Grok 4.6: Benchmarks, Pricing, Context & API

SpaceXAI · Proprietary

A high-agentic-score frontier model priced like a challenger, with a smaller context window and thinner provider ecosystem than its main rivals. Source-checked August 28, 2026.

Compare this model Open live ranking

Live rank: #4
Intelligence v4.1: 60.9
Best blended / 1M: $3.00
Context: 500K

The answer in 60 seconds

The value challenger in the frontier group

Grok 4.6 earns a serious trial because its live agentic and general scores sit near the leaders while its published token price is materially lower. That makes it compelling for agent loops and high-volume reasoning. The trade-offs are a 500K context ceiling, modest observed throughput, and an ecosystem that is less portable than OpenAI or Anthropic across enterprise clouds.

Best for

— Agentic workloads where cost per completed task matters
— Reasoning and coding systems already using xAI tools
— Teams willing to benchmark a lower-priced frontier alternative

Look elsewhere when

— You require a one-million-token input window
— Cloud-provider portability or mature enterprise integrations are mandatory

Independent data

What the current measurements say

Metrics come from the live Artificial Analysis-backed WhatLLM feed and refresh hourly. Prices and speed are best observed values across matching endpoints; they can come from different providers.

Intelligence Index: 60.9
Coding Index: 76.8
Agentic Index: 58.7
Input / output per 1M: $2.00 / $6.00
Fastest observed: 59.9 tok/s
Lowest observed TTFT: 4.150s

Benchmarks worth reading together

The strongest indexed configuration is shown. Different tests measure different jobs; no single score is a universal model verdict.

GDPval-AA Elo: 1736
GDPval-AA: 62%
GPQA Diamond: 95%
Humanity's Last Exam: 43%
SciCode: 54%
AA-LCR: 75%

One family, 4 measured configurations

Effort settings belong on one canonical page. Compare the quality and latency curve here instead of treating each setting as a separate model.

Configuration	Intelligence	Coding	Agentic	Fastest	Lowest TTFT
high	60.9	76.8	58.7	59.7 tok/s	33.901s
xhigh	60.0	75.9	56.6	59.9 tok/s	34.251s
medium	59.0	74.4	56.3	58.1 tok/s	27.156s
low	51.7	66.3	47.8	55.6 tok/s	4.150s

Source truth

What Grok 4.6 actually is

API model: grok-4-6; Provider-confirmedSource ↓
Context: 500,000 tokens; Provider-confirmedSource ↓
Reasoning: Low through xhigh effort configurations; Provider-confirmedSource ↓
Official token price: $2 input / $6 output per 1M tokens; higher beyond 200K context; Provider-confirmedSource ↓
Safety evidence: Dedicated public Grok 4.6 model card; Provider-confirmedSource ↓

Editorial analysis

The decision behind the ranking

We have not claimed hands-on use here. The analysis below combines provider-confirmed facts with the independent data shown above; each section labels that distinction.

WhatLLM analysis

The interesting number is agentic value, not the tie on overall score

A leaderboard tie can hide the commercial difference. Grok 4.6 combines a frontier-band Agentic Index with a lower blended token price than the premium Anthropic and OpenAI tiers. If an agent makes many calls, retries, and tool decisions, that price gap compounds faster than a small movement in the overall index.

The right test is a completed-task harness with tool costs and retries included. A cheap token that causes another loop is not cheap; a lower-priced model that finishes in fewer steps is.

Open question

Where the challenger story still needs proof

Grok’s official release and model card provide useful first-party evidence, but independent operational evidence is still younger than it is for the longest-running API families. Reliability under sustained concurrency, tool-specific error recovery, and provider portability deserve workload tests before a broad migration.

Boundaries

What the headline score misses

01The 500K context window is half the size of several current frontier peers, and requests beyond 200K use higher-context pricing.
02Current independent speed observations are not class-leading.
03The model card bars autonomous high-stakes decisions without human oversight and domain-expert validation.
04Jailbreak results improved on one harness and worsened on another; “safer” is not a supported blanket conclusion.
05The model is closed and third-party hosting choices are limited.

Alternatives

The next models to test

Compare the shortlist →

GPT-5.6 Sol

The broad API flagship: frontier coding and tool use with more controllable effort than its single leaderboard row suggests.

GLM-5.3

A low-priced, agentic coding specialist whose strongest release is hosted rather than open weight.

Claude Opus 5

The practical premium leader: close to the capability ceiling without Fable-level pricing or routing complexity.

Evidence register

Sources and claim status

Checked August 28, 2026. Pricing and availability should be verified again before procurement.

Direct answers

Questions people ask about Grok 4.6

Is Grok 4.6 a frontier model?

On the current WhatLLM and Artificial Analysis signals, yes: its strongest effort configurations sit in the same broad capability band as the leading closed models. Production fit still depends on your tools, latency, and error costs.

What is the Grok 4.6 context window?

SpaceXAI documents a 500,000-token context window.

Why choose Grok 4.6 over GPT-5.6 Sol?

Grok can be attractive when token economics and agentic performance matter more than a one-million-token window or OpenAI’s wider tool and cloud ecosystem. Test both on completed tasks rather than comparing one benchmark.

Assembling the frontier