Is Grok 4.7 better than Grok 4.6?

Yes, on every benchmark xAI published on September 21, 2026: CursorBench 4.0 rose from 40.4% to 46.3%, DeepSWE v1.1 from 65.2% to 71.0%, EEBench from 53.0% to 64.0%, AA Briefcase v1.1 from 1,546 to 1,657, Terminal-Bench 4.0 from 20.3% to 38.0%, the Harvey Legal Agent Benchmark from 15.8% to 19.6% and HealthBench Professional from 48.5% to 56.7%. Artificial Analysis independently measured its hallucination rate on AA-Omniscience falling from 34% to 29% and its Intelligence Index score at 46 versus Grok 4.6's lower mid-pack score. Grok 4.7 also uses a new, larger base model rather than Grok 4.6's post-training-only refresh.

Does Grok 4.7 cost more than Grok 4.6?

Per token, no. Both are $2.00 per million input tokens, $0.50 cached and $6.00 per million output tokens under 200,000 prompt tokens, doubling above that, with a fast variant at twice the speed for twice the price. Per task, yes: Artificial Analysis found Grok 4.7 (xhigh) emitted about 81,000 output tokens per Intelligence Index task against about 36,000 for Grok 4.6 (xhigh), because it 'works longer on difficult tasks.' Expect a hard agentic task to cost roughly twice what it did on 4.6 at the same effort setting.

Should I upgrade from Grok 4.6 to Grok 4.7?

Upgrade if you run long agentic coding or terminal tasks (Terminal-Bench 4.0 nearly doubled), legal or electrical-engineering workflows, or anything where Grok 4.6's hallucinations hurt. Stay on Grok 4.6 for high-volume, short, latency-sensitive calls — classification, extraction, chat — where 4.6 already scored well and 4.7's extra thinking tokens would double your bill for no visible gain. Cursor kept Grok 4.6 on its menu for exactly this reason. A sensible pattern is to route long tasks to 4.7 and keep 4.6 as the cheap default.

Is Grok 4.6 being deprecated?

As of September 22, 2026 xAI has announced no deprecation date for Grok 4.6. It remains available in the xAI API at the same price as Grok 4.7, Cursor explicitly kept it selectable, and xAI's launch post positions 4.7 as an addition 'served at the same price and speed as Grok 4.6.' Given that Grok 4.5 shipped July 16 and Grok 4.6 in August, xAI's roughly six-week cadence suggests 4.6 will be superseded in tooling menus within a few months, so plan migrations rather than rushing them.

Quick Answer

Grok 4.7 vs Grok 4.6: Should You Upgrade? (Sep 2026)

Published: September 22, 2026

The short answer

Upgrade for long agentic tasks; keep Grok 4.6 for short, high-volume calls. Grok 4.7, released September 21, 2026, beats Grok 4.6 on all seven benchmarks in xAI’s launch table and hallucinates less, at an identical $2/$6 per million tokens. The catch is that it reaches those scores by thinking more than twice as long, so per-task cost roughly doubles. Same rate card, different bill.

Side by side

	Grok 4.7	Grok 4.6
Released	September 21, 2026	August 2026
Base model	New, larger base model	Same 1.5T V9 foundation as Grok 4.5, improved post-training
Price (<200K prompt)	$2.00 in / $0.50 cached / $6.00 out per MTok	Same
Price (≥200K prompt)	$4.00 / $1.00 / $12.00	Same
Fast variant	2x output speed at 2x price	2x output speed at 2x price
Context window	500,000 tokens	500,000 tokens
Knowledge cutoff	May 2026	Earlier 2026
Output tokens per AA index task (xhigh)	~81,000	~36,000
AA-Omniscience hallucination rate	29%	34%
AA Intelligence Index v4.3.2	46 (rank 16/655)	Lower; mid-pack
Availability	Cursor, Grok Build, xAI API, GitHub Copilot, Vercel AI Gateway	Same surfaces; still selectable in Cursor

