SpaceXAI released Grok 4.7 on 21 September, a model for programming, agent tasks and knowledge work built on a larger base than Grok 4.6 and trained with longer reinforcement learning on harder, multi hour tasks.1 ForkLog 2026-09-22 Unveiled 21 Sep; CursorBench 4.0 46.3 vs 40.4 (Grok 4.6), 41.7 (GPT-5.6 Sol), 51.8 (Fable 5.1); AA Intelligence Index 46, up two; 81,000 vs 36,000 output tokens; 2 and 6 dollars per million; 500,000 context; Fast variant in Cursor and Grok Build only. Open source The company reports 46.3 percent on CursorBench 4.0, up from 40.4 for Grok 4.6, ahead of GPT-5.6 Sol at 41.7 and behind Claude Fable 5.1 at 51.8.1 ForkLog 2026-09-22 Unveiled 21 Sep; CursorBench 4.0 46.3 vs 40.4 (Grok 4.6), 41.7 (GPT-5.6 Sol), 51.8 (Fable 5.1); AA Intelligence Index 46, up two; 81,000 vs 36,000 output tokens; 2 and 6 dollars per million; 500,000 context; Fast variant in Cursor and Grok Build only. Open source Pricing is unchanged at 2 dollars per million input tokens and 6 dollars per million output tokens.2 ITdaily 2026-09-22 Grok 4.5 in July, 4.6 in August, 4.7 in September; DeepSWE 71, EEBench 64, LatchBio 62.4; same 2 and 6 dollar pricing. Open source On the independent Artificial Analysis Intelligence Index the model scores 46, two points above its predecessor.1 ForkLog 2026-09-22 Unveiled 21 Sep; CursorBench 4.0 46.3 vs 40.4 (Grok 4.6), 41.7 (GPT-5.6 Sol), 51.8 (Fable 5.1); AA Intelligence Index 46, up two; 81,000 vs 36,000 output tokens; 2 and 6 dollars per million; 500,000 context; Fast variant in Cursor and Grok Build only. Open source We assess with high confidence that Grok 4.7 is a real but incremental improvement that keeps SpaceXAI in the second tier, and with moderate confidence that its unchanged sticker price understates its cost per task.
What changed, and what the tables hide
The vendor numbers show gains across the board: 71 percent on DeepSWE v1.1, 64 percent on EEBench and 62.4 percent on the LatchBio biosafety benchmark.2 ITdaily 2026-09-22 Grok 4.5 in July, 4.6 in August, 4.7 in September; DeepSWE 71, EEBench 64, LatchBio 62.4; same 2 and 6 dollar pricing. Open source Artificial Analysis’s coding agent index, run with SpaceXAI’s Grok Build harness, rose from 47 to 56.1 ForkLog 2026-09-22 Unveiled 21 Sep; CursorBench 4.0 46.3 vs 40.4 (Grok 4.6), 41.7 (GPT-5.6 Sol), 51.8 (Fable 5.1); AA Intelligence Index 46, up two; 81,000 vs 36,000 output tokens; 2 and 6 dollars per million; 500,000 context; Fast variant in Cursor and Grok Build only. Open source The model offers a 500,000 token context window and four reasoning levels, and a Fast variant at double speed and double price available only in Cursor and Grok Build.1 ForkLog 2026-09-22 Unveiled 21 Sep; CursorBench 4.0 46.3 vs 40.4 (Grok 4.6), 41.7 (GPT-5.6 Sol), 51.8 (Fable 5.1); AA Intelligence Index 46, up two; 81,000 vs 36,000 output tokens; 2 and 6 dollars per million; 500,000 context; Fast variant in Cursor and Grok Build only. Open source
Two caveats matter. First, SpaceXAI’s comparison tables tested Grok 4.7 at its highest reasoning setting against Grok 4.6 at a lower one, so the headline deltas mix model improvement with extra compute.3 Next 2026-09-22 Tables compared xhigh to high; rival prices cited; 23 Sep update: independent Terminal-Bench 4.0 at 26 vs 38 claimed; Astra 60, Fable 5.1 55; 240M tokens in testing vs 88M median. Open source Second, the extra compute shows up in the output. Artificial Analysis found Grok 4.7 at its highest setting used about 81,000 output tokens per complex task, against 36,000 for Grok 4.6.1 ForkLog 2026-09-22 Unveiled 21 Sep; CursorBench 4.0 46.3 vs 40.4 (Grok 4.6), 41.7 (GPT-5.6 Sol), 51.8 (Fable 5.1); AA Intelligence Index 46, up two; 81,000 vs 36,000 output tokens; 2 and 6 dollars per million; 500,000 context; Fast variant in Cursor and Grok Build only. Open source At an unchanged per token price, more than doubling the tokens per task roughly doubles the bill for the same job.
