Why this matters: the cheapest frontier model has a hidden tier that doubles the bill on long prompts.
September 27, 2026 — xAI launched Grok 4.7 on September 21 as its new flagship for coding and knowledge work at $2 per million input tokens and $6 per million output — the cheapest published frontier rate — while adding a tier that doubles costs for prompts above 200,000 tokens. Days later, Elon Musk said on X that SpaceX could field a model matching Anthropic's Claude Fable and OpenAI's GPT-6 within two to three months — and reach "pole position" in AI in about six.
Image: Original illustration of a neural-network robot coding amid swirling coins, symbolizing Grok 4.7's price-war positioning. Credit: Red Eye Eagle News / original AI illustration.
Editor's Highlight: Grok 4.7 is less a capability leap than a strategy statement: "good enough and cheap" wins the agent economy — while long-context work got more expensive this week.
Table of Contents
- Key Facts
- What Grok 4.7 brings
- Benchmarks: leads in legal and engineering, trails the kings
- The $2/$6 sticker price — and the 200K-token catch
- Musk's six-month bet
- A new safety stack
- Bottom Line
- FAQ
- Sources
Key Facts
| Detail | Information |
|---|---|
| Model | Grok 4.7 (API id: grok-4.7) |
| Launched | September 21, 2026 |
| Base price | $2 input / $6 output per 1M tokens |
| Long-context tier | $4 input / $12 output over 200K prompt tokens |
| Context window | 500,000 tokens |
| Availability | xAI API, Cursor, Grok Build, OpenRouter, Vercel, Cloudflare |
| Musk's claim | Fable/GPT-6-level model in 2–3 months; "pole position" in ~6 |
What Grok 4.7 brings
Grok 4.7 uses a larger base model than Grok 4.6, with a longer reinforcement-learning run weighted toward multi-hour tasks, trained to understand the Grok Bot harness natively. The company says it checks its own work more carefully and handles long context better.
Specs: 500,000-token context window, text and image input with text output, four reasoning-effort levels (low, medium, high, xhigh), and tools for function calling, web search, X search and code execution.
Benchmarks: leads in legal and engineering, trails the kings
All launch benchmarks are vendor-reported by xAI, mixing effort settings across models — treat them as the company's own scorecard, not an independent ranking.
On that scorecard, Grok 4.7 leads in EEBench (electrical engineering, 64.0%) and the Harvey Legal Agent Benchmark (19.6%), ahead of every rival listed. CursorBench 4.0, focused on longer coding tasks, came in at 46.3%, up from 40.4% for Grok 4.6.
But Fable 5.1 Max tops four of seven benchmarks — Terminal-Bench 4.0 (57.9% vs. 38.0%) — and GPT-5.6 Sol Max leads DeepSWE; an independent rerun measured just 26% on Terminal-Bench against xAI's claimed 38%.
Analysis Note: The price comparisons favor xAI because its models are cheap, not because they are the smartest. A vendor's chart is the opening bid, not the verdict.
The $2/$6 sticker price — and the 200K-token catch
At face value Grok 4.7 undercuts the field: $2/$6 per million input/output tokens, against $4/$20 for GPT-5.6 Sol and $10/$50 for Fable 5.1 in xAI's own comparison.
But for prompts reaching 200,000 tokens, xAI bills the entire request at a long-context rate — roughly double:
| Prompt size | Input per 1M | Output per 1M |
|---|---|---|
| Under 200K tokens | $2.00 | $6.00 |
| 200K tokens or more | $4.00 | $12.00 |
This is a per-request cliff, not a marginal rate: a 199K-token prompt and a 200K-token prompt can have meaningfully different bills. Cached input moves from $0.50 to $1.00 in the same way, and Grok 4.6 had no such tier — so watch the 200K threshold before scaling.
A Grok 4.7 Fast variant — twice the output speed at twice the price — runs only in Cursor and Grok Build, and GitHub Copilot added Grok 4.7 across its paid tiers.
