Grok 4.7: xAI Launches 2.1T Model at Same Low Price
SpaceXAI released Grok 4.7 on Monday, September 21, 2026, calling it “a notable improvement over Grok 4.6 at the same price and speed.” The model ships with 2.1 trillion parameters (up 40% from Grok 4.6), costs $2 per million input tokens and $6 per million output tokens, and is available immediately in Cursor, Grok Build, the xAI API, and GitHub Copilot.
The launch ends at least five public delays since late July. Elon Musk first said Grok 4.7 was “four weeks out” on July 26, then pushed it to “3 to 4 weeks,” then “10 days” on September 1, then “needs a few more days to cook” on September 11. It finally shipped eleven days after that last update.
What Changed from Grok 4.6
The core upgrade is a new, larger base model. Where Grok 4.6 and the 4.7 roadmap anticipated a substantially heavier parameter count, the final build landed at 2.1 trillion, up from 1.5 trillion in Grok 4.6 and built with a longer reinforcement learning run on a harder task mix.
xAI says the training was “weighted toward problems that take many hours to complete.” The stated outcome: the model spends more time working through difficult tasks before giving up, double-checks its own outputs more carefully, and manages longer context more reliably.
SpaceXAI also folded in supplemental training data from its sister company: Starlink satellite telemetry, SpaceX manufacturing records, and engineering failure logs. The pitch is a model with stronger reasoning about hardware and physical systems compared to models trained exclusively on internet text. That shows up in EEBench, the electrical-engineering benchmark, where Grok 4.7 scores 64%, more than seven points above Claude Fable 5.1’s 56.4%.
Benchmark Breakdown
Here is the full comparison from xAI’s launch post, alongside Artificial Analysis’ independent Intelligence Index score:
| Benchmark | Grok 4.7 | Grok 4.6 | GPT-5.6 Sol | Fable 5.1 |
|---|---|---|---|---|
| Price (input / output) | $2 / $6 | $2 / $6 | $4 / $20 | $10 / $50 |
| CursorBench 4.0 | 46.3% | 40.4% | 41.7% | 51.8% |
| DeepSWE v1.1 | 71.0%* | 65.2% | 72.7% | 70.0% |
| EEBench | 64.0% | 53.0% | 39.4% | 56.4% |
| AA Briefcase v1.1 (Elo) | 1,657 | 1,546 | 1,487 | 1,678 |
| Terminal-Bench 4.0 | 38.0% | 20.3% | 37.3% | 57.9% |
| Harvey Legal Benchmark | 19.6% | 15.8% | 2.5% | 6.7% |
| HealthBench Professional | 56.7% | 48.5% | 60.5% | 62.1% |
| AA Intelligence Index (v4.3.2) | 46 | n/a | n/a | 53 |
*DeepSWE high-effort score. Source: xAI launch post; Artificial Analysis Intelligence Index.
A few patterns stand out.
Where Grok 4.7 leads: Legal work (Harvey Legal Benchmark) and electrical engineering (EEBench). On the Harvey benchmark, Grok 4.7 scores nearly three times higher than Fable 5.1. On EEBench, it leads the entire frontier field. If your enterprise workflows involve legal document analysis or hardware-adjacent reasoning, the case for Grok 4.7 is stronger than the overall index score suggests.
Where it trails: Autonomous coding (Terminal-Bench 4.0) and long-horizon office work (AA Briefcase). The Terminal-Bench gap is especially wide: Fable 5.1 scores 57.9% versus Grok 4.7’s 38%. The Decoder’s independent analysis put the agentic coding gap between Grok 4.7 and GPT-6 Astra even wider, with GPT-6 at 60% versus Grok 4.7’s 26% on that same benchmark.
Price-performance position: At $2 input and $6 output, Grok 4.7 is five times cheaper per token than Fable 5.1 ($10/$50). For enterprise teams routing high-volume coding assist, document Q&A, or early-stage triage, the cost difference is meaningful even if Grok 4.7 is not the top performer on most tasks.
Safety and Safeguards
Grok 4.7 ships with an entirely new safeguard stack. On HackerBench v0.3, it allowed 3.3% of risky dual-use cyber prompts through while keeping legitimate security work unblocked. On LatchBio’s biosafety benchmark, it scored 62.4%, the highest any model has posted on that test. xAI says it is providing select cybersecurity partners with invite-only access to Grok 4.7’s red-team capabilities for defensive research.
This matters for enterprise procurement teams working through security review. A model with both high utility on benign tasks and low pass-through on dual-use prompts reduces the compliance surface area compared to models that block either too much or too little.
Availability
Grok 4.7 is live today on the following platforms:
- xAI API: model ID
grok-4.7, 500,000-token context, Responses API and Chat Completions - Cursor: available on all paid tiers; a fast variant (2x speed, 2x price) available in the IDE only
- Grok Build: the default coding agent model
- GitHub Copilot: rolling out across Pro, Pro+, Max, Business, and Enterprise SKUs in VS Code, JetBrains, Xcode, Eclipse, Visual Studio, and the CLI
- Gateways: OpenRouter, Vercel AI Gateway, Cloudflare Workers AI
One practical note from the official documentation: set a prompt_cache_key (Responses API) or x-grok-conv-id header (Chat Completions) on every request. Without it, multi-turn conversations route across cold servers and lose the cache, effectively paying full input price for every turn in an agentic loop.
What This Means for Enterprise AI Teams
The honest read on Grok 4.7 is “strong value, not frontier leader.” It posts a 46 on the Artificial Analysis Intelligence Index against 53 for both Fable 5.1 and GPT-6. On the most demanding agentic coding benchmarks, the gap to the top two models is still wide.
That said, three use cases have a concrete reason to consider Grok 4.7 today.
Legal and compliance workflows. A 19.6% Harvey Legal benchmark score is not just a win over GPT-5.6 Sol (2.5%) and Fable 5.1 (6.7%), it is a category-wide lead. If you are routing contract review, regulatory analysis, or compliance document work through an API model, the numbers here are difficult to ignore.
Hardware and engineering tasks. The SpaceX training data shows up in EEBench. Enterprises in manufacturing, energy, or engineering-adjacent sectors may find Grok 4.7 genuinely better on domain-specific tasks than its overall index score implies.
Cost-sensitive, high-volume routing. At five times the cost difference, routing lower-stakes workloads (first-pass triage, bulk classification, draft generation) through Grok 4.7 and reserving Fable 5.1 or GPT-6 for high-stakes final steps is a defensible architecture. The 500,000-token context window supports long-document workflows without truncation.
The roadmap Musk sketched: Grok 4.8 as a step up, Grok 4.9 targeting Fable/Astra performance class, and Grok 5 as a potential frontier leader. No dates are confirmed, and Grok 4.7 itself was delayed five times, so timeline estimates carry uncertainty. What is confirmed is that Grok 4.7 is the live frontier option from xAI right now.
For enterprise teams building agentic systems and considering model routing strategy, we can walk through the tradeoffs for your specific workflow.
Sources: xAI official launch post (primary), xAI developer docs (primary), Decrypt (secondary), The Decoder (secondary), GitHub Copilot changelog (primary), CryptoBriefing (secondary).