Grok 4.7 Shows Why Cheap AI Tokens Can Cost More

xAI launched Grok 4.7 on September 21 with 2.1 trillion parameters and an API price of $2 per million input tokens and $6 per million output tokens. That sticker price is the hook. Agent loops decide how deep it goes.

The model arrived after repeated shifts in Elon Musk’s projected release window, as documented in Decrypt’s account of the launch. xAI called Grok 4.7 a “notable improvement” over Grok 4.6 at the same price and speed. It is available through the Grok app, Cursor, Grok Build and xAI’s API.

Grok 4.7 also carries an unusual training pitch. xAI supplemented its training with SpaceX material including Starlink telemetry, manufacturing records and engineering failure logs. The proposed advantage is better reasoning about hardware and physical systems than a model trained mostly on internet text. The available benchmark results do not establish that broad claim, although the model performs competitively in several knowledge-work and engineering tests.

Launch-week winner labels offer little help to anyone approving an API budget. Leaderboards turn model releases into fantasy football for people with access tokens. What buyers need is the cost of an acceptable completed task, especially when an AI agent can make repeated tool calls while working through one prompt.

Cheap Tokens Can Still Produce an Expensive Task

A token is a small unit of text processed or generated by a model. API providers usually charge for each million tokens, like a utility charging by the kilowatt-hour. An agentic task may involve many model requests: inspect a file, call a tool, read the response, revise the plan, call another tool, and continue until the job is done or the robot equivalent of “works on my machine” appears.

That creates three separate units buyers should track:

1. Price per token, the rate printed on the menu.
2. Tokens per request, how much the model consumes during one API call.
3. Tokens and requests per completed task, the amount that reaches the bill.

Theo Browne highlighted the collision between those units in his analysis of Grok 4.7. Citing Artificial Analysis, he said Grok 4.6 at its extra-high reasoning setting generated about 38,000 output tokens per task, while Grok 4.7 generated about 81,000. Dividing 81,000 by 38,000 gives roughly 2.1 times as many output tokens per task in that benchmark.

At an unchanged output-token rate, 2.1 times the tokens would produce roughly 2.1 times the output-token charge. That inference assumes comparable tasks, settings and billing treatment. Input tokens, caching, failed attempts and other system costs can change the final total, so the ratio should not be treated as a universal Grok bill multiplier.

Michael Truell of Cursor offered a narrower production measurement in the public exchange recounted by Browne: Grok 4.7 reportedly used 5% more tokens for median requests and 20% to 30% more for P99 requests, meaning requests near the heavy end of the distribution.

Those figures can coexist with the benchmark result. A prompt that triggers more requests can consume far more tokens across the whole job even if each request grows by only 5% or 30%. Think of a delivery service that raises the cost of each trip slightly while sending twice as many vans. The per-trip increase looks modest; accounting still starts sweating.

We do not have independent access to xAI’s or Cursor’s production telemetry, and the Artificial Analysis result represents a benchmark workload rather than every deployment. Buyers can resolve the ambiguity for their own use by logging total tokens, request count, completion rate and wall-clock time per task. Comparing API menu prices alone leaves out the variable the agent controls: how long it keeps working.

Grok 4.7 Has a Jagged Capability Profile

The benchmark picture resists a single rank. On GDPval, which evaluates economically valuable work such as legal memos, spreadsheets and presentations, Grok 4.7 scored 1,695 Elo against Claude Fable 5.1’s 1,735. On AA-Briefcase, a multi-hour office-work test, Grok recorded 1,657 against Fable’s 1,678.

Coding results spread out further. In Matthew Berman’s benchmark breakdown, Grok 4.7 scored 38% on Terminal Bench 4.0, compared with 57.9% for Fable 5.1 and 58.2% for GPT-6 Astra. On DeepSWE, it posted 71%, close to GPT-5.6 Sol at 72.7% and Fable at 70%. Berman observed that xAI’s displayed DeepSWE table omitted Astra, although Astra appeared elsewhere in the release material. The 74.1% Astra result he added would lead that table.

Legal work produced the funkiest plot twist. Grok scored 19.6%, ahead of the named frontier rivals in Berman’s comparison: Sol at 2.5%, Fable at 6.7% and Astra at 5.4%. Muse Spark 1.2 scored 42%, however, so Grok did not lead the full field shown.

The comparison supports a narrower description than “best” or “second best.” Grok 4.7 appears competitive on broad knowledge work and unusually strong against several frontier rivals on that legal test, while its Terminal Bench result shows a large coding-agent gap. Each benchmark uses its own harness, prompts and scoring method, so those rankings may change inside a company’s codebase or document workflow.

A separate review from TheAIGRID reached a similarly mixed judgment. The reviewer confirmed the $2 and $6 API rates and described the model as pleasant on some engineering tasks, while criticizing its front-end and 3D performance and reporting loops in use. One reviewer’s front-end and 3D workload can expose failure modes, but it cannot estimate how often every developer will hit them.

The Delay History Explains the Temperature

Grok 4.7 inherited expectations from more than its version number. Decrypt’s reconstruction of Musk’s posts records at least five shifts in the projected release window, progressing from “four weeks out” to “a few weeks,” “3 to 4 weeks,” “10 days,” and then “needs a few more days to cook.” Decrypt also reported that, days before release, Musk wrote that Grok 4.7 should land “roughly on par with” Claude Opus 5.0, rather than the newer Opus 5.1, with multimodal performance still needing work.

The recent release history makes another near-frontier result easier to interpret. Grok 4.5 arrived in July with what Decrypt described as the industry’s largest training cluster, yet placed third behind models from Anthropic and OpenAI. Grok 4.6 subsequently trailed the frontier group on coding autonomy. Across those releases, larger models and substantial compute have coexisted with results below first place.

That history suggests a plausible product strategy. A model can succeed without topping every benchmark if it is cheap enough, widely distributed and strong on the workloads its customers run. Grok already has distribution through X, its standalone app, Cursor and other xAI surfaces. For a high-volume workflow that tolerates retries or accepts a lower success rate, the low token price may beat a pricier frontier model.

For tasks where one wrong answer creates expensive rework, a higher-priced model could cost less overall by succeeding sooner. The useful comparison is success-adjusted cost: total spend divided by completed, acceptable tasks. It also needs latency and human-review time. Otherwise, “cheap” describes the ingredient while ignoring how much the recipe uses.

Grok 4.7’s launch leaves buyers with a better shopping rule than any leaderboard crown: meter the whole task, because the cheapest token can become expensive when an agent keeps ordering seconds.