xAI shipped Grok 4.7 on September 21, 2026, holding the line at $2 per million input tokens and $6 per million output tokens, the same rate card it used for Grok 4.6. That decision, not the model itself, is what’s rattling the coding-assistant market this week. Our earlier coverage of the launch detailed the specs and the benchmark gap to GPT-6. This piece looks at what happens next: how Cursor, Vercel, and enterprise buyers are reacting, and what the pricing standoff means for Anthropic and OpenAI heading into Q4.
Grok 4.7 Lands Into an Already Crowded Price War
Forkast News framed the timing bluntly, noting Grok 4.7 arrived nine days after Elon Musk publicly endorsed Dario Amodei’s call for a coordinated AI slowdown, then priced the new model at what Forkast calls 80% cheaper than comparable Anthropic and OpenAI tiers. Whether or not that figure holds up model-for-model, the optics are hard to miss: a company whose CEO backed the idea of pumping the brakes just cut the effective cost of running long coding-agent sessions.
VentureBeat‘s take on the same day was more measured. Its reporting pairs the “coding gains with the same affordable pricing” framing with a warning that high token consumption threatens real-world ROI, meaning the headline $2/$6 rate can get eaten alive by long-context agent sessions before a team ever notices. That tension between sticker price and effective cost is the real story of Grok 4.7’s launch, more than the model card itself. It’s the same tension we flagged when covering Grok 4.5’s debut earlier this year, where xAI first tried the same low-price, high-volume playbook against OpenAI and Anthropic.
What xAI Actually Shipped
Grok 4.7 keeps the 500,000-token context window xAI introduced with Grok 4.6, and it keeps the tiered pricing structure the company has used since that release. According to pricing data compiled by BenchLM.ai, prompts under 200,000 tokens run at $2 per million input tokens, $0.50 per million cached input tokens, and $6 per million output tokens. Cross the 200,000-token mark and every rate doubles: $4 input, $1 cached input, $12 output. xAI also ships a faster variant at double the standard rate, $4/$12, which multiple distribution channels have set as the default option, according to IT Brief’s reporting on the rollout.
On capability, AI/TLDR’s release notes put Grok 4.7 at 46.3% on a benchmark it calls CursorBench 4.0, up from 40.4% for Grok 4.6. Separately, AI Weekly reports a 71% score on DeepSWE v1.1, a benchmark built around long-duration software engineering tasks rather than one-shot code completion. Those are respectable gains for a model that xAI is not charging any more to run, and they matter because DeepSWE-style benchmarks are designed to punish models that lose track of context across a long session, exactly the failure mode enterprise buyers worry about most when handing an agent a real production codebase.
Grok 4.7 Pricing at a Glance
| Prompt Length | Input (per 1M tokens) | Cached Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|---|
| Under 200,000 tokens | $2.00 | $0.50 | $6.00 |
| 200,000+ tokens | $4.00 | $1.00 | $12.00 |
| Fast variant (default in some tools) | $4.00 | n/a | $12.00 |
Source: BenchLM.ai pricing data and IT Brief reporting on Grok 4.7’s distribution channels, both dated September 21-22, 2026. Note that the 500,000-token context window means a single long agent session can cross the 200,000-token threshold and quietly double the effective bill, which is exactly the concern VentureBeat raised.
Cursor and the Rush to Integrate
The fastest-moving part of this story is distribution. IT Brief reports Grok 4.7 is already live inside Cursor across desktop, web, and iOS, plus command-line tools and software development kits, alongside the standalone Grok API and third-party model routers. That’s a full-surface rollout, not a limited beta, and it means developers who already have Cursor open can pick Grok 4.7 from a model dropdown today without touching a new API key.
Vercel moved just as fast. BriefFlash reports Grok 4.7 landed on Vercel’s AI Gateway at a promotional 40% discount through September 27, cutting the effective rate to $1.20 per million input tokens and $3.60 per million output tokens, well below even xAI’s own list price. For teams already routing model calls through AI Gateway, that’s a low-friction reason to test Grok 4.7 against whatever they’re currently running through GPT-6 Astra or Claude for coding tasks.
RuntimeWire’s coverage adds useful framing here: it describes Grok 4.7 as built specifically for “longer agent work,” which lines up with why coding tools are prioritizing it over chat-focused models. Multi-step refactoring, test-writing loops, and long debugging sessions are exactly the workloads where a 500,000-token context window and per-token pricing start to matter more than raw benchmark scores.
The Benchmark Gap Nobody’s Papering Over
Cheap and fast doesn’t mean it’s winning on capability. As we reported when Grok 4.7 launched, xAI’s own model trailed GPT-6 Astra by roughly 34 points on aggregate benchmark testing, a gap wide enough that no amount of pricing strategy fully closes it for teams that need top-tier reasoning on hard problems. Grok 4.7’s 71% on DeepSWE v1.1 and 46.3% on CursorBench 4.0 are real improvements over Grok 4.6, but neither figure puts it in the same tier as the frontier models from Anthropic or OpenAI on the hardest coding benchmarks.
