Grok 4.7 vs Grok 4.5: Faster at the Same Price, and 748 Prompt Tokens You Never Sent

Grok 4.7 and Grok 4.5 have the same list price: $2 input / $6 output per million tokens. On our three graded tasks, both passed all 15 calls. Grok 4.7 was faster on every one of them.

It also cost 1.4× more to run the same battery, and the price list cannot tell you why. Grok 4.7 bills 748 more prompt tokens than Grok 4.5 on every call, for text we never sent. And on a ten-word "reply with only OK" prompt, it refused in 2 of 3 runs.

How we tested

Three script-graded tasks: strict JSON extraction from a messy support ticket, a four-step arithmetic word problem (answer: 580), and an instruction-adherence task (exactly three bullets, each under 12 words, no preamble). Five runs per task per model, temperature 0.7, max_tokens 1500, run on 2026-10-19 through our own provider handlers. Costs are xAI's own billed figure (cost_in_usd_ticks), which equals list rates from shared/pricing.ts with reasoning priced as output and cached input at the cached rate. Separately, we sent a ten-word probe prompt three times to each model to measure token overhead.

Both models passed 15 of 15. On these tasks there was no accuracy difference to report. Everything below is about speed, tokens and cost.

Speed and cost per task

Median of five runs. "Before" is the cost Promptster displayed when we ran this, which left out the reasoning tokens xAI reports outside completion_tokens and priced cached input at the full rate. "Now" is xAI's billed cost, which Promptster displays since 2026-10-22 (more on that below).

Grok 4.7 was 1.9× faster on extraction and 20–30% faster on the other two. It also cost more on every task: 1.38× across all 15 calls on xAI's billed cost. The old display put the gap at 1.93×.

The answers themselves were near-identical. Both returned the same four JSON fields on every run, both got 580 every time (both like LaTeX: \(48 \times 1.25 = 60\)), and both produced three short bullets with no preamble.

The 748 tokens

Here are the prompt tokens each model reported for the same input:

The difference is exactly 748 tokens on every prompt. Our text is counted the same way on both models, and Grok 4.7 adds a fixed block on top. For comparison, OpenAI's GPT-6 models count that ten-word probe as 18 prompt tokens. Claude 5.5 counts it as 25.

Two details in the raw usage object point to text added on xAI's side rather than anything in our request:

• prompt_tokens_details.cached_tokens was 1,152 on the very first Grok 4.7 probe call (384 on Grok 4.5). We had never sent that prompt before, so the cache hit has to be a prefix shared across requests.
• The overhead does not depend on what we send. A fixed 748-token difference across four unrelated prompts is what you would see from a longer system prompt on the provider's side.

At the full input rate, 748 tokens would be $0.0015 per call. xAI serves them from its cache, though, at $0.50 per million for Grok 4.7, so they cost about $0.0004 per call. Across our 15 graded calls Grok 4.7's input cost $0.0081 more than Grok 4.5's, which is 69% of the $0.0117 gap. The other 31% is output and reasoning. Grok 4.7 cost more mainly because of what it was sent before our prompt, not because of what it wrote.

That cache discount is in xAI's favour, and it makes short prompts cheaper than the $2 headline rate suggests: on the JSON task a median of 1,280 of Grok 4.7's 1,382 prompt tokens were cached. Our first figures priced every prompt token at list rate. xAI reports the exact bill in cost_in_usd_ticks, and it matches list rates with cached input at $0.50 per million, so the costs above use it.

Reasoning tokens, and a gap in our own numbers

Both models reason before answering, and xAI reports that work in completion_tokens_details.reasoning_tokens, outside completion_tokens. On the first Grok 4.7 probe call: 1,254 prompt + 18 completion + 329 reasoning = 1,601 total_tokens.

Median reasoning tokens per task:

Grok 4.7 reasons less than 4.5 on two tasks and about 55% more on the arithmetic problem.

To be clear about our own product: until 2026-10-22, Promptster's displayed cost counted completion_tokens only, so it left xAI's reasoning tokens out, and it priced the cached prefix at the full input rate. Promptster now counts reasoning tokens, cached input and per-request fees, and for xAI it shows the cost xAI reports. The two errors pulled in opposite directions, which is why Grok 4.5's cost went up between the columns above and Grok 4.7's went down.

The refusal

The probe prompt was ten words: "Reply with only the word OK and nothing else, please." Grok 4.5 answered OK three times out of three. Grok 4.7 did this:

Two refusals in three tries, on a prompt with nothing in it to refuse. Run 2 spent 991 reasoning tokens and ten seconds deciding not to say "OK".

Some context. This is n=3 on a probe, not a graded task, and Grok 4.7 passed all 15 real calls, including the three-bullet task, which is just as strict about format. We are not claiming Grok 4.7 is unreliable at following instructions. We are claiming that something in the text added on the provider's side makes it suspicious of "reply with only X" phrasing. That is a common pattern in production: classifiers, health checks, yes/no gates and smoke tests all ask for exactly one word.

It also reverses the latency result. On the probe, Grok 4.5's median was 0.7 s and Grok 4.7's was 4.3 s. Its speed advantage on real tasks disappears when it stops to argue with the instruction.

What we'd do

For JSON extraction and short structured answers, Grok 4.7 is the faster choice at a modest, predictable premium. On extraction, 1.6 s against 3.0 s is a real difference on an interactive path.

Budget Grok 4.7 per call, not per token. A fixed 748-token overhead hurts short prompts most. On our ten-word probe it made up 60% of the prompt-token count. On a 10,000-token document it would barely register.

If you depend on one-word replies, test Grok 4.7 before switching. Run the exact prompts your code relies on several times and check for refusals. A refusal here returns HTTP 200 with finish_reason: "stop", so it looks like a success. Our empty-response troubleshooting post covers the same trap on other providers.

Grok 4.5 is still the cheaper of the two at the same list price. Our August test found it cheap against the flagships, and its smaller overhead keeps it cheap against its successor.

Try it yourself

• Run the ten-word probe — the prompt Grok 4.7 refused twice; run it a few times
• Run the three-bullet task — both Grok models passed it 5/5

No account and no API key needed. To see prompt, completion and total tokens side by side on your own prompts, use Promptster.

Run 2026-10-19 against grok-4.7 and grok-4.5 through Promptster's provider handlers: three script-graded tasks, five runs each, temperature 0.7, max_tokens 1500, plus a ten-word token-overhead probe run three times per model. Costs are xAI's reported billed cost (cost_in_usd_ticks), repriced on 2026-10-22 from the stored usage objects. They equal list rates from shared/pricing.ts ($2/$6 for both) with reasoning tokens priced as output and cached input at $0.50 (Grok 4.7) or $0.30 (Grok 4.5) per million. Latency is wall-clock from one machine and is not a host SLA. Raw responses and usage objects for every call are in content/blog/data/q4-2026-battery/ and content/blog/data/q4-2026-token-overhead/.