xAI shipped Grok 4.7 on Monday, September 21, 2026, and the pitch is simple: pay the same price as last generation, get a bigger model. What the launch didn’t come with was a clean benchmark sweep. Multiple outlets, led by The Decoder, published test results the same week showing Grok 4.7 landing well behind Anthropic’s Claude Fable 5.1 and OpenAI’s GPT-6 Astra on several frontier reasoning and agentic-coding benchmarks, even as it undercuts both on price by a wide margin.
The result is an odd split for a launch day: xAI gets to claim the cheapest frontier-class model on the market, while reviewers get to claim it’s not actually frontier-class on some of the tasks that matter most to paying developers. Both things appear to be true at once, and that tension is what’s driving the search traffic and the trade coverage today.
xAI’s Grok 4.7 Launch: The Key Facts
Grok 4.7 succeeds Grok 4.6 and runs on what xAI describes as a larger base model, positioned specifically as a coding and knowledge-work tool rather than a general chat upgrade. It’s live now through the Grok API, through grok.com, and in the iOS and Android Grok apps, alongside distribution through Cursor and, per OfficeChai, GitHub Copilot. Coverage from The Neuron frames the release as another shot in what’s become an open price war among frontier labs rather than a pure capability play.
xAI kept the headline number unchanged from Grok 4.6: $2 per million input tokens and $6 per million output tokens for prompts under 200,000 tokens. That’s the detail every report leads with, because it means developers who already budgeted for Grok 4.6 get a capability bump without a line-item increase. The catch, as multiple benchmark trackers have now shown, is that “bigger” doesn’t mean “better across the board.”
How Grok 4.7 Pricing Works
Grok 4.7’s pricing has four moving pieces: a standard tier, a long-context tier, cached-token discounts, and a faster paid variant. For prompts below 200,000 tokens, xAI charges $2 per million input tokens and $6 per million output tokens, with cached input billed separately at $0.50 per million tokens. Cross the 200,000-token line and the model shifts to a long-context tier: $4 per million input tokens, $12 per million output tokens, and $1 per million cached input tokens. The model carries a 500,000-token context window overall, so the long-context tier applies to a meaningful chunk of real coding and document-review workloads, not just edge cases.
xAI also sells a “fast” variant of Grok 4.7 aimed at latency-sensitive coding agents. It runs at roughly double the output speed of the standard model, and it costs exactly double: $4 per million input tokens and $12 per million output tokens on the sub-200K tier. That’s a straightforward speed-for-money trade rather than a different model tier, which keeps the pricing structure easy to reason about even if it adds a decision developers have to make on every deployment.
| Model / Tier | Input ($/M tokens) | Output ($/M tokens) | Context Window | Source |
|---|---|---|---|---|
| Grok 4.7 (standard, <200K) | $2.00 | $6.00 | 500K tokens | xAI / Folkfox pricing scorecard |
| Grok 4.7 (long-context, ≥200K) | $4.00 | $12.00 | 500K tokens | xAI / Folkfox pricing scorecard |
| Grok 4.7 Fast (standard tier) | $4.00 | $12.00 | 500K tokens | xAI API docs |
| Grok 4.7 via OpenRouter | $1.60 | $4.80 | 500K tokens | OpenRouter listing |
| GPT-5.6 Sol (max) | $4.00 | $20.00 | Varies by tier | BitsMinds benchmark table |
| Claude Fable 5.1 (max) | $10.00 | $50.00 | Varies by tier | BitsMinds benchmark table |
Two things jump out of that table. First, Grok 4.7’s list price undercuts GPT-5.6 Sol on output tokens by more than 3x and Claude Fable 5.1 by roughly 8x on output tokens, which is the basis for outlets like Forkast describing it as sharply cheaper than comparable Anthropic and OpenAI tiers. Second, OpenRouter is reselling Grok 4.7 below xAI’s own direct API price, at $1.60 input and $4.80 output, plus a $0.40-per-million cache-read rate and $5 per 1,000 web-search calls. That gap between xAI’s list price and third-party router pricing is unusual and worth watching if it persists past launch week.
Grok 4.7 vs GPT-6 Astra and Claude Fable 5.1: The Benchmark Gap
This is where the “bargain” framing runs into trouble. On the Artificial Analysis Intelligence Index (version 4.3.2), Grok 4.7 scores 46, landing mid-pack, while Claude Fable 5.1 and GPT-6 Astra both score 53. That’s a real gap on a composite reasoning benchmark, not a rounding error, and it’s the number The Decoder’s report leans on most heavily to argue that Grok 4.7 is not actually competing at the frontier despite the marketing.
