SpaceXAI Sells Grok 4.7 on Price–Performance — Its Own Numbers Show Second Place Behind Fable 5.1

SpaceXAI Sells Grok 4.7 on Price–Performance — Its Own Numbers Show Second Place Behind Fable 5.1

SpaceXAI's new model for coding and knowledge work arrived on September 21, after at least five scrapped launch dates. The company sells it on price–performance, with an output-token price roughly one-fifth of Fable 5.1's, while the company's own benchmarks place the model second on key metrics.

SpaceXAI, the company behind Grok, launched the model Grok 4.7 on September 21, 2026, aimed at coding and knowledge work. Pricing is set at $2 per million input tokens and $6 per million output tokens (Unite.AI). The model is available with no waitlist from day one, in Cursor, Grok Build, third-party model routers, the xAI API, the Grok app, X, and via the voice assistants in Tesla's dashboards (Android Headlines via MSN).

The Price That Carries the Story

The company's own pricing table compares Grok 4.7 with its competitors: GPT-5.6 Sol costs $4/$20, and Claude Fable 5.1 costs $10/$50 per million input/output tokens. That means Grok 4.7's output-token price of $6 is roughly one-fifth of Fable 5.1's $50, and less than a third of GPT-5.6 Sol's $20. The comparison is thus against GPT-5.6 Sol — not GPT-6, which is not priced in the table (Unite.AI).

For latency-sensitive workloads, SpaceXAI offers a variant that processes queries twice as fast, at twice the price (SiliconANGLE). Grok 4.6 remains at the same price, $2/$6, according to the same table.

What the Numbers Show — and Who Computed Them

All benchmark figures are the company's own, self-reported, and not independently verified.

On CursorBench 4.0, which SpaceXAI says stresses longer coding tasks, Grok 4.7 is reported at 46.3 percent — up from 40.4 for Grok 4.6, and ahead of GPT-5.6 Sol at 41.7 — but clearly behind Fable 5.1 at 51.8 (Unite.AI).

On agent benchmarks, Grok 4.7 lands in second place: 1695 on GDPval versus Fable 5.1's 1735, and 1657 on AA-Briefcase versus 1678 (Decrypt). SiliconANGLE reports that the model beats Fable 5.1 on the Harvey Legal Agent Benchmark and EEBench, but trails OpenAI Group PBC's newest model, GPT-6 Astra, on EEBench (SiliconANGLE).

The comparisons are not fully like-for-like. Unite.AI notes that Grok 4.7 was run at the "xhigh" setting, while the competitors were run at "max" — a detail that complicates the interpretation of the differences.

Safety — the Company's Own Claims

SpaceXAI says Grok 4.7 was built with an entirely new safety stack, and that it is the strongest model the company has tested on refusals and jailbreak resistance. The company reports 62.4 percent on the LatchBios biosafety benchmark and 3.3 percent pass-through on its own HackerBench v0.3. These figures, too, are self-reported (Unite.AI).

Delayed, and Talked Down by the Founder Himself

The launch followed at least five occasions on which Elon Musk pushed back the timeline since late July, according to Decrypt, with ever-later time estimates — including "ten days" as of September 1 and "a few more days" on September 11 (Decrypt). Musk also lowered expectations himself: according to Decrypt, he wrote a few days before launch that Grok 4.7 would land roughly at the level of Anthropic's Claude Opus 5.0 — not the newer Opus 5.1 — and that multimodal performance still needed work (Decrypt).

Decrypt also reports that the model has 2.1 trillion parameters, a 40 percent increase from Grok 4.6's 1.5 trillion, and that it was trained with supplementary engineering data from SpaceX, including Starlink telemetry. These are company figures relayed via Decrypt and Android Headlines, with no primary documentation confirming them.

What Remains

The open tension in the story is that SpaceXAI markets Grok 4.7 as leading on price–performance, while the model's own numbers — and Musk's own caveats — place it second behind Fable 5.1 on key benchmarks. Both can be true at once: you get near-best performance at a fraction of the price. How large the gap is in practical use on real workloads, and what it would take to close it, remains unknown — and no independent evaluation of the self-reported figures exists yet.