The cadence is the strategy
Grok 4.5 shipped in July, Grok 4.6 in August and Grok 4.7 in September, roughly one model a month.2 ITdaily 2026-09-22 Grok 4.5 in July, 4.6 in August, 4.7 in September; DeepSWE 71, EEBench 64, LatchBio 62.4; same 2 and 6 dollar pricing. Open source The price has held at 2 and 6 dollars throughout, against the 4 and 20 dollars SpaceXAI cites for GPT-5.6 Sol and 10 and 50 dollars for Claude Fable 5.1.3 Next 2026-09-22 Tables compared xhigh to high; rival prices cited; 23 Sep update: independent Terminal-Bench 4.0 at 26 vs 38 claimed; Astra 60, Fable 5.1 55; 240M tokens in testing vs 88M median. Open source The pattern is a challenger’s: ship often, stay close enough on capability to be on the shortlist, and compete on price for buyers who do not need the top model. Distribution through Cursor matters to that plan, because it puts Grok in front of developers who choose models per task rather than per vendor.1 ForkLog 2026-09-22 Unveiled 21 Sep; CursorBench 4.0 46.3 vs 40.4 (Grok 4.6), 41.7 (GPT-5.6 Sol), 51.8 (Fable 5.1); AA Intelligence Index 46, up two; 81,000 vs 36,000 output tokens; 2 and 6 dollars per million; 500,000 context; Fast variant in Cursor and Grok Build only. Open source
The monthly cadence also explains the token inflation. A lab shipping every four weeks has limited time to improve a model’s efficiency between releases, and the cheapest way to lift benchmark scores within a release cycle is to let the model think longer. That is consistent with what Artificial Analysis measured: a two point move on its index alongside more than double the output tokens per task.1 ForkLog 2026-09-22 Unveiled 21 Sep; CursorBench 4.0 46.3 vs 40.4 (Grok 4.6), 41.7 (GPT-5.6 Sol), 51.8 (Fable 5.1); AA Intelligence Index 46, up two; 81,000 vs 36,000 output tokens; 2 and 6 dollars per million; 500,000 context; Fast variant in Cursor and Grok Build only. Open source It is also consistent with the comparison tables pairing the new model’s highest effort setting against the old model’s lower one.3 Next 2026-09-22 Tables compared xhigh to high; rival prices cited; 23 Sep update: independent Terminal-Bench 4.0 at 26 vs 38 claimed; Astra 60, Fable 5.1 55; 240M tokens in testing vs 88M median. Open source None of this makes the gains fake. It means the improvement is partly purchased at inference time, and the customer pays for that purchase per token.
The safety claims
SpaceXAI also says it redesigned the model’s safety system, reporting that Grok 4.7 allowed 3.3 percent of dangerous queries on HackerBench v0.3.1 ForkLog 2026-09-22 Unveiled 21 Sep; CursorBench 4.0 46.3 vs 40.4 (Grok 4.6), 41.7 (GPT-5.6 Sol), 51.8 (Fable 5.1); AA Intelligence Index 46, up two; 81,000 vs 36,000 output tokens; 2 and 6 dollars per million; 500,000 context; Fast variant in Cursor and Grok Build only. Open source ITdaily reports the company’s claim of the highest resistance to unwanted prompts and jailbreaks in dual use domains such as cybersecurity and biotechnology.2 ITdaily 2026-09-22 Grok 4.5 in July, 4.6 in August, 4.7 in September; DeepSWE 71, EEBench 64, LatchBio 62.4; same 2 and 6 dollar pricing. Open source These are vendor reported figures on benchmarks with limited outside replication, and they should be weighed as claims until independent red teams publish comparable results. They do signal that SpaceXAI is now competing on safety posture in public, a change from earlier releases framed almost entirely around capability.
Who gains and who loses
Cursor and other multi model tools gain a cheaper option for routine coding work and leverage over the leading labs’ pricing. Cost sensitive developers gain if their tasks run at lower effort settings, where the token inflation matters less. Anthropic and OpenAI face mild pressure at the bottom of their price ladders rather than at the top, since Grok 4.7 still trails their leading models on the vendor’s own table.1 ForkLog 2026-09-22 Unveiled 21 Sep; CursorBench 4.0 46.3 vs 40.4 (Grok 4.6), 41.7 (GPT-5.6 Sol), 51.8 (Fable 5.1); AA Intelligence Index 46, up two; 81,000 vs 36,000 output tokens; 2 and 6 dollars per million; 500,000 context; Fast variant in Cursor and Grok Build only. Open source Buyers who budget on per token price rather than per task cost are the ones most exposed to a surprise.
The counter case
The strongest case against our reading is that a monthly cadence compounds. If each release adds five or six points on agentic coding benchmarks at constant price, SpaceXAI closes the gap with the leaders within two or three releases, and the second tier label expires quickly. The independent index moving only two points argues against that pace, but one release is a thin sample.1 ForkLog 2026-09-22 Unveiled 21 Sep; CursorBench 4.0 46.3 vs 40.4 (Grok 4.6), 41.7 (GPT-5.6 Sol), 51.8 (Fable 5.1); AA Intelligence Index 46, up two; 81,000 vs 36,000 output tokens; 2 and 6 dollars per million; 500,000 context; Fast variant in Cursor and Grok Build only. Open source
What to watch
- The Terminal-Bench claim will not survive independent testing. We expect third party reruns to come in well below the 38 percent claimed, closer to the mid 20s. That would put Grok 4.7 far behind GPT-6 Astra and Claude Fable 5.1. Watch token use too: a model this verbose on agent tasks will look expensive per solved problem.
What happened next (updated Oct 13, 2026)
This section was added after publication. In an update on 23 September, Next reported that Artificial Analysis measured Grok 4.7 at 26 percent on Terminal-Bench 4.0, against the 38 percent SpaceXAI had claimed, with GPT-6 Astra at 60 and Claude Fable 5.1 at 55, and that the model generated about 240 million tokens during testing against a median of 88 million.3 Next 2026-09-22 Tables compared xhigh to high; rival prices cited; 23 Sep update: independent Terminal-Bench 4.0 at 26 vs 38 claimed; Astra 60, Fable 5.1 55; 240M tokens in testing vs 88M median. Open source We found no reporting of a further Grok release through 13 October.