The picture gets more complicated on task-specific benchmarks, where Grok 4.7 actually wins in places. On CursorBench 4.0, a benchmark built around real coding-agent workflows inside the Cursor editor, Grok 4.7 scored 46.3%, ahead of GPT-5.6 Sol’s 41.7%. On DeepSWE v1.1, a software-engineering benchmark, Grok 4.7 hit 71.0%, edging out Claude Fable 5.1 Max at 70.0%, though it still trailed GPT-5.6 Sol Max at 72.7%. Those are close results, well within the range where prompt design and tooling integration can swing the outcome.
Terminal-Bench 4.0 is where the gap turns into a canyon. Grok 4.7 scored just 26% on this benchmark, which tests multi-step command-line and agentic tool use, compared with 60% for GPT-6 Astra and 55% for Claude Fable 5.1. That’s not a close call by any reading, and it’s the single stat most likely to matter to teams evaluating Grok 4.7 for autonomous coding agents rather than single-shot code completion.
| Benchmark | Grok 4.7 | GPT-6 Astra | Claude Fable 5.1 | GPT-5.6 Sol |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index v4.3.2 | 46 | 53 | 53 | — |
| Terminal-Bench 4.0 | 26% | 60% | 55% | — |
| CursorBench 4.0 | 46.3% | — | — | 41.7% |
| DeepSWE v1.1 | 71.0% | — | 70.0% (Max) | 72.7% (Max) |
| Cost per task (ArtificialAnalysis, “high” tier) | $2.73 (lowest) | Higher | Higher | Higher |
Read together, the tables tell a consistent story: Grok 4.7 is genuinely competitive, sometimes ahead, on tasks that resemble a single coding session with a human in the loop. It falls sharply behind on tasks that require sustained, multi-step autonomous tool use, which is exactly the direction agentic coding products are heading. That split matters more than the raw index score, because it points to where xAI’s engineering effort has and hasn’t landed yet.
Why Grok 4.7 Trails on Terminal-Bench but Leads on Cost
xAI has never framed Grok as an agentic-first model the way OpenAI has positioned GPT-6 Astra or Anthropic has positioned Claude Fable 5.1, both of which shipped with heavy emphasis on long-horizon tool use and autonomous task completion. Grok’s roadmap has leaned more on raw reasoning throughput and cost efficiency, which tracks with the Terminal-Bench result. A model tuned for single-turn accuracy and price competitiveness isn’t automatically good at chaining dozens of terminal commands without losing state, and Grok 4.7’s 26% score suggests that gap hasn’t closed with this release.
The cost side of the ledger is the flip side of the same design choice. Artificial Analysis puts Grok 4.7’s effective cost per task at $2.73 in its “high” reasoning-effort tier, the lowest of the models it compared, with per-task prices across the tested field varying by up to 1.4x. For workloads where accuracy differences are small and volume is high, that cost advantage can outweigh a benchmark deficit. For workloads where a single dropped step means a broken deploy or a corrupted database migration, the Terminal-Bench gap is the number that matters, not the price.
The Coding Push: Cursor and GitHub Copilot Integration
xAI is leaning hard into developer distribution rather than consumer chat growth with this release. Grok 4.7 shipped built into Cursor across desktop, web, and iOS, plus Cursor’s command-line tools and SDKs, and OfficeChai reports it rolled into GitHub Copilot essentially at launch. That’s a deliberate strategy: rather than compete purely on chatbot mindshare against ChatGPT and Gemini, xAI is trying to become a default model option inside the tools developers already have open all day.
It’s a similar playbook to what Zhipu ran with GLM-5.2, undercutting incumbent pricing while pushing hard into IDE and agent integrations rather than standalone apps. The difference is that GLM’s benchmark story was closer to parity with GPT-class models on cost-adjusted performance, while Grok 4.7’s Terminal-Bench score gives critics an easy talking point about whether cheap and capable can coexist in the same release.
Cached Tokens and Long-Context Billing
For teams running Grok 4.7 against large codebases or long document chains, the 200,000-token pricing cliff is worth planning around. Below that line, cached input costs $0.50 per million tokens; cross it, and cached input jumps to $1.00 per million while both input and output rates double. Folkfox’s pricing scorecard confirms these numbers are consistent across xAI’s launch page, its API pricing page, and its developer docs, so there’s no ambiguity about the tiers, just a planning cost for anyone whose average prompt sits near the boundary.
The “Fast” Variant Tradeoff
The fast variant is a clean lever rather than a hidden fee: pay double, get roughly double the output speed. For latency-sensitive agent loops where wall-clock time is the bottleneck, not token cost, that’s a reasonable option. For high-volume batch jobs where price per token is the deciding factor, the standard tier remains the better default, and it’s the one most of the pricing coverage treats as the headline number.
