

As of mid-September, Grok 4.7 still had not shipped. @mark_k noted on September 15 that it "might" arrive that day "unless something causes further delays," extending the slipping-date pattern the article already documents. No release notes, pricing, context-window figure, or independent benchmark had surfaced.

Update — 2026-09-20

New pre-release evidence now points to internal evaluation rather than a public launch. LuminaBench reported that grok-4.7-build appeared at the xhigh setting across multiple benchmark evaluations before the results were removed under an embargo. A separate purported KernelBench chart showed mixed results, with Grok 4.7 trailing Opus 5.0 in the post's comparisons and falling substantially behind on several kernels. These incomplete leaks provide no methodology or final-model confirmation.

Deployment signals also accumulated on September 17: Google Cloud quota entries and a base_model:grok-4.7 listing reported by LuminaBench suggest that an integration may be in preparation. Quota visibility is not public availability, however, and the model still has no confirmed launch, published benchmark suite, or official access terms.

Update — 2026-09-21

On September 21, LuminaBench reported that an xAI-authenticated model catalogue briefly exposed grok-4.7 with low, medium, high, and xhigh reasoning variants. The listing adds specificity to the earlier infrastructure traces, but it does not establish public access, pricing, or a completed launch. NFT Chen's summary likewise said no official xAI announcement had appeared.

Building similar long-context chat and reasoning workloads today? On kie.ai you can try Grok 4.6, Grok 4.5, and Gemini 3.7 Flash.