Monday morning, xAI dropped a new flagship model into the middle of one of the most chaotic AI launch weeks we’ve had all year. And honestly? This one might be the most interesting of the bunch.
Grok 4.7 is xAI’s new model for coding, agentic work, and knowledge tasks. The pitch is unusual in an industry that usually reserves fanfare for price hikes: same price as Grok 4.6 — $2 per million input tokens, $6 per million output — but built on a bigger base model with a longer, harder training run.
I’ve spent the last couple of days digging through xAI’s docs, their benchmark table, and some sharp independent analysis to figure out what actually changed, what it really costs (spoiler: it’s not quite as simple as “$2/$6”), and whether your startup should care. Here’s everything I found.
What Grok 4.7 Actually Is (And Isn’t)
First, some context. This week has been absolutely packed with model releases — OpenAI shipped cheaper GPT-6 Sol and Luna, Anthropic shipped Claude Opus 5.5, and as Fast Company noted, most of those are repackaged versions of existing flagships with new price tags.
Grok 4.7 is different. It’s a legitimately new flagship model — new base model, not a fine-tune of 4.6, trained with a longer reinforcement learning run weighted toward tasks that take hours to finish, not seconds. That’s xAI’s description from their announcement, and it’s as specific as they get: no parameter count published, no architecture details.
Here’s what the official xAI docs confirm:
- Model ID:
grok-4.7— live now through the xAI API, Cursor, Grok Build, OpenRouter, Vercel, and Cloudflare - Price: $2.00 per million input tokens, $6.00 per million output tokens
- Context window: 500,000 tokens
- Knowledge cutoff: May 2026
- Modalities: text and image input, text output
- Reasoning effort: four levels — low, medium, high (the default), and xhigh
- Self-verification: upgraded ability to check its own output, per xAI
- Native Grok Bot harness support: wired into the conversational agent framework xAI uses for knowledge work
What it isn’t: an open model. The weights stay closed — API access only. And despite the “SpaceXAI” branding on some of xAI’s docs, it’s the same xAI model pipeline behind it.
The Price Story Everybody’s Getting Wrong
Real talk: “$2/$6” is true and also misleading. There are three pricing wrinkles that actually change what you pay, and I haven’t seen a single launch-day headline get all three right.
The headline rate is the easy part
$2 per million input tokens and $6 per million output tokens is genuinely competitive for a Western frontier coding model. For comparison, Claude Fable 5.1 runs about $10/$50 — that’s five times the input cost and over eight times the output cost. GPT-5.6 Sol sits around $4/$20, roughly double Grok 4.7’s input and triple the output. The recent price-war releases from Anthropic and OpenAI moved the market, and xAI is answering with straight-up price-performance.
If you’re running an AI product and burning through output tokens — which is where the real money goes in coding and agent workloads — that 3x–8x difference against the competition is not small. This is the kind of margin math we broke down in our guide to what an AI business actually costs to run: API bills are often your biggest variable cost, so the per-token rate deserves real attention.
The 200,000-token cliff
Here’s the wrinkle most people will miss. Grok 4.7 inherits Grok 4.6’s tiered pricing: once a single request’s prompt crosses 200,000 tokens, the whole request gets billed at double the rate — $4 input, $1 cached input, $12 output per million tokens.
That matters more than you’d think. A 500K context window practically invites you to dump entire codebases into the prompt. Do that in an agentic coding session and congratulations, your “$2 model” just billed you at $4. Cached input ($0.50 under the threshold) helps if your system prompts and repo snapshots repeat across calls — caching is basically free money for agent loops — but the cliff doubles that rate too.
The practical takeaway: Digital Applied’s teardown of the launch nailed it — the number to watch for agentic coding isn’t the headline rate, it’s what share of your sessions cross 200,000 prompt tokens. Price the switch at $4/$12 if a meaningful chunk of your traffic runs long.
The fast variant nobody can buy (directly)
There’s also a Grok 4.7 Fast: same model, twice the token rates, twice the output speed. But it’s only available inside Cursor and Grok Build — not on the public xAI API. So if your coding harness offers a “fast” toggle, that toggle is a 2x bill. No judgment, just know what you’re clicking.
The route price vs. the list price
One more wrinkle, and this one’s in your favor. The day after launch, Grok 4.7’s route on OpenRouter was about 20% under xAI’s list price — roughly $1.60/$4.80 instead of $2/$6. Routes change without notice, so don’t build your business plan around it, but if you access the model through a router instead of directly, you may simply pay less. Meanwhile GitHub — which rolled Grok 4.7 out to Copilot across all plans (Pro, Pro+, Max, Business, and Enterprise) on launch day — bills at provider list pricing under usage-based billing. Same model, different bills depending on the door you walk through.
“Served at the same price and speed as Grok 4.6, it is highly competitive in its class.” — xAI, Introducing Grok 4.7, September 21, 2026
That’s xAI’s own framing, and honestly, it’s fair — with the asterisk that “competitive in its class” is doing a lot of work once you look at the benchmarks.
