Frontier models got cheaper again this week. xAI shipped Grok 4.7 on 21 Sep 2026 while freelancers were already juggling Claude and GPT desks, spend caps, and client quotes that assume last month’s rates. The launch post talks price-performance and coding benches. Useful for vendors. Less useful when you need a job → desk → meter answer before Friday.
This is a try / API / stay / skip sheet. Not a benchmark dump. Not a rewrite of our Claude vs OpenAI API cost ladder. Cursor workspace habits stay on the Cursor Projects post. Here the only question is whether grok-4.7 earns a third seat — inside Cursor first, via the public Grok API only when the SOW needs it, or not at all.
Affiliate disclosure: some tool or platform links may later be affiliate or referral. Prices below are from xAI’s live pricing and models docs as of 25 Sep 2026 (SAST) — re-check before you quote a client. Not legal advice; not a guarantee of clients, savings, or income.
What shipped 21 Sep 2026 — Grok 4.7 in plain English
On 21 Sep 2026, xAI published Introducing Grok 4.7. The model ID on the Grok API is grok-4.7. Launch positioning is coding and knowledge work: longer runs on hard tasks, more self-checks, same list price and speed as Grok 4.6 for the standard path. The models docs put it bluntly — for everything else including code, use Grok 4.7. Treat that as a vendor claim, not a Mr1Tech bake-off.
Availability on the launch post: Cursor, Grok Build, the Grok API, plus third-party coding harnesses, routers, and cloud platforms. Context window is 500k tokens. Knowledge cut-off is May 2026 — so realtime facts still need search tools if the brief needs them. Do not invent ChatGPT-style consumer plan gates for Grok; quote what Cursor’s own plan screen and the xAI Console show you at signup.
There is also a Fast variant: same model, faster infrastructure, twice the standard token rates. Docs are clear that Fast is Cursor and Grok Build only. It is not on the public xAI API. If your only key is an API key, Fast is not a desk you can flip on from the console. Grok Build’s free tier also does not include Fast — read the Build screen before you promise a client “the fast path.”
For freelancers the practical read is simple: standard grok-4.7 is the API and Cursor try path; Fast is an IDE/Build option you only meter after you know Cursor’s own billing language for that variant.
The price sheet freelancers actually need
USD list rates for grok-4.7 match Grok 4.6 on the xAI pricing page. Under 200k prompt tokens: $2.00 input / $0.50 cached input / $6.00 output per 1M tokens. At or above 200k prompt tokens, the whole request bills at the long-context row: $4.00 / $1.00 / $12.00. That cliff is the line freelancers miss when they dump a giant repo into one prompt “because context is 500k.”
US regional endpoint https://us.api.x.ai/v1 runs at 1.1× global token rates (docs example under 200k: $2.20 / $0.55 / $6.60). Batch discount tables list older models at 20% off; grok-4.7 is not on that batch-discount list as of this draft’s re-pull — do not quote a batch saving you have not confirmed on the model detail page.
Grok 4.7 Fast (Cursor / Grok Build only): under 200k $4 / $1 / $12; above 200k $6 / $1.50 / $18 per 1M. Cursor may bill its own fast variant through your Cursor plan — read Cursor’s meter language; do not invent agent quotas here. Server-side tools (web search, code execution, and friends) add separate invocation rows on the same pricing page if you enable them. Agents API–style harness meters on OpenAI are a different product; do not mash those into an xAI quote.
Rand clients still see USD API invoices. Card and billing-country quirks at Console or Cursor signup are real for international freelancers — verify at signup; do not invent decline rates or FX markups.
Cursor desk vs API desk
Try the model switch inside an existing Cursor plan first. Same repo, same brief, one finished job you already shipped on Claude or GPT. That answers quality and rework without opening a second vendor invoice. Cursor Projects habits (folders, rules, ship flow) are covered on the sibling post; this section is only which model you point at the job.
Move to the public Grok API when the deliverable needs a client-owned key, a hard USD spend cap you control outside the IDE, or a harness / router that is not Cursor. Self-hosted or partner sandboxes on other vendors do not change xAI’s token table. If the SOW only says “use AI to go faster,” a third API key is optional packaging — same caution as any other paid AI desk (see the Microsoft 365 Premium Copilot buy-or-skip framing). Long Cursor agent runs still need a stable uplink; for mobile-data packaging see fibre vs mobile data for AI work.
