xAI holds Grok's $2 token price for a model 40 Elo points behind Fable 5.1
Grok 4.7 arrived on Monday after five walked-back timelines, with 40% more parameters and Grok 4.6's list price intact. Cursor's own cost chart still puts its price per task above GPT-6 Astra and Claude Sonnet 5.
The Investor · Invest desk
What happened
- xAI shipped Grok 4.7 on Monday after Musk walked the timeline back at least five times since late July, from "four weeks out" to "needs a few more days to cook" on September 11.
- The model runs on 2.1 trillion parameters, up 40% from Grok 4.6's 1.5 trillion, with supplemental training on Starlink telemetry, SpaceX manufacturing records and engineering failure logs.
- xAI held the starting price at $2 per million input tokens and $6 per million output, and sells a variant with twice the output speed at twice the price.
- On the benchmarks xAI published, Grok 4.7 came second to Claude Fable 5.1 on GDPval and AA-Briefcase, and second to GPT-6 Astra on the EEBench electrical-engineering test.
- Against its predecessor the model gained ground, scoring 46.3% on CursorBench 4.0 versus 40.4%, and 38% on Terminal Bench 4.0 versus 20.3%.
Why it matters
- Who has to decide decision Teams comparing list prices per token and teams comparing the cost of a finished task will pick different vendors from the same published numbers, so the choice of accounting unit now decides the contract.
- What it costs cost By holding price while the base model grew by 600 billion parameters, xAI pays for the extra serving capacity itself instead of passing it to customers.
- What it limits constraint With no dates attached to Grok 4.8, 4.9 or 5, anyone standardising on xAI has to justify the decision on the model that shipped Monday rather than the roadmap behind it.
Next story · Build · Sep 21
Cursor's task-cost table shrinks Grok 4.7's 7.5x list-price gap to about 1.9xConfirmed8 publishers
Forty Elo points is the whole distance between Grok 4.7's 1695 on GDPval and Claude Fable 5.1's 1735 [4][22]. GDPval scores head-to-head on the system chess uses [13], so that gap converts to about 56 wins in 100 for Fable [23]. On AA-Briefcase the margin is 21 points, or roughly 53 in 100 [5][24].
The base model grew by 600 billion parameters. The list price did not move [7][29]. Musk called Grok 4.7 "a strong combination of intelligence, speed & low cost" on X [20], and xAI's own release note described it as "a notable improvement over Grok 4.6 at the same price and speed" [19].
Price is charged per token, and tokens are what the new model consumes more of. xAI says Grok 4.7 spends longer working through hard problems and double-checks its own answers more often than Grok 4.6 did [8]. On CursorBench 4.0, which plots accuracy against the cost and token count each task burns, Decrypt reports Grok 4.7 landing pricier per task than GPT-6 Astra and Claude Sonnet 5, with Fable 5.1 winning at every price point on the chart [32].
A team billed by consumption pays for the extra verification steps. The per-token discount can still leave the customer paying more per task.
The two published accounts of the same benchmark set do not agree. Crypto Briefing reports that xAI's results put Grok 4.7 ahead of GPT 5.6 Sol on several evaluations, with Sol and Fable 5.1 ahead on others [18]. Against Grok 4.6, the Terminal Bench 4.0 gain works out at 87% in relative terms, while CursorBench 4.0 moved 5.9 points [30][31]. Grok 4.7 also posted 71% on DeepSWE 1.1 at high reasoning effort [17], and, on the safety side, allowed 3.3% of risky dual-use prompts through on HackerBench 0.3 and scored 62.4% on a LatchBio biosafety benchmark [26].
The months of delay went into a larger base model and a longer reinforcement learning run on tasks that take hours to finish [11]. They did not go into multimodal, which Musk said still needs work when he set expectations at "roughly on par with" Anthropic's Claude Opus 5.0, not the newer Opus 5.1 [25]. He then sketched Grok 4.8 as a meaningful step up, Grok 4.9 in "Astra/Fable class" and Grok 5 as a possible frontier leader, with no release dates. "We shall see," he wrote [33].
I'd expect distribution to decide more of this than either leaderboard. Grok 4.7 shipped with no waitlist into the Grok app, Cursor, Grok Build and the xAI API [9], and reaches developers through third-party coding tools, model routers and cloud platforms [10]. The counter-case is the one Cursor's chart makes: agent work bills by consumed tokens, and a model that reasons longer gives back its token-price advantage exactly where the spend concentrates. An independent cost-to-completion measurement at matched accuracy would settle it. If Grok 4.7 finishes the same task for less money, Decrypt's claim that xAI has consistently undercut Anthropic and OpenAI on cost per token holds [21]; if it does not, xAI is selling second place at a premium per task.
What to watch
- An independent cost-per-completed-task measurement at matched accuracy, testing whether the $2 per million token price survives task-level accounting.
- A dated release for Grok 4.8, or for the Grok 4.9 that Musk placed in "Astra/Fable class."
- What the selected cybersecurity partners given red-team access report against the 3.3% HackerBench 0.3 pass-through figure.
Information, not investment advice. This story is general information about companies and markets, drawn from cited reporting. It is not a recommendation to buy, sell or hold any security; check the cited sources before acting.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence45
- Adoption35
- Hype gap+15
- Incentives70
- Confidence55
Perspective Coverage
3 publishers- Builder
- Builder 42%
- Operator
- Operator 27%
- Investor
- Investor 31%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
xAI released Grok 4.7 on Monday, after at least five delays since late July.