OpenRouter Pricing Undercuts xAI’s Own List Price
One detail that hasn’t gotten much attention outside model-router circles: OpenRouter lists Grok 4.7 at $1.60 per million input tokens and $4.80 per million output tokens, which is 20% below xAI’s own direct list price on both ends. That’s not typical for a same-day launch, and it usually reflects either promotional routing credits, volume-based wholesale pricing, or a router absorbing margin to win developer traffic away from direct API integration. Whichever it is, developers price-shopping Grok 4.7 have a real incentive to check router pricing before defaulting to xAI’s own endpoint.
Market Reaction and Competitive Pressure
The launch lands in a market where every major lab is now shipping within weeks of each other. Anthropic’s Claude Fable 5.1 and OpenAI’s GPT-6 Astra have both been out long enough to accumulate benchmark history, and GPT-6 Astra’s training run alone reportedly used 100,000 GPUs, a scale that helps explain why it’s outperforming cheaper rivals on the hardest agentic tests. Grok 4.7 arriving at flat pricing with a mixed benchmark record won’t dislodge either from the top of the leaderboard, but it does put fresh pressure on mid-tier providers who were already competing mostly on price.
DeepSeek’s V4 Pro, for instance, scores 53 on a comparable intelligence index but raised its own prices 3.6x earlier this year, moving in the opposite direction from xAI. That divergence, one lab cutting effective cost while another raises it, suggests the market hasn’t settled on whether frontier-adjacent performance justifies premium pricing or whether volume and distribution matter more. Grok 4.7’s launch is effectively a bet on the second answer.
Historical Context: From Grok 4.6 to Grok 4.7
xAI has shipped Grok updates on a roughly quarterly cadence since Grok 4 first launched, with each release framed around narrowing the gap to GPT and Claude flagship models rather than leapfrogging them outright. Grok 4.6 held that same $2/$6 pricing structure, and TradingKey and BitsMinds both noted that Grok 4.7’s decision to keep pricing flat while adding capability was a deliberate signal that xAI isn’t trying to monetize incremental gains, at least not yet. That’s a different approach from OpenAI and Anthropic, both of which have introduced new pricing tiers alongside capability jumps in GPT-6 Astra and Claude Fable 5.1.
The pattern of “same price, bigger model, mixed benchmarks” has repeated across at least two Grok release cycles now, which suggests xAI is optimizing its release cadence around predictable pricing rather than benchmark-topping headlines. Whether that’s a sustainable strategy against labs with far larger training budgets is one of the open questions this launch raises rather than answers.
How This Fits the Broader Price War in Frontier AI
The Neuron’s framing of this launch as a “price war” rather than a “capability war” reflects a real shift in how frontier labs are competing. When Anthropic and OpenAI both anchor flagship pricing well above xAI’s $2/$6 structure, on Claude Fable 5.1’s $10/$50 max tier and GPT-5.6 Sol’s $4/$20 tier, they’re implicitly betting that measurable performance gaps on hard benchmarks justify the premium. xAI’s bet is that most production workloads don’t actually need Terminal-Bench-level agentic reliability, and that cost-sensitive developers will tolerate the gap in exchange for a 3-8x price cut on output tokens.
That’s not a new argument in the broader AI industry. Even as pricing gets more aggressive, some lab leaders have publicly urged the industry to slow down capability races rather than accelerate them, a tension worth watching as more labs ship cost-cutting releases without matching benchmark gains. A detailed breakdown of Grok 4.7’s benchmark suite against GPT and Claude models is available from Kingy AI.
What Enterprises and Developers Should Weigh
For teams evaluating Grok 4.7 today, the decision comes down to workload shape rather than a single leaderboard number. High-volume, single-turn coding assistance, code review, documentation generation, straightforward chat, is exactly where Grok 4.7’s CursorBench and DeepSWE scores hold up against far pricier rivals. Long-horizon autonomous agents that chain dozens of terminal commands or tool calls without human correction are where the Terminal-Bench gap becomes a real reliability risk, not just a benchmark footnote.
- Budget-constrained teams running high-volume, single-step coding tasks: Grok 4.7’s price-to-performance ratio on CursorBench and DeepSWE is hard to beat.
- Teams building autonomous, multi-step coding agents: the Terminal-Bench gap versus GPT-6 Astra and Claude Fable 5.1 is large enough to warrant direct testing before committing.
- Long-document or large-codebase workloads: plan around the 200,000-token pricing cliff, since it doubles both input and output rates.
- Cost-sensitive deployments already using a model router: check OpenRouter’s listed rates, which currently undercut xAI’s direct API pricing.
Predictions: Where the Grok vs GPT-6 vs Claude Race Goes Next
A few things look likely to play out over the next two to three release cycles. First, expect xAI to prioritize a Terminal-Bench-style fix in the next point release; a 34-point gap against GPT-6 Astra is too visible to leave unaddressed, and it’s already the single most-cited number in this week’s coverage. Second, expect OpenRouter-style price undercutting to keep spreading as more labs release models through third-party routers, putting pressure on direct API pricing across the board, not just at xAI.
