xAI (now shipping models under the SpaceXAI banner) released Grok 4.7 on September 21, 2026, exactly 40 days after Grok 4.6. The pitch is familiar: more capability at the same token price. The independent numbers tell a more nuanced story. Grok 4.7 scores 46 on the Artificial Analysis Intelligence Index v4.3.2, sitting in the frontier pack but a clear step below the joint leaders, GPT-6 Astra and Claude Fable 5.1, at 53 each.
Where Grok 4.7 gets genuinely interesting is the economics: bargain token rates, but also one of the highest output-token counts per task of any model tested, which quietly inflates real-world costs.
Here is the full breakdown, including the complete Artificial Analysis domain and benchmark data, what xAI’s own launch table claims, and how Grok 4.7 compares to its predecessor, the top-tier leaders, and Xiaomi’s MiMo V2.6 Pro, which matches its intelligence score at a fraction of the price.
What Is Grok 4.7 and What’s New?
Grok 4.7 is xAI’s new flagship model for long-running coding and knowledge work. According to the official launch post, it uses a larger base model than Grok 4.6 (widely reported as a 2.1 trillion parameter architecture, though xAI has not confirmed the figure) and received a longer reinforcement-learning run weighted toward problems that take hours rather than minutes. xAI also trained it natively on the Grok Bot agent harness, targeting conversations, research, and document production outside of pure software repositories.
The developer documentation confirms the core specs:
- Context window: 500,000 tokens
- Knowledge cutoff: May 2026
- Modalities: text and image input, text output
- Tooling: function calling, web search, X search, code execution
- Reasoning effort: low, medium, high, xhigh
Availability is broad for a day-one launch: the xAI API, Cursor (across all plans), Grok Build (where it is now the default model), OpenRouter, Vercel, and Cloudflare. A fast variant at twice the token price and twice the output speed exists but is limited to Cursor and Grok Build.
There is also a US regional API endpoint carrying a 10% token-price premium.
The Full Artificial Analysis Breakdown
Most launch coverage cites Grok 4.7’s headline Intelligence Index score and stops there. The full Artificial Analysis profile (all figures at xhigh reasoning effort) shows where the model is actually strong, and where it visibly struggles. The table below lines Grok 4.7 up against three rivals benchmarked in the same window: Xiaomi’s MiMo V2.6 Pro, Kimi K3, and the Step 5 Preview. Cells marked with a dash have no published result for that model yet.
| Benchmark | Grok 4.7 | MiMo V2.6 Pro | Kimi K3 | Step 5 Preview |
|---|---|---|---|---|
| AA Intelligence Index v4.3.2 | 46 | 46 | 44 | 44 |
| Economics Index | 56 | 53 | 53 | 53 |
| Legal Index | 54 | 45 | 49 | 46 |
| Strategy & Ops Index | 54 | 48 | 50 | 47 |
| Finance & Accounting Index | 52 | 46 | 47 | 45 |
| Engineering Index | 47 | 47 | 43 | 45 |
| Healthcare & Medical Index | 47 | – | – | – |
| AA-LCR v1.1 | 77% | 86.3% | 88.7% | 88.3% |
| AutomationBench-AA | 66% | 58.6% | 58.3% | 51.0% |
| GDPval-AA v2.1 | 60% | 59% | 51% | 53% |
| AA-Briefcase v1.1 | 58% | 54.7% | 51.0% | 47.8% |
| SciCode | 57% | 60.9% | 59.5% | 58.9% |
| Humanity’s Last Exam | 43% | 49.4% | 46.9% | 46.5% |
| Terminal-Bench 4.0 | 26% | 34.8% | 12.6% | 33.3% |
| GDP.pdf | 20% | 19.2% | 22.0% | 14.8% |
| CritPt | 18% | 26.6% | 23.4% | 20.9% |
| MLCR-AA | 15% | – | – | – |
| AA-Omniscience Accuracy | 47% | 35% | 48% | 42% |
| AA-Omniscience Non-Hallucination Rate | 71% | 59% | 47% | 57% |
The pattern is clear: Grok 4.7 is a strong professional-knowledge model, particularly in economics, legal, and business automation, but it stumbles on deep multi-step agentic tasks and frontier-stress evaluations. Its 71% non-hallucination rate on AA-Omniscience, meanwhile, means it still confidently fabricates roughly three in ten answers.
The 26% Terminal-Bench 4.0 score deserves special attention because xAI’s own launch table reports 38.0% for the same benchmark. The difference comes down to harness and reasoning settings, and it is a useful reminder that vendor tables and independent testing do not measure quite the same thing.
The Verbosity Problem: Cheap Tokens, Expensive Tasks
xAI kept Grok 4.6’s token pricing: $2 per million input tokens, $6 per million output, and $0.50 per million cached input, with rates doubling for prompts above 200,000 tokens. On paper, that is among the cheapest frontier-model rates available, roughly a fifth of what GPT-6 Astra and Claude Fable 5.1 charge at $10/$50.
Token price, however, is not what you pay. Task cost is. And Grok 4.7 is one of the most verbose models in Artificial Analysis testing, generating more output tokens per Intelligence Index task than almost anything else in the database. The result is roughly $3.74 in compute per Intelligence Index task.
That is cheaper than Qwen 3.8 Max, cheaper than Claude Opus 5 at higher reasoning levels, and cheaper than Claude Fable 5.1 at high or max effort (where Fable runs about $7.63 per task). But Claude Fable 5.1 at medium reasoning or lower undercuts it, and GPT-6 Astra’s low token use keeps Astra near $3.26 per task despite its far pricier tokens. The same trap shows up in Anthropic’s budget tier, where Claude Haiku 5.5 generated 440 million output tokens on the same index against a 100 million median.
