Invest1 publisher2 min readPublished
Musk grades his unreleased Grok 4.7 against a superseded Anthropic model
Musk said the delayed Grok 4.7 should be roughly on par with Opus 5.0, not 5.1. Parity with what Anthropic and OpenAI already sell he put two versions further up the ladder, at Grok 4.9.
The Investor · Invest desk
What happened
- Musk said on X that xAI's unreleased Grok 4.7 lines up more favorably with Anthropic's Opus 5.0 than with the newer Opus 5.1, Anthropic's Fable line or OpenAI's Astra.
- The model was originally scheduled for release on September 12, and Musk attributed the slip to reinforcement learning that may have penalized response length too heavily.
- Musk said the follow-up, Grok 4.8, is a 2.5-trillion-parameter model trained on xAI's new C++ stack that finishes training this week and then starts reinforcement learning.
- Third-party assessments put Anthropic's Claude Fable 5.1 at 52.6% on Terminal-Bench-Science, against published pricing of $10 per million input tokens and $50 per million output.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- decision A buyer comparing models this week has Anthropic's list price and nothing comparable from xAI. Grok 4.7 has no confirmed token pricing, and the pitch that it could be up to 10 times cheaper is unverified.
- constraint Musk puts the version that matches what Anthropic and OpenAI already sell at 4.9, two releases past the one that has not shipped. For the interim, xAI is competing on a schedule.
- exposure Reinforcement learning is the stage that delayed 4.7, and 4.8 still has it ahead. The model Musk calls a noticeable improvement has to clear the same step that broke the last one.
- precedent Cryptopolitan wrote that the admissions are unlike the hype Musk typically stokes pre-launch, which sets up the Grok 5 promise as the next claim buyers will discount.
Musk's release ladder doubles the size of Grok 4.6 across four models. It goes roughly 1.5 trillion parameters in 4.6, about 2.1 trillion in the delayed 4.7 supplemented with SpaceX engineering data [9], 2.5 trillion in the 4.8 he said finishes training this week [12], and a 3-trillion run with upgraded internal training software and cleaner data after that [14][1]. The step from 4.7 to 4.8 buys about 19 percent more parameters [2]. Musk described the current 2.5-trillion model as better than an earlier 2.1-trillion version trained with Jax that suffered mistakes "only corrected mid run" [18].
Ten times cheaper than Anthropic's published $10 per million input tokens and $50 per million output would be $1 and $5 [3]. Cryptopolitan reports that Grok 4.7 has no confirmed benchmark scores or token pricing, and that the claim it could be up to 10 times cheaper than a competing model is unverified [11].
Asked how the model stacks up against the competition, Musk wrote that it "should be roughly on par with Opus 5.0, not 5.1", and added that it was "better in some ways, worse in others" [2][3]. He said multimodal performance still needs work [5]. On Friday he said the model needed "a few more days to cook" [6].
By his own placement, the version that reaches "Astra/Fable class" is 4.9 [13]. Grok 5 is where he put the ceiling, saying it would be "maybe better than anything" and hedging with "we shall see", with no date attached [15]. Cryptopolitan did not report funding or compute-spend figures for xAI, Anthropic or OpenAI.
If 4.7 ships within days with published scores level with Opus 5.1 and a price near $1 and $5, the concession was expectation management ahead of a launch. If the 2.5-trillion run clears reinforcement learning without the length-penalty problem, 4.8 absorbs 4.7. The delay then costs xAI a few weeks of comparison instead of customers. For now the measurable baseline for xAI is Grok 4.6 [19].
What to watch
- Whether Grok 4.7 ships at all or gets folded into 4.8 once the 2.5-trillion-parameter run clears reinforcement learning.
- Any published Grok token price sheet, and whether it lands near the $1 and $5 that a 10x discount to Anthropic implies.
- Whether Anthropic or OpenAI change pricing on Opus 5.1 or Astra before xAI quotes anything.