When Grok 4.7 earns a seat
Give it a seat when coding or knowledge-work jobs already felt close on Grok 4.6 at the same list price, and you want one controlled A/B without rewriting your Claude/GPT stack. xAI frames gains on longer coding tasks, multi-hour office-style work, and legal-adjacent drafts as vendor benchmarks — useful direction, not a win-rate you put in a client proposal. CursorBench and DeepSWE numbers on the launch post are marketing context; your acceptance criteria on a real brief still win.
It also earns a seat when the client already accepts a multi-model desk and you can log meter + rework on one non-secret job before you change the default. Pricing that starts at $2 / $6 per 1M under the 200k cliff is competitive on paper next to many frontier rows; the seat still has to beat your finished Claude or GPT output on the same brief, not a launch chart. If you already run harnesses or routers that list grok-4.7, try there with a spend cap before you rebuild the whole stack around xAI.
Skip inventing “average freelancer Grok bills.” One logged job row beats a guess. If Grok Build is already on your desk for throwaway prototypes, that is a try surface — still not a reason to paste production secrets into a personal key.
When to stay on Claude or GPT
Stay when your Claude or GPT desk already finishes the job inside budget and review time. Sibling Anthropic and OpenAI tier choosers are separate desks this week — do not rewrite them here. If you already metered cache-heavy Anthropic loops or budgeted an OpenAI Agents harness, adding grok-4.7 as a third meter without a failed row is noise, not savings.
Stay when the SOW locks a vendor, a DPA, or a residency path that does not include xAI. Stay when Fast marketing tempts you but you only hold a public API key — Fast is not that path. Stay when the brief needs knowledge after May 2026 and you have not enabled search tools; cut-off is a product limit, not a vibe. Stay when review time on your current stack is already the billable bottleneck — a second model that needs the same human pass does not cut the quote.
Cross-link the live ladder when you need Claude vs OpenAI $/MTok only: API cost: Claude vs OpenAI for freelancers. This post does not replace that sheet.
Buy/skip matrix (sheet)
| Job | Try grok-4.7 in Cursor | Public API + spend cap | Stay Claude / GPT | Skip |
|---|---|---|---|---|
| One finished coding job, non-secret | Yes — first | Only if SOW needs a key | If already inside budget | — |
| Client-owned key / harness outside Cursor | Weak | Yes — hard USD cap | If DPA locks vendor | If residency forbids |
| “Use Fast — twice the speed” | Cursor / Build only | No — not public API | — | If you only have an API key |
| Giant context dump near/over 200k | Maybe — watch cliff | Meter whole request at long rates | Often cheaper split | If no reason for one prompt |
| Third vendor with no failed row | — | — | Yes | Yes — clarify first |
| Knowledge after May 2026, no search tools | Weak | Weak without tools | Yes if your stack has live data | Until tools or newer cut-off |
Get the 1-page Cursor try / API / stay / skip chooser
Optional printable of this free guide. Soft link until checkout. Not a “paid summary.” Not a guarantee of clients, savings, or income.
This week’s action — one finished job, then one sheet row
Pick one finished job you already shipped on Claude or GPT. Same brief, same acceptance criteria, non-secret materials only. Re-run it on Grok 4.7 in Cursor first — standard, not Fast, unless you already know how Cursor bills the fast variant. Log rework minutes and any API meter if you escalate past the IDE. Note whether you tripped the 200k prompt cliff. Fill one sheet row: keep as a Cursor option / open public API with a spend cap / stay on Claude or GPT / skip Fast claims and third-vendor noise.
If the rework surprises you, the surprise is the product — not a reason to hide the meter from the next quote. Do not change your default model mid-client week on a launch post alone. Sources: Introducing Grok 4.7 (21 Sep 2026), xAI pricing, models — re-check model IDs and Fast / regional rows at paste time.
FAQ
Is Grok 4.7 Fast on the public API? No. Docs list Fast for Cursor and Grok Build only, at 2× standard token rates. An xAI API key alone does not unlock that path.
Does the 200k cliff apply to the whole request? Yes. Once the prompt hits the threshold, the long-context row bills input, cached input, and output for that request. Split prompts when you can.
Should I replace Claude or GPT this week? Only if one logged A/B on a real job wins on quality plus meter. Otherwise keep the desk that already ships and treat Grok as an optional third seat.