Two factors drive Grok 4.7’s effective cost above its sticker price. The first is confirmed: raw verbosity. The model writes long, which means even “cheap” tokens add up across a multi-hour agent run. The second is speculation, not confirmed: a potentially lower cache-hit rate in agent workloads, where long, highly variable contexts may reuse fewer prompt prefixes. Watch real workloads on this one.
Cursor’s independent testing paints a consistent picture. At high effort, Cursor measured Grok 4.7 at 43.9% on CursorBench 4.0 with an average task cost of $4.69, versus $5.20 for Grok 4.6.
Claude Fable 5.1 and Claude Opus 5 scored higher (49.2% and 44.7%) but cost about $9 per task. Grok 4.7 at xhigh hits 46.3% at $6.01 per task. The model lives on the price-performance frontier, but it earns that position through low token prices rather than token efficiency.
Grok 4.7 vs Grok 4.6: What Actually Improved
xAI’s launch table reports gains on all seven of its tracked benchmarks. The standouts, measured at xhigh against Grok 4.6 at high effort:
- Terminal-Bench 4.0: 38.0%, up from 20.3% (the biggest jump, though harness-dependent as noted)
- EEBench (electrical engineering): 64.0%, up from 53.0%
- CursorBench 4.0: 46.3%, up from 40.4%
- Harvey Legal Agent Benchmark: 19.6%, up from 15.8%, and the strongest legal-agent score in xAI’s table, ahead of GPT-5.6 Sol (2.5%) and Fable 5.1 (6.7%)
- DeepSWE v1.1: 71.0% at high effort, versus 65.2% for Grok 4.6
Since the effort settings differ between the two columns, the cleanest comparison is Cursor’s like-for-like high-effort run: 43.9% versus 40.4%, at a slightly lower average task cost.
That is a real but modest generational gain. The bigger story is that Grok 4.7 sustains long agent runs better, which xAI credits to its longer reinforcement-learning training. The safety stack is also entirely new, with xAI reporting 62.4% on LatchBio’s biosafety benchmark and a 3.3% dual-use prompt leak rate on its own HackerBench v0.3 test, both vendor-reported figures.
Where Grok 4.7 Lands: Frontier, But Not on the Podium
The top of the market is settled. GPT-6 Astra and Claude Fable 5.1 share first place on the Intelligence Index at 53 points, and both price at $10 per million input tokens and $50 per million output. That podium has since been rearranged: Anthropic’s Claude Opus 5.5, released on September 22, took first place on the Intelligence Index and undercut both models by 60% on list price per token.
As we covered in our GPT-6 versus Claude Fable 5.1 comparison, those two trade blows by specialty: Fable leads knowledge and factual accuracy, Astra leads math, computer use, and cybersecurity.
Grok 4.7 is not competing with them for raw capability, and the agentic-coding gap is stark: 26% on Terminal-Bench 4.0 under independent testing versus 60% for GPT-6 Astra and 55% for Claude Fable 5.1. Even the much cheaper DeepSeek V4.1 Flash edges past it at 27%.
But Grok 4.7 is unquestionably a frontier model in the tier below, and it reaches that tier at about one-fifth the token price of the leaders and a lower cost per task than either of them at serious reasoning effort. For teams that want frontier-class professional work without flagship-tier spending, that trade is the entire value proposition.
Same Intelligence Score, Different Price: Xiaomi MiMo V2.6 Pro
The most striking comparison of the week is not with the models above Grok 4.7, but with one that landed within the same 24-hour window. Xiaomi’s MiMo V2.6 Pro also scores 46 on the Artificial Analysis Intelligence Index v4.3.2, matching Grok 4.7 exactly, and it is the top-ranked Chinese model and top open-weights model globally.
The economics diverge dramatically. MiMo V2.6 Pro charges roughly $0.43 per million input tokens and $0.87 per million output tokens, works out to about $0.13 per Intelligence Index task, and runs at 134 tokens per second with a 1M-token context window.
Grok 4.7’s $3.74 per task is more than 28 times higher for the same headline intelligence. The models are not identical in profile: Grok 4.7 is stronger in professional and legal work, while MiMo trades on efficiency and open weights. But if raw intelligence per dollar is the criterion, MiMo V2.6 Pro resets the math. The other open-weight flagship of the season is covered in our Hy4 Preview vs MiMo V2.6 Pro comparison, which weighs Tencent’s entry against the same price-per-intelligence bar.
We cover the model in depth in our full Xiaomi MiMo V2.6 Pro breakdown.
Should You Use Grok 4.7?
Grok 4.7 is a situational recommendation rather than a default one.
- For long agentic coding on a budget: yes, with caveats. Cursor’s data shows solid price-performance at $4.69 to $6.01 per task, and it beats GPT-5.6 Sol on CursorBench at lower cost. Just do not expect it to match Fable 5.1 or Astra on raw agentic depth.
- For legal, economics, and business-automation work: genuinely compelling. The Harvey legal score leads xAI’s comparison table outright, and AutomationBench-AA at 66% and Economics at 56 are strong independent results.
- For hallucination-sensitive applications: hold off. A 71% non-hallucination rate on AA-Omniscience leaves meaningful room for confident errors.
- For intelligence per dollar above all: look at MiMo V2.6 Pro first. The wider Chinese AI pricing gap explains why the cost floor keeps dropping.
The bottom line: Grok 4.7 buys you a real frontier model at mid-market prices, but verbosity means the sticker price understates the bill, and the intelligence ceiling is a full tier below the GPT-6 Astra and Claude Fable 5.1 class.
For the right workload, that is a trade worth making, and the Step 5 preview, DeepSeek V4.1 Flash and MiMo-V2.6-Pro suggest the price floor will keep falling.