Third, watch for Anthropic and OpenAI to hold their premium pricing rather than match Grok 4.7’s cuts, betting that the benchmark gap alone is enough to retain enterprise customers who can absorb the cost difference. Fourth, expect more coding tools beyond Cursor and GitHub Copilot to add Grok 4.7 as a selectable model option within the next month, given how aggressively xAI is pushing IDE distribution. Fifth, and more speculatively, if the cost-per-task advantage Artificial Analysis measured holds up under independent testing, expect at least one more mid-tier lab to follow xAI’s flat-pricing strategy rather than raising prices alongside capability gains, the opposite of what DeepSeek did with its V4 Pro price hike earlier this year.
Competitive Landscape Comparison
Stacking Grok 4.7 against its two closest frontier rivals makes the trade-off explicit. GPT-6 Astra leads on the hardest agentic benchmark by a wide margin and carries a training investment, reportedly a 100,000-GPU run, that Grok 4.7 can’t match on paper. Claude Fable 5.1 sits close behind GPT-6 Astra on the composite index and comfortably ahead of Grok 4.7 on Terminal-Bench, while pricing itself at the top of the market. Grok 4.7 is the value play of the three: cheapest by a wide margin, competitive on single-turn coding tasks, and clearly behind on sustained agentic reliability. None of the three is a strict upgrade over the others across every metric, which is unusual for a market that spent most of 2025 consolidating around one or two clear leaders.
That fragmentation is arguably the bigger story than any single benchmark. A year ago, choosing a frontier model was mostly a question of which lab’s flagship was current. Today it’s a workload-by-workload calculation involving price tiers, context-window cliffs, router discounts, and task-specific benchmark deltas, exactly the kind of complexity that pushes procurement decisions toward whoever publishes the clearest pricing table, which right now looks like xAI.
Frequently Asked Questions
What is Grok 4.7 and when did it launch?
Grok 4.7 is xAI’s newest model, positioned for coding and knowledge work, succeeding Grok 4.6. It launched on Monday, September 21, 2026, and is available through the Grok API, the Grok apps, and integrations including Cursor and GitHub Copilot.
How much does Grok 4.7 cost per million tokens?
Standard pricing for prompts under 200,000 tokens is $2 per million input tokens and $6 per million output tokens, with cached input at $0.50 per million tokens. Prompts at or above 200,000 tokens bill at $4 input, $12 output, and $1 cached input per million tokens.
Is Grok 4.7 better than GPT-6 Astra or Claude Fable 5.1?
It depends on the task. Grok 4.7 outperforms GPT-5.6 Sol on CursorBench 4.0 and edges out Claude Fable 5.1 Max on DeepSWE v1.1, but it trails both GPT-6 Astra (60%) and Claude Fable 5.1 (55%) badly on Terminal-Bench 4.0, where it scores just 26%. On the Artificial Analysis Intelligence Index, Grok 4.7 scores 46 versus 53 for both rivals.
Why is Grok 4.7 cheaper than other frontier models?
xAI kept Grok 4.7’s pricing identical to Grok 4.6 rather than raising it alongside the capability increase. That puts it well below Claude Fable 5.1’s max tier ($10 input / $50 output per million tokens) and GPT-5.6 Sol’s max tier ($4 input / $20 output), which several outlets have described as roughly an 80% discount versus comparable Anthropic and OpenAI tiers.
What is Terminal-Bench and why does Grok 4.7 score low on it?
Terminal-Bench 4.0 tests multi-step, autonomous command-line and tool-use tasks, the kind of sustained agentic work coding assistants increasingly need to handle without human correction. Grok 4.7’s 26% score suggests xAI’s tuning has prioritized single-turn reasoning and cost efficiency over long-horizon agent reliability, unlike GPT-6 Astra and Claude Fable 5.1.
Can I use Grok 4.7 in Cursor or GitHub Copilot?
Yes. Grok 4.7 is integrated into Cursor across desktop, web, and iOS, plus Cursor’s command-line tools and SDKs, and according to OfficeChai it rolled out inside GitHub Copilot at launch.
Is Grok 4.7 cheaper through OpenRouter than directly from xAI?
Yes. OpenRouter lists Grok 4.7 at $1.60 per million input tokens and $4.80 per million output tokens, about 20% below xAI’s own direct API pricing of $2 input / $6 output, plus a separate $0.40-per-million cache-read rate.
What is Grok 4.7’s context window?
Grok 4.7 supports a 500,000-token context window. Pricing shifts to a higher tier once a prompt crosses 200,000 tokens, so very long documents or large codebases cost roughly double per token compared with shorter prompts.
