Grok 4.6 costs, limits, and evals vs models I pay for

When people mix up Grok Build, Grok 4.6, and grok-build-0.1, I do not blame them. xAI shipped a product, a frontier model, and a cheaper coding model in the same week, and the names overlap on purpose.

This is not a how-to for chat. I already wrote that: How I Actually Use Grok Bot. This post is the bill and the scoreboard. I copied the eval numbers from xAI’s 12 August 2026 launch post. I copied the prices from the live model cards on 30 August 2026. I did not round them.

I pay for Grok 4.6 and I use Grok Build. “The models I actually pay for” in the title means that usage, compared honestly to xAI’s own table — not a private invoice dump.

Three different things with “Grok” and “Build” in the name

Grok Build is the coding-agent product. The homepage is x.ai/build. The install line they publish is:

text

curl -fsSL https://x.ai/cli/install.sh | bash

As of the 4.6 launch, Grok Build is powered by Grok 4.6. It is a CLI / computer agent: plan mode, skills, MCP, subagents, diffs. It is not the chat window on grok.com.

Grok 4.6 is the frontier model (grok-4.6). 500k context. Text and image in, text out. That is what Grok Build and Cursor were pointed at on launch day. Pricing starts at $2 / $6 per 1M tokens under 200k prompt, and doubles once the prompt is ≥ 200k. Details below; I re-fetched them at write time.

**grok-build-0.1** is a different API model. 256k context. Cheaper and faster. Aliases on the card include grok-code-fast-1. Do not conflate the Grok Build product with this model id. 4.6 is the default in Grok Build / Cursor. grok-build-0.1 is the cheaper coding model you call from the API when you want that tradeoff.

I keep those three names separate in every internal note. Mixing them is how you mis-read a bill.

What Grok Build is, in practice

I treat Grok Build as the agent that sits on the machine: repo, terminal, skills, MCP servers. Chat Grok is the hosted assistant I use for thinking and drafting. Different jobs.

The product page is the source for what it claims to do: skills and AGENTS.md, plan mode with blocked edits until you approve, plugins, Q&A when a task is ambiguous, parallel subagents, including worktree-isolated ones. I do not treat the marketing list as a benchmark. I treat it as the product surface.

xAI said Grok 4.6 is available in Cursor and Grok Build, and that they offered 2× included usage inside those two products for the first week after launch. That week is over. I am not freezing a promo as a current limit.

If you want the chat workflow, use the live how-to. This page is costs, limits, and the official eval table.

Official evals (12 August 2026)

I am copying xAI’s launch table. Competitor figures in that table are from their comparison — xAI says they are drawn from the respective developers’ system cards or leaderboards. I am not inventing scores and I am not “adjusting” them.

Source: Introducing Grok 4.6, dated 12 August 2026. Columns are Grok 4.6 High, Grok 4.5 High, GPT-5.6 Sol Max, Fable 5 Max.

AA Intelligence Index — 61 / 56 / 61 / 62
GDPVal-AA v2 — 1753 / 1526 / 1728 / 1741
CursorBench v3.2 — 69.9% / 66.7% / 67.2% / 70.5%
DeepSWE v1.1 — 65.9% / 54% / 73% / 70%
FrontierCode v1.1 (Extended) — 61.3% / 56.6% / 60.6% / 63.6%
APEX-Agents — 57.5% / 47.1% / 56.7% / 59.2%
Terminal-Bench v3.0 — 26% / 15.7% / 34.6% / 34.1%
APEX-SWE — 56.4% / 53.6% / — / 58.8%
AA-Briefcase — 1577 / 1313 / 1502 / 1574
Harvey LAB (Vals) — 15.8% / 12.9% / 2.5% / 11.3%

xAI’s note on that table: best score per evaluation in bold; third-party model scores are the best of self-reported or publicly available results.

What I actually take from that as someone who pays the bill:

On the composite AA Intelligence Index, they show 4.6 High tied with GPT-5.6 Sol Max at 61, with Fable 5 Max at 62.
4.6 is a clear step up from 4.5 High on every row they published.
It is not the top score on DeepSWE, Terminal-Bench, CursorBench, or FrontierCode in their own table. I am not going to pretend otherwise.
Harvey LAB is the row where 4.6 High is the best of the four numbers they showed.

That is the official launch table. If a later card disagrees, trust the later card.

Costs I re-fetched on 30 August 2026

Headline, as locked: $2 / $6 per 1M under 200k, doubles above, 500k context.

