Analysis · 2026-08-14 · 8 min read

Grok 4.6: $2/$6 per Million Tokens, 2M Context, and Frontier Agentic Coding

xAI's Grok 4.6 matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index at $2/M input and $6/M output. Full benchmark table, price-per-point math versus Fable 5 Max and GPT-5.6, and when the 2x-priced Fast variant is worth it.

TL;DR

Grok 4.6 launched August 12, 2026 at $2.00 per million input tokens and $6.00 per million output tokens, with a 2M-token context window.
It scores 61 on the Artificial Analysis Intelligence Index — a tie with GPT-5.6 Sol Max and one point behind Fable 5 Max (62).
That makes it, on published rates, the cheapest model at the 61-point tier: roughly a third of GPT-5.6's output price.
A Fast variant exists at exactly 2x the price ($4/$12) — same intelligence, lower latency.
Weak spot: DeepSWE v1.1 (65.9%) and Terminal-Bench v3.0 (26%) trail both GPT-5.6 and Fable 5 by a wide margin.

!xAI

1. The Pricing

Blended assumes a 3:1 input-to-output token ratio. Verify current rates in the live AI Model Pricing Table before you budget.

The number that matters here is output. Grok 4.6 bills output at 3x input — a tame ratio. GPT-5.6 bills output at 6x input, on a base that's already 3x higher. For a reasoning-heavy agent that emits long chains of thought, that compounding is the whole bill.

2. The Benchmarks, Honestly

xAI's published eval table, with the strongest competitor scores alongside:

Source: x.ai/news/grok-4-6. Third-party scores are best of self-reported or public results.

Read it as three groups:

Where Grok 4.6 leads: GDPVal-AA (real economic work product), AA-Briefcase (knowledge work), and Harvey LAB (legal reasoning, where it more than triples GPT-5.6's 2.5%). These are *knowledge-work* wins, not coding wins.

Where it ties: AA Intelligence Index, CursorBench, FrontierCode, APEX-Agents — all within a couple of points of the frontier. For most agentic coding in an IDE loop, you will not feel the gap.

Where it loses: DeepSWE (65.9% vs GPT-5.6's 73%) and especially Terminal-Bench v3.0 (26% vs ~34%). Long autonomous shell sessions remain the weak lane. If your agent lives in a terminal unsupervised, this is the number to care about, and it argues for escalation to a different model.

3. Price Per Intelligence Point

The interesting math isn't the benchmark — it's the benchmark divided by the blended price.

Grok 4.6 delivers a frontier-tier score at roughly a fifth of GPT-5.6's cost per point. That is the entire story of this release. xAI didn't win the intelligence race — it matched the leaders and then undercut them by 3–4x.

For a workload of 600M input and 30M output tokens a month:

Same traffic, $3,300/month of difference between the cheapest and most expensive frontier option. Run your own split in the cost calculator or side by side in the Model Comparison Calculator.

4. The 2M Context Window

Grok 4.6 ships a 2-million-token context window — 5x GPT-5.6's 400K and 4x Fable 5's 500K. Two practical consequences:

Whole-repo agents become viable. A 2M window holds a mid-size codebase plus dependency docs plus the conversation, without a retrieval layer in the middle.
Big contexts are still expensive. 2M tokens of input at $2/M is $4.00 per call before a single output token. A 40-turn agent run that re-sends a large context every turn will bankrupt you no matter how cheap the per-token rate is.

The fix is the same as always: stable prefixes and prompt caching, plus trimming the window instead of filling it. Our prompt caching guide covers the prefix rules; the Agent Loop Cost Estimator shows how re-sent context dominates multi-turn spend.

5. Should You Pay 2x for Fast?

The Fast variant is $4/$12 — identical capability, lower latency. It's worth it only when latency has a dollar value:

Yes: interactive coding assistants, live chat, anything where a human waits on the token stream.
No: batch pipelines, nightly agents, background enrichment, evaluation runs. There, latency is free and you're paying double for nothing.

The routing rule is simple: Fast for humans in the loop, standard for machines in the loop. Most teams that deploy both find 80%+ of their traffic belongs on the standard tier.

6. Where Grok 4.6 Fits in a Routing Stack

The Coinbase pattern — a gateway that routes by task class — applies cleanly:

Default lane: Grok 4.6 standard for agentic coding, research, drafting, and knowledge work. Frontier score, lowest price at the tier.
Latency lane: Grok 4.6 Fast for anything a human is watching.
Escalation lane: GPT-5.6 or Fable 5 for long autonomous terminal work and the hardest SWE tasks, where Grok's DeepSWE and Terminal-Bench gaps are real.
Cheap lane: an open model like Kimi K3 or a flash tier for classification, extraction, and summarization that never needed frontier reasoning.

Grok 4.6 is available in the SpaceXAI API, Grok Build, Cursor, and through OpenRouter, Vercel, and Cloudflare.

Bottom Line

Grok 4.6 is a price event more than a capability event. It matches GPT-5.6 on composite intelligence, beats both frontier rivals on knowledge-work and legal evals, and does it at $2/$6 per million with a 2M context. It genuinely trails on autonomous terminal work — so keep an escalation path — but for the majority of agentic and knowledge workloads, it resets what the frontier tier should cost.

Compare it against every other model on the LLM Leaderboard, or check its live rates in the pricing table.