Grok 4.7 Pushes SpaceXAI Into the Top Four AI Labs: Benchmarks, Pricing and the Token-Cost Catch
Posted on 22nd Sep 2026 06:05:07 in Artificial Intelligence, Machine Learning
Tagged as: Grok 4.7, xAI, SpaceXAI, AI models, coding agents
SpaceXAI released Grok 4.7 on September 21, 2026, calling it its most powerful model for coding and knowledge work. Within hours, the independent evaluator Artificial Analysis published its own results and delivered the more consequential headline: the score brings SpaceXAI into the top four AI labs for the first time.
The launch lands in a month that has already produced OpenAI's GPT-6 Astra, Anthropic's Claude Fable 5.1, Google's Gemini 3.8 Flash and Meta's Muse Spark 1.3. Grok 4.7 is xAI's answer, and the pitch is simple: xAI says the model is twice as fast at half the price of comparable frontier models.
What xAI Shipped in Grok 4.7
xAI says Grok 4.7 runs on a new, larger base model than Grok 4.6 and was trained with a longer reinforcement learning run on a harder mix of tasks, weighted toward problems that take many hours to complete. Elon Musk said on X that the model ships with 2.1 trillion parameters, roughly 40 percent more than Grok 4.6, although xAI's documentation does not confirm that figure.
Three engineering changes stand out:
- Better self-verification. xAI says the model checks its own work more carefully before answering, an improvement that shows up most in long document and presentation tasks.
- Long-context handling. The 500,000-token context window is unchanged from Grok 4.6, but xAI claims better management of long threads and multi-step reasoning.
- Native harness training. Grok 4.7 was trained to understand xAI's Grok Bot harness natively, making it more consistent in conversational and general knowledge work.
Reasoning effort is configurable from low to xhigh, and the model accepts text and image input with a knowledge cutoff of June 2026. Availability was broad from day one: Cursor, Grok Build, the xAI API, third-party coding harnesses, model routers and cloud platforms, plus a GitHub Copilot rollout for Pro, Pro+, Max, Business and Enterprise subscribers.
The Benchmarks: A Real Jump in Agentic Work
Artificial Analysis evaluated Grok 4.7 at xhigh reasoning effort and scored it 46 on the Intelligence Index, two points above Grok 4.6, enough to place SpaceXAI in the top four labs. The gain is concentrated almost entirely in agentic knowledge work, and there the movement is large:
- AA-Briefcase, a private benchmark of long-horizon professional work, rose 111 Elo to 1,657, putting Grok 4.7 just behind Claude Opus 5 and Claude Fable 5.1 at the frontier.
- GDPval-AA, which asks models to produce documents, spreadsheets and slides, rose 90 Elo to 1,695.
- The Coding Agent Index climbed nine points to 56 when paired with xAI's first-party Grok Build agent, overtaking GPT-5.6 Sol. DeepSWE v1.1 improved from 65 to 73 percent, Terminal-Bench 4.0 from 18 to 33 percent and SWE-Atlas-QnA from 58 to 63 percent. In native harnesses, only Claude Fable 5.1, GPT-6 Astra and Claude Opus 5 now rank ahead.
xAI's own table makes the price argument explicit. On the vendor's numbers, Grok 4.7 scores 46.3 percent on CursorBench 4.0 and 64 percent on EEBench, an electrical engineering test, at $2 input and $6 output per million tokens, against $10 and $50 for Fable 5.1 and $4 and $20 for GPT-5.6 Sol. On the same table, EEBench improves from 53 percent on Grok 4.6 to 64 percent, and the Harvey Legal Agent Benchmark from 15.8 to 19.6 percent.
The picture is not uniformly positive. Outside agentic knowledge work, Artificial Analysis found Grok 4.7 broadly matching Grok 4.6, with small regressions on the long-context reasoning test AA-LCR, down 3.7 points, and on AutomationBench-AA, down 1.1 points.
The Catch: It Thinks Longer, and You Pay for It
The gains come with a heavier token bill. Artificial Analysis measured roughly 81,000 output tokens per Intelligence Index task at xhigh effort, against 36,000 for Grok 4.6 at high effort and 27,000 for GPT-6 Astra at maximum effort. That is 125 percent and 196 percent more, respectively. Tasks averaged about 7.1 minutes each, with output speed near 188 tokens per second on long prompts.
List pricing is unchanged from Grok 4.6: $2 per million input tokens, $6 per million output tokens, and $0.50 per million cached input tokens. Two details matter for budgeting. Requests above 200,000 prompt tokens step up to higher rates, and the fast variant costs double for twice the output speed.
The practical consequence is that a cheap per-token price does not automatically mean a cheap task. A model that spends three times as many tokens to finish the same job can erase a large unit-price advantage. Cost per completed task, not headline price, is the number to watch.
Why the Harness Now Matters as Much as the Model
The security research firm XBOW, which had early access to Grok 4.7, published the most interesting independent finding of the launch. In its standard exploit-crafting harness, Grok 4.7 performed slightly worse than Grok 4.6. Inside xAI's Build orchestration, its performance improved substantially: Build-based systems produced 68 correct findings against 42 for Grok 4.6 in the same setup.
The reason appears to be behavioural. Grok 4.7 issues shorter, more atomic actions, preferring short shell commands over long Python scripts, which suits the fast feedback loop of an agentic coding environment. XBOW also found that xhigh reasoning offered no measurable advantage over high effort for offensive security work, and that a failure mode where Grok 4.6 would reason in circles without acting, roughly 0.85 percent of runs, disappeared entirely in 4.7.
The lesson for anyone building with agents is that prompts and models are no longer the whole story. The orchestration layer increasingly has to be matched to the model, not the other way around.
Safety and the Guardrail Race
xAI says Grok 4.7 ships with an entirely new safeguard stack and is the strongest model the company has tested on refusals and jailbreak resistance. The company reports 62.4 percent on LatchBio's biosafety benchmark and says the model blocked all but 3.3 percent of risky dual-use prompts on its internal HackerBench v0.3 while rarely refusing legitimate security work. Select cybersecurity partners now have invite-only access to the model's red-team capabilities.
Independent data supports the direction, if not the marketing. Artificial Analysis measured a hallucination rate of 29 percent, down from 34 percent on Grok 4.6, though factual accuracy stayed broadly flat at 47 percent.
The safety framing matters commercially as well. Earlier in September, OpenAI's GPT-6 Astra became the first model to trigger that lab's critical cybersecurity threshold, and Google shipped a defenders-only Gemini 3.8 Flash Cyber. Cyber capability is now a headline feature and a liability at the same time.
What It Means for Businesses and Builders
For teams building on AI, the Grok 4.7 launch reinforces three practical points:
- Run your own evaluation. Vendor and independent benchmarks disagreed in useful ways here. A model can lead on CursorBench and EEBench while merely matching its predecessor on general reasoning.
- Budget per task, not per token. The token-efficiency gap between Grok 4.7, Grok 4.6 and GPT-6 Astra changes the arithmetic of any high-volume agent workload.
- Match the harness to the model. XBOW's split results are the clearest evidence yet that orchestration is a first-class engineering decision, not a wrapper.
With Grok 4.7 holding frontier pricing at $2 and $6 per million tokens while comparable models charge up to $10 and $50, the wider story of September 2026 is competition compressing prices even as capability climbs. For Indian developers and small businesses weighing AI tools, that competition is the most valuable part of the news.