That’s the core trade xAI is asking developers to make: accept a capability gap in exchange for a cost structure that makes running the model at scale, across thousands of agent sessions a day, dramatically cheaper. For some workloads, code review, boilerplate generation, first-pass test writing, that trade is close to free money. For workloads that need the strongest possible reasoning, the gap still matters.
Price and Capability, Side by Side
| Model | Standard Input / Output (per 1M tokens) | Context Window | Notable Coding Benchmark |
|---|---|---|---|
| Grok 4.7 (xAI) | $2 / $6 | 500,000 tokens | 71% on DeepSWE v1.1 (AI Weekly) |
| GPT-6 Astra (OpenAI) | $10 / $50 | Not directly comparable in these reports | Reported ~34-point aggregate lead over Grok 4.7 in our prior coverage |
| Claude (Anthropic) | Not published in flat per-token terms in current reporting | Varies by release | Anthropic recently cut cache costs 75% on its Fable 5.1 release |
Sources: xAI/BenchLM.ai pricing data, our reporting on GPT-6 Astra’s $10/$50 launch pricing, our coverage of Grok 4.7’s benchmark gap to GPT-6, and our report on Claude Fable 5.1’s cache pricing cut. Anthropic has not published a single flat per-token rate comparable to xAI’s structure in the sources reviewed for this piece, so that cell is deliberately left general rather than estimated.
Why “80% Cheaper” Is a Real Number With Real Limits
Forkast’s 80% figure compares Grok 4.7’s list price against Anthropic and OpenAI’s frontier tiers. It’s a real gap. GPT-6 Astra’s launch pricing of $10 per million input tokens and $50 per million output tokens, which we covered when OpenAI announced it, is five times Grok 4.7’s input rate and more than eight times its output rate. That’s not a rounding error, and it’s the number driving most of this week’s coverage.
But Vandata Team’s analysis makes the more important point for anyone actually budgeting a coding-agent deployment: nominal price per token isn’t the same as cost per completed task. A model that needs more retries, longer context to reach the same result, or the doubled-rate tier past 200,000 tokens can end up costing more per finished pull request than a pricier model that gets there in fewer tokens. Nobody in the current reporting has published a controlled cost-per-task comparison between Grok 4.7 and its rivals, so treat the 80% figure as a starting point, not the full picture.
How We Got Here: xAI’s Pricing Ladder
Grok 4.7 didn’t appear in a vacuum. BenchLM.ai’s pricing archive shows xAI moving through several pricing tiers over the past year: Grok Code Fast 1 launched as an ultra-cheap, narrow coding tool at $0.20 per million input tokens and $1.50 per million output tokens. Grok 4.1 Fast followed at $0.20/$0.50. Grok 4.3 moved up to $1.25/$2.50 as capability grew. Grok 4.5 then jumped to the $2/$6 tier it has now held across three releases.
That pattern says something about strategy: xAI spent its early coding-model era racing to the bottom on price with narrow, specialized models, then pivoted to a flat, higher rate for its general frontier line once it needed the margin to fund bigger context windows and broader capability. Holding $2/$6 flat across Grok 4.5, 4.6, and 4.7 while benchmark scores climb is effectively xAI eating its own cost increases rather than passing them to developers, a bet that market share now matters more than per-token margin.
The Musk-Amodei Irony
Forkast’s framing of the nine-day gap between Musk backing a slowdown and xAI shipping an aggressive price cut isn’t just color. It highlights a real split inside the industry between public rhetoric about pacing AI development responsibly and the competitive pressure to ship faster, cheaper models the moment a rival looks vulnerable. Dario Amodei has been publicly vocal about wanting labs to slow down and coordinate; xAI’s actual product decisions this week point the opposite direction.
None of the current reporting suggests Musk walked back his endorsement, and there’s no indication xAI sees a contradiction between backing a slowdown conversation and competing hard on price today. But the juxtaposition is exactly the kind of detail that keeps showing up in coverage of this launch, and it’s worth watching whether Anthropic or OpenAI respond to the pricing move with commentary of their own.
What This Means for Anthropic and OpenAI
Neither company has announced a pricing response as of this writing. GPT-6 Astra’s $10/$50 rate, covered in our earlier report on its launch and how to try it, gives OpenAI room to compete on capability rather than price, especially given the benchmark gap our prior coverage documented. Anthropic’s recent move to cut cache costs 75% on Claude Fable 5.1 shows the company is already trimming costs on its own terms, separate from any direct reaction to xAI.