Live sources:

grok-4.6

From the model card, same day:

Context: 500,000 tokens
Modalities: text, image → text
Function calling, structured outputs, reasoning: yes
Batch API: not supported

Per 1M tokens:

Prompt under 200k: input $2.00, cached input $0.50, output $6.00
Prompt ≥ 200k: input $4.00, cached input $1.00, output $12.00

xAI’s rule, quoted in substance: once the prompt reaches 200k tokens, the whole request is billed at the higher band.

Priority / fast is 2× standard rates when the response confirms "service_tier": "priority". If it was served at the default tier, you pay standard. I am not going to freeze a “fast” SKU name beyond that.

grok-build-0.1

Context: 256,000 tokens
Per 1M, prompt under 200k: $1.00 / $0.20 cached / $2.00 output
Prompt ≥ 200k: $2.00 / $0.40 / $4.00

That is the cheaper/faster coding model. Still not the Grok Build product.

Tools

From the pricing page: web search, X search, and code execution / code interpreter are $5 / 1k calls, on top of tokens. File attachments search is $10 / 1k. Collections search is $2.50 / 1k. Image generation uses Imagine rates, not this table.

Rate limits (Tier 0, the default)

Tiers are spend-based on the API since 1 January 2026. Tier 0 is $0 cumulative. Once you qualify for a tier, it does not downgrade.

grok-4.6 T0: 150 RPS, 50M TPM
grok-build-0.1 T0: 37 RPS, 10M TPM

Higher tiers scale those caps. I am not going to paste the whole ladder here; the rate-limits doc is the live table.

I did not find a knowledge-cutoff date on the Grok 4.6 model card I fetched. I am not inventing one.

How I actually read this as a buyer

I do not use evals to pick a religion. I use them to sanity-check marketing.

Grok 4.6 is the model I pay frontier rates for when the job is long-running agent work, coding across a repo, or knowledge work that needs the 500k window. grok-build-0.1 is what I reach for when the job is “cheap, fast, coding-shaped” and 256k is enough. Grok Build the product is the harness. Chat Grok is still the thing I described in the how-to.

I also do not freeze consumer plan prices in this post. Those move. If you are on grok.com rather than the API, look at the live plans page.

Competitor list prices belong on that vendor’s own pricing page. If I cannot cite the vendor, I skip the number. This post is xAI’s card plus xAI’s eval table.

How I route work across the three names

This is the part that is mine, not xAI’s table.

Chat Grok (the how-to) is for thinking, outlines, and research preparation. I do not point it at a repo and walk away.

Grok Build is for work that should happen on a computer: plan, edit, test, MCP, subagents. The model inside it, as of the 4.6 launch, is Grok 4.6. I still approve plans. I still read diffs. The product is a harness, not a replacement for review.

**API grok-4.6** is what I pay when I need that model in my own loop — long context, agentic coding, knowledge work — and I am okay with frontier token rates. I watch the 200k band. A single fat prompt is how you accidentally double the whole request.

**API grok-build-0.1** is what I pay when the job is coding-shaped, 256k is enough, and I want the cheaper band. I do not pretend it is Grok 4.6 because the product and the model id share a word.

A concrete bill shape, using only the published rates (not a usage dump):

2M fresh input + 1M output on grok-4.6 under 200k: (2 × $2) + (1 × $6) = $10
The same tokens on grok-build-0.1: (2 × $1) + (1 × $2) = $4
The same tokens on grok-4.6 if the prompt had already crossed 200k: (2 × $4) + (1 × $12) = $20

Plus $5 / 1k if the agent actually called web search, X search, or code execution. Token math without tool calls is how people under-forecast agents.

I do not use that arithmetic to declare a winner against Claude or GPT. I use it so I know which SKU I am in before I start a long run.

What the eval table is not

It is not a reason to rewrite the how-to. It is not CursorBench as a purchasing department. It is xAI’s launch comparison, with their footnote about competitor figures.

If you want the chat workflow, stay on the live how-to. This URL stays the card.

What I will not do with these numbers

I will not round 69.9% to 70%. I will not silently drop the Harvey LAB row because it is awkward for a competitor. I will not write “Grok 4.6 wins coding” when their own table has GPT-5.6 Sol Max and Fable 5 Max ahead on DeepSWE and Terminal-Bench.

I will not merge this into the live how-to. That URL already has sessions. This slug is a separate row.

Sources

Introducing Grok 4.6 — https://x.ai/news/grok-4-6
Grok Build — https://x.ai/build
Grok 4.6 model card — https://docs.x.ai/developers/models/grok-4.6
Grok Build 0.1 model card — https://docs.x.ai/developers/models/grok-build-0.1
Pricing — https://docs.x.ai/developers/pricing
Rate limits — https://docs.x.ai/developers/rate-limits
Live sibling: How I Actually Use Grok Bot

Grok 4.6 costs, limits, and evals vs the models I actually pay for