- [2]
Musk walked the Grok 4.7 timeline back at least five times since late July: "four weeks out," then "a few weeks," then "3 to 4 weeks," then "10 days" on September 1, then "needs a few more days to cook" on September 11.
- [3]
Grok 4.7 scored second to Claude Fable 5.1 on GDPval and AA-Briefcase, and second to GPT-6 Astra on EEBench, an electrical-engineering benchmark.
- [4]
Grok 4.7 hit 1695 on GDPval; Claude Fable 5.1 topped the chart at 1735.
- [5]
On AA-Briefcase, built by Artificial Analysis, Grok 4.7 posted 1657 against Fable 5.1's 1678.
- [6]
Grok 4.7 costs $2 per million input tokens and $6 per million output tokens.
- [7]
xAI is keeping the same starting price for Grok 4.7 as Grok 4.6, at $2 per million input tokens and $6 per million output tokens; a faster version offering twice the output speed is available at twice the price.
- [8]
xAI says Grok 4.7 spends longer working through hard problems and double-checks its own answers more often than Grok 4.6 did.
- [9]
There is no waitlist for Grok 4.7; it is live in the Grok app, Cursor, Grok Build, and the xAI API.
- [10]
Grok 4.7 is available through Cursor, Grok Build and the Grok API, as well as third party coding tools, model routers and cloud platforms.
- [11]
Grok 4.7 uses a larger base model than Grok 4.6 and underwent a longer reinforcement learning run focused on more difficult tasks that can take hours to complete; xAI said the training improved its ability to verify its own work, manage longer context and operate within the Grok Bot environment.
- [12]
Grok 4.7 runs on 2.1 trillion parameters, up 40% from the 1.5 trillion in Grok 4.6.
- [13]
GDPval measures performance on real, economically valuable knowledge work using tasks vetted by working professionals, and scores it as an Elo rating, the same head-to-head ranking system chess uses.
- [14]
xAI folded supplemental training data from SpaceX into Grok 4.7: Starlink satellite telemetry, manufacturing records, and engineering failure logs.
- [15]
Grok 4.7 scored 46.3% on CursorBench 4.0 compared with 40.4% for Grok 4.6.
- [16]
Grok 4.7's Terminal Bench 4.0 score rose to 38% from Grok 4.6's 20.3%.
- [17]
Grok 4.7 scored 71% on DeepSWE 1.1 at high reasoning effort, 1,657 on AA Briefcase and 64% on EEBench.
- [18]
xAI's results put Grok 4.7 ahead of GPT 5.6 Sol on several of those evaluations, though GPT 5.6 Sol and Fable 5.1 remained ahead on others.
- [19]
xAI called Grok 4.7 "a notable improvement over Grok 4.6 at the same price and speed."
- [20]
Musk followed up on X, calling Grok 4.7 "a strong combination of intelligence, speed & low cost."
- [21]
According to Decrypt, xAI has consistently undercut Anthropic and OpenAI on cost per token even as it trails them on raw capability, betting that "good enough, cheap, and everywhere" beats "best, but pricier" for the bulk of everyday use.
- [22]
Fable 5.1 leads Grok 4.7 on GDPval by 40 Elo points.
- [23]
A 40-point Elo gap implies an expected head-to-head win rate of about 56% for Fable 5.1 against Grok 4.7.
- [24]
The 21-point AA-Briefcase gap (1678 against 1657) implies an expected head-to-head win rate of about 53% for Fable 5.1.
- [25]
Days before launch Musk wrote that Grok 4.7 should land "roughly on par with" Anthropic's Claude Opus 5.0, not the newer Opus 5.1, with multimodal performance still needing work.
- [26]
xAI said Grok 4.7 allowed 3.3% of risky dual use prompts through on HackerBench 0.3 and scored 62.4% on a biosafety benchmark from LatchBio.
- [27]
xAI is providing selected cybersecurity partners with access to the model's red team capabilities for defensive research.
- [28]
Grok 4.5 launched in July with the biggest training cluster in the industry and third-place scores behind Claude and OpenAI's models.
- [29]
Grok 4.7 adds 600 billion parameters over Grok 4.6.
- [30]
The Terminal Bench 4.0 move from 20.3% to 38% is a gain of 17.7 points, or 87% in relative terms.
- [31]
The CursorBench 4.0 move from 40.4% to 46.3% is a gain of 5.9 points, or about 15% in relative terms.
- [32]
CursorBench 4.0 plots accuracy against the cost and token count each task burns through; Grok 4.7 lands in the middle, pricier per task than GPT-6 Astra and Claude Sonnet 5, and short of Fable 5.1, which wins at every price point on the chart.
- [33]
In the same post Musk sketched Grok 4.8 as a meaningful step up, Grok 4.9 in "Astra/Fable class," and Grok 5 as a possible frontier leader; those upgrades have no release date, and he wrote "We shall see."
Sources
3 independent publishers whose own reporting we read for this story.
- cryptobriefing.comxAI launches Grok 4.7 with stronger coding and knowledge work performance
1 article · September 21, 2026
- cryptopolitan.comxAI's Grok 4.7 outscores Anthropic and OpenAI models on legal agent work for less
1 article · September 22, 2026
- decrypt.coxAI Launches Grok 4.7. It's Bigger, But Late to the AI Frontier Party
1 article · September 21, 2026