The more interesting pressure point may be internal. We previously reported that Anthropic was already weighing a new model release in response to GPT-6 Astra’s benchmark position. A three-way price and capability fight between xAI, OpenAI, and Anthropic, playing out inside the same developer tools like Cursor and Vercel’s AI Gateway, gives enterprise buyers leverage they didn’t have a year ago: the ability to switch coding models with a dropdown menu rather than a procurement cycle.
Market Impact for Enterprise AI Buyers
For engineering leaders running coding agents at scale, the practical takeaway from this week isn’t “switch to Grok 4.7.” It’s that the model-selection decision has become genuinely dynamic. With Vercel’s AI Gateway and similar routers now supporting multiple frontier coding models behind one interface, teams can A/B test Grok 4.7 against GPT-6 Astra or Claude on their own codebase and their own task mix, rather than betting a budget on vendor claims.
That shift matters more than any single price cut. Model routing infrastructure is turning coding-agent selection into a commodity decision made per-task rather than a single annual vendor choice, which puts sustained pressure on every lab to keep both price and benchmark scores moving in the right direction at the same time. VentureBeat’s ROI warning is the practical corollary: run your own cost-per-completed-task numbers before switching, because list price alone won’t tell you which model is actually cheaper for your workload.
There’s also a budgeting wrinkle specific to Grok 4.7’s structure that’s easy to miss on a first read of the pricing page. Because the doubled rate kicks in at 200,000 tokens rather than at some higher ceiling, and the context window stretches to 500,000 tokens, a team that lets agent sessions run long without checkpointing can slide into the expensive tier without any explicit decision to do so. Finance and platform teams evaluating Grok 4.7 alongside GPT-6 Astra or Claude should treat that threshold the same way they’d treat a cloud-compute autoscaling limit: something to monitor and cap, not something to discover on next month’s invoice.
Five Predictions for the Next Quarter
- Expect at least one of OpenAI or Anthropic to announce a pricing adjustment or a new discounted tier for coding-specific workloads before the end of 2026, mirroring xAI’s flat-rate approach.
- Model routers like Vercel’s AI Gateway will keep running short-term promotional pricing on new frontier models to drive trial volume, following the pattern set with Grok 4.7’s 40%-off window.
- Cost-per-completed-task benchmarking, rather than raw per-token pricing, will become a more common way outlets and analysts compare coding models, given the gap VentureBeat and Vandata Team have already flagged.
- xAI will likely keep its $2/$6 rate flat again for at least one more major Grok release, continuing the pattern set across Grok 4.5, 4.6, and 4.7.
- Coverage of AI “slowdown” commentary from lab leaders will face more scrutiny each time it’s followed by an aggressive product or pricing move, as happened with the Musk-Amodei timeline this week.
Frequently Asked Questions
What is Grok 4.7 and when did it launch?
Grok 4.7 is xAI’s latest frontier model for coding and knowledge work, released September 21, 2026, according to multiple outlets including AI/TLDR and TradingKey. It succeeds Grok 4.6 and keeps the same 500,000-token context window.
How much does Grok 4.7 cost to use?
Standard pricing is $2 per million input tokens, $0.50 per million cached input tokens, and $6 per million output tokens for prompts under 200,000 tokens, per BenchLM.ai’s pricing data. Prompts of 200,000 tokens or more double to $4/$1/$12. A faster variant runs at $4 input and $12 output per million tokens.
Is Grok 4.7 really 80% cheaper than Claude and GPT-6?
Forkast News reported that figure comparing Grok 4.7’s list price to Anthropic and OpenAI’s frontier tiers. Against GPT-6 Astra’s published $10/$50 rate, Grok 4.7’s $2/$6 input rate is indeed a fraction of the cost. Anthropic has not published a single comparable flat rate in current reporting, so a precise percentage against Claude specifically isn’t available.
Can I use Grok 4.7 inside Cursor?
Yes. IT Brief reports Grok 4.7 is available inside Cursor on desktop, web, and iOS, as well as through command-line tools, SDKs, the Grok API, and third-party model routers.
Does Grok 4.7 beat GPT-6 or Claude on coding benchmarks?
No. Our earlier reporting found Grok 4.7 trailing GPT-6 Astra by roughly 34 points on aggregate benchmark testing. Grok 4.7 does show gains over its own predecessor, moving from 40.4% to 46.3% on CursorBench 4.0 and scoring 71% on DeepSWE v1.1, according to AI/TLDR and AI Weekly.
What is the Vercel AI Gateway discount on Grok 4.7?
BriefFlash reports a 40% promotional discount on Vercel’s AI Gateway through September 27, 2026, bringing the effective rate to $1.20 per million input tokens and $3.60 per million output tokens.
Will Anthropic or OpenAI cut prices in response?
Neither company has announced a direct pricing response as of this report. Anthropic has already cut cache costs 75% on its Claude Fable 5.1 release, and we previously reported Anthropic was weighing a new model in response to GPT-6 Astra’s benchmark position, though that move wasn’t framed as a response to Grok 4.7 specifically.