Benchmarks, With the Receipts
Here’s where I want to be careful, because vendor benchmark tables are marketing documents with numbers in them. xAI published a table stacking Grok 4.7 against Grok 4.6, GPT-5.6 Sol Max, and Claude Fable 5.1 Max. Every figure was measured and published by xAI — their harness, their runs, their competitor scores. Digital Applied’s breakdown of the table is worth reading precisely because it labels the provenance of every row.
The shape of the results is consistent: Grok 4.7 beats Grok 4.6 on every row, by 3.8 to 17.7 points on the percentage metrics. It leads all three other models on EEBench (electrical engineering, 64.0%) and the Harvey legal agent benchmark (19.6%). The headline jump is Terminal-Bench 4.0 — multi-hour terminal work — from 20.3% on 4.6 to 38.0% on 4.7. That last one matters most for agentic coding, because it’s the closest thing to “turn the agent loose in a terminal and see what happens.”
But it trails Claude Fable 5.1 on the coding rows most buyers weigh first — CursorBench 4.0 (46.3% vs. Fable’s 51.8% per xAI’s runs) and Terminal-Bench 4.0 (38.0% vs. 57.9%) — plus the office-work and clinical rows. On DeepSWE v1.1 it beats Fable 5.1 and trails GPT-5.6 Sol.
And here’s the important caveat: these are all xAI-run numbers. Independent runs by Artificial Analysis put Grok 4.7 at 46 on their Intelligence Index — mid-pack, with Claude Fable 5.1 and GPT-6 at 53 each — while noting the model burns roughly twice as many output tokens per task as Grok 4.6. That’s the hidden cost of those gains: longer reasoning traces mean more output tokens, and output tokens are the expensive half of the bill.
So the honest read: a real step up from 4.6, genuinely competitive on price-performance, but not the top of the coding leaderboard. xAI’s actual claim is about price-performance on coding tasks — and at $2 against $10 input, that claim is about the denominator, not the numerator.
How to Actually Use Grok 4.7 in Your Startup
Okay, enough scoreboard-watching. Here’s the practical part — the four ways to actually get your hands on this thing this week.
1. In your coding editor (the easy path)
This is where Grok 4.7 will reach most developers first. If you use Cursor, it’s already there — Cursor and Grok Build are xAI’s own distribution, and they’re the only places you can toggle the Fast variant at 2x speed and 2x price. GitHub Copilot users get it through the standard model picker now that the rollout has hit all plans, billed at list pricing under usage-based billing.
No integration work, no API keys, no billing dashboards to learn. If you’re a solo founder or a tiny team and your editor is where the code happens, this is probably your move. Just check your Copilot or Cursor usage report after a week so the output-token burn doesn’t surprise you.
2. Through the xAI API (the builder path)
Building an AI feature into your product? The API is OpenAI-compatible, so if your stack already talks to OpenAI’s API, pointing it at https://api.x.ai/v1 with your XAI_API_KEY is the whole migration. The model ID is grok-4.7, and you control reasoning depth with the reasoning_effort field (low, medium, high, xhigh — high is the default).
Here’s roughly what a call looks like:
curl https://api.x.ai/v1/chat/completions \ -H "Authorization: Bearer $XAI_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "grok-4.7", "reasoning_effort": "high", "messages": [ {"role": "system", "content": "You are a senior full-stack engineer reviewing this codebase for bugs and security issues."}, {"role": "user", "content": "Review the attached module and list every issue you find, ranked by severity."} ] }'The reasoning-effort knob is the interesting part. xAI reports its headline coding scores at xhigh — but xhigh burns more output tokens, which costs more money. For everyday work, the default high is the sensible starting point; reach for xhigh on the genuinely hard problems (the multi-hour debugging sessions, the architecture refactors), where the extra reasoning actually pays for itself.
3. Through OpenRouter (the price-shopper path)
If you want the cheapest route to the model, check OpenRouter’s price on the day you switch. On September 22nd it sat about 20% under list. The route also exposes the 500K context and the four effort levels. Just remember routes can change pricing without notice — record what you paid and reconcile monthly, especially if you’re running agent loops at volume.
4. In your agent stack (the ambitious path)
Grok 4.7 was explicitly trained for long-horizon agentic work, and xAI ships it with native support for their Grok Bot harness plus agentic tool calling. If you’re building a coding agent or a knowledge-work agent for your startup, this is the model xAI wants you to build on — and the 500K context window plus cached input at $0.50 means multi-step loops with big context can stay affordable, as long as you stay under that 200K-per-request cliff.
A quick note for teams already on the AI tooling stack most founders run in 2026: Grok 4.7 doesn’t require you to rip anything out. It’s a drop-in option in the harnesses you already use. The switch is a config change, not a migration.