Musk's six-month bet
On X this week, Musk said he is "cautiously optimistic that SpaceX will have a Fable/GPT-6 level model in 2 to 3 months" — and could reach "pole position" in AI in about six months if its pace holds.
His argument: the AI effort is younger than its rivals', but SpaceX brings computing power online fast. "Hardware is hard," he said, "and SpaceX has shown it can do it rapidly."
He also sketched a roadmap — Grok 4.7 on par with Claude Opus 5.0 (not 5.1), a 2.5-trillion-parameter Grok 4.8 next, Grok 4.9 in the class of Astra and Fable, and Grok 5 "better than anything."
SpaceX's AI segment generated about $2.56 billion in second-quarter revenue, and the company acquired Cursor in August in a stock deal based on a $60 billion valuation.
Newsroom Focus: Treat the roadmap as a claim, not a commitment: Musk said in June that SpaceX would ship brand-new models monthly through 2026 — that cadence has not materialized, though Grok 4.5, 4.6 and 4.7 did ship.
A new safety stack
Grok 4.7 ships with an entirely new safeguard stack — the strongest xAI says it has tested on refusals and jailbreak resistance. It topped LatchBio's biosafety benchmark at 62.4%, and on the company's own HackerBench v0.3 it let 3.3% of risky dual-use prompts through. Both figures are company-reported, not independently verified.
Bottom Line
Grok 4.7 is the frontier market's cheapest credible coding model — a larger base model, longer-horizon training, and a $2/$6 sticker price that undercuts Anthropic and OpenAI several-fold, at least on paper.
What to watch: whether Musk's "pole position" timeline survives the benchmark gauntlet, whether the 200K-token cliff rankles developers, and how rivals answer a price war they are suddenly losing on the invoice.
FAQ
What is Grok 4.7 and when did it launch?
Grok 4.7 is xAI's flagship model for coding, agentic tasks and knowledge work, launched September 21, 2026, with a larger base model than Grok 4.6 and a longer reinforcement-learning run on multi-hour tasks.
How much does Grok 4.7 cost?
The base API price is $2 per million input tokens and $6 per million output tokens. The catch: prompts of 200,000 tokens or more are billed entirely at a doubled rate of $4 input and $12 output per million.
How does it benchmark against rivals?
xAI's scorecard puts it ahead in electrical engineering (EEBench 64.0%) and legal agent work (Harvey 19.6%), but behind Fable 5.1 on Terminal-Bench and GPT-5.6 Sol on DeepSWE — and an independent rerun scored it 12 points below xAI's Terminal-Bench claim.
What did Musk claim about SpaceX's AI future?
On X, Musk said SpaceX could field a Fable/GPT-6-level model in 2–3 months and reach "pole position" in about six, citing fast compute buildout — and that Grok 5 might be "better than anything."
Sources
- MarkTechPost, "SpaceXAI Releases Grok 4.7: A Larger Base Model at the Same $2/$6 Price as Grok 4.6" — launch specs, benchmarks, pricing, safety figures: https://www.marktechpost.com/2026/09/21/spacexai-releases-grok-4-7/
- Stocktwits, "SPCX Stock Rises Overnight: Musk Says SpaceX Could Match Anthropic's Fable, OpenAI's GPT-6 Soon" — Musk's X posts, roadmap claims, business context: https://stocktwits.com/news-articles/markets/equity/spcx-stock-musk-spacex-match-anthropic-fable-openai-gpt-6-lead-ai-6-months/cZMaDa7RBaO
- EveryDev.ai, "Weekly AI Dev News Digest: September 19–25, 2026" — pricing tier, Copilot rollout, price-war context: https://www.everydev.ai/p/news-weekly-ai-dev-news-digest-september-19-25-2026
- Simply AI, "xAI's Grok 4.7 Is the Cheapest Frontier Model Ever (With a Catch)" (YouTube) — explainer video embedded above: https://www.youtube.com/watch?v=cljcyFOFcP8
- Original AI illustration generated for Red Eye Eagle News (hero image).
Join the conversation
Comments
Post a Comment