Benchmarks: every row moved

Benchmark	Grok 4.7 (xHigh)	Grok 4.6 (High)	Gain
CursorBench 4.0	46.3%	40.4%	+5.9 pts
DeepSWE v1.1	71.0% (high effort)	65.2%	+5.8 pts
EEBench	64.0%	53.0%	+11.0 pts
AA Briefcase v1.1	1,657	1,546	+111
Terminal-Bench 4.0	38.0%	20.3%	+17.7 pts
Harvey Legal Agent Benchmark	19.6%	15.8%	+3.8 pts
HealthBench Professional	56.7%	48.5%	+8.2 pts
GDPval (Elo, xAI chart)	1,695	1,605	+90

The Terminal-Bench 4.0 jump is the headline: multi-hour terminal work nearly doubled. That is where “a longer reinforcement-learning run on a harder mix of tasks, weighted toward problems that take many hours” shows up. Caveat: Artificial Analysis’s own Terminal-Bench 4.0 run put Grok 4.7 at 26%, so the absolute level depends heavily on harness and effort setting — but the relative gain over 4.6 is consistent across sources.

What actually changed under the hood

Grok 4.6 was a post-training refresh: xAI reused Grok 4.5’s foundation and improved supervised fine-tuning and RL, targeting 4.5’s hallucination problem. Grok 4.7 is a different animal — “a new, larger base model,” a longer RL run, explicit training to verify its own work and manage longer context, and native understanding of the Grok Bot harness. It is the first genuinely new base since Grok 4.5 shipped on July 16, 2026.

Musk’s July claims of 2.1 trillion parameters and SpaceX engineering data in training are not in xAI’s launch post; the parameter count remains unpublished.

The token-cost trap

Same per-token price does not mean same cost. Artificial Analysis measured ~81,000 output tokens per Intelligence Index task for Grok 4.7 versus ~36,000 for Grok 4.6 at matched xhigh effort, at ~188 tokens per second. Worked example at $6 per million output tokens:

1,000 hard agentic tasks on Grok 4.6: 36M output tokens ≈ $216
1,000 hard agentic tasks on Grok 4.7: 81M output tokens ≈ $486

You pay ~2.25x for the quality gain. Against Claude Fable 5.1 at $50 per million output, that is still cheap, but if you set a budget on 4.6 you will blow it on 4.7 unless you lower the effort setting.

Upgrade decision by workload

Workload	Recommendation	Why
Long agentic coding in Cursor / Grok Build	Upgrade	Terminal-Bench 4.0 20.3% → 38.0%; CursorBench +5.9 pts
Multi-hour office / document tasks	Upgrade	AA Briefcase +111, GDPval +90 Elo; xAI says documents and presentations improved
Legal, electrical engineering	Upgrade	Grok 4.7 leads even Fable 5.1 Max on Harvey Legal and EEBench
Anything hurt by hallucinations	Upgrade	34% → 29% on AA-Omniscience
Short chat, classification, extraction at volume	Stay on 4.6	Extra thinking tokens double cost for little gain
Latency-critical UX	Stay on 4.6 or use 4.7 fast variant	4.7 spends ~7 minutes per hard index task at xhigh
Clinical reasoning	Neither — use Fable 5.1 / GPT-5.6 Sol	Grok 4.7’s 56.7% HealthBench trails both

Migration checklist

Swap the model ID (grok-4.7) in one non-critical route; the API surface is unchanged.
Cap effort — start at high, not xhigh; Artificial Analysis found the top two reasoning levels score about the same.
Re-baseline your token budget for ~2x output on hard tasks.
Keep 4.6 as fallback in your router until 4.7 has run a week of production traffic.
Check the ≥200K prompt premium — long-context prompts double in price on both models, so trim context rather than assuming the 500K window is free.

For how Grok 4.7 stacks up against the frontier pair, see Grok 4.7 vs Claude Fable 5.1 vs GPT-6 Astra vs GPT-5.6 Sol; for the model itself, What is Grok 4.7?.