Where It Fits Against Claude, GPT-6, and DeepSeek
Look, nobody buys models in a vacuum. Here’s the competitive picture as of this week:
- vs. Claude Fable 5.1 ($10/$50): Fable leads on raw coding benchmarks — CursorBench, Terminal-Bench, the works. Grok 4.7 costs a fifth of the input price and roughly an eighth of the output price. If your workload is token-heavy and Fable-level scores aren’t mission-critical, the math heavily favors Grok. If you need the absolute best agentic coding, Fable still wears the crown.
- vs. GPT-6 family: The new Sol and Luna models undercut on price in this week’s AI price war, while the flagship Astra leads benchmarks (73% on OSWorld-style tasks, near the top of Artificial Analysis). Grok 4.7 sits between — cheaper than Astra, pricier than Luna, with its own benchmark profile.
- vs. DeepSeek V4.1-Flash: DeepSeek’s model edges past Grok 4.7 on some independent coding tests at aggressive pricing, with an MoE architecture designed for cheaper multi-step agent loops. If your startup’s biggest line item is agent-loop tokens, DeepSeek deserves a slot in your evaluation set too.
And here’s the line from Digital Applied’s related cost-per-task analysis that stuck with me: Grok 4.7 lists at under a third of Claude Opus 5.5’s output price, yet can cost more per completed task in independent tests because it uses more tokens per task. List price is not task price. Run your own numbers.
Should Your Startup Switch? Three Questions
A new row on a vendor’s benchmark table is not a reason to change the model behind your product. These three questions are — and only your own usage can answer them:
- Does the gain hold on your tasks, at the effort you actually run? xAI’s coding scores are at xhigh; your harness probably defaults to high. Re-run 20–50 of your representative tasks at both effort levels with your current model as the control. If you don’t have an eval set, this week is a good excuse to build one — it pays for itself on every future model decision.
- What share of your sessions cross 200,000 prompt tokens? Pull last month’s usage from your billing export. If a meaningful share runs long, price the switch at $4/$12, not $2/$6, and compare against your incumbent’s long-context rate rather than its headline price.
- Which price will you actually pay — list, route, or harness rate? The API list, the OpenRouter route, and Copilot’s list-price billing are three different numbers for the same model. The one in your contract is the one that matters.
If the answers favor the switch, roll it out to one team for two weeks with the previous model still configured as a fallback. That’s how you turn a launch post into a decision you can defend.
FAQ
Is Grok 4.7 free to use?
Not through the API — it’s $2 per million input tokens and $6 per million output tokens. But if you have a GitHub Copilot plan, Grok 4.7 is now available in the model picker on all plans at no extra charge beyond your normal usage-based billing. Cursor users can also access it within their existing subscriptions.
How does Grok 4.7 compare to Grok 4.6?
Same price, same speed, bigger model. It’s built on a new, larger base model with a longer reinforcement learning run focused on multi-hour tasks. xAI’s own numbers show gains on every benchmark — most notably Terminal-Bench 4.0 jumping from 20.3% to 38.0%. If you’re on 4.6, there’s very little reason not to move to 4.7.
What’s the catch with the 200,000-token pricing tier?
When a single request’s prompt crosses 200,000 tokens, the entire request is billed at double the rate ($4 input, $1 cached input, $12 output per million tokens). Long agentic coding sessions with big codebases in context can cross this line easily. Cached input and keeping prompts lean help; blindly dumping repos into a 500K context window does not.
Can Grok 4.7 browse the web or see images?
It accepts text and image input (up to 20MB per image), but it has no knowledge of events past its May 2026 training cutoff — you need to enable xAI’s server-side web search or X search tools for realtime data. Image input works for screenshots, diagrams, and UI mockups; output is text only.
Should a non-technical startup founder care about this?
Honestly? At the level of “which model should I pick,” probably not yet — that’s a decision for your technical co-founder or your agent vendor. But at the level of “AI coding is getting dramatically cheaper,” yes: a flagship-class coding model at $2/$6 means the AI-powered products and services you could build or buy just got more affordable to run. That’s the part that affects your margins.
The Bottom Line
Grok 4.7 is the real deal in a week full of repackaged releases — a genuinely new flagship model that got meaningfully better at long-horizon coding work without moving the price. The $2/$6 headline is competitive, the 500K context is generous, and it’s already sitting in the tools your developers use every day.
But the honest version has three asterisks: it still trails Claude Fable 5.1 on the coding benchmarks that matter most, the 200K-token pricing cliff can quietly double your bill, and it burns more tokens per task than its predecessor — so list price and task price aren’t the same thing.
My take? If you’re building with AI agents or shipping code with AI assistance, put Grok 4.7 in your evaluation set this week. Run your own tasks, check your own token usage, and let your numbers decide. The vendors will keep launching models every eleven days or so — your job is to be the person who measures instead of the person who just switches. If you’d rather not hand-roll the integration yourself, an AI integration studio like AISquadX can wire it up for you.
