Grok API Pricing: One Model, Seven Different Rates

On 18 August 2026, Amazon Bedrock started serving xAI’s Grok 4.6. That matters less for what it added than for what it exposed: the same model now has more than one price, and the gap between the cheapest way to buy it and the most expensive is a factor of four. Nothing about the model changes. The invoice does.

This post prices Grok 4.6 the way a WordPress site owner actually meets it — as the engine behind a support chatbot answering questions about your own content. Every rate below was read from a vendor’s own page on 3 September 2026, and every token count comes from this site’s stored transcripts rather than from a configuration screen. That distinction cost us a published figure last week, and the correction is still sitting on that post.

The rate cards, side by side

xAI sells Grok 4.6 directly. AWS resells it on Bedrock. They do not charge the same thing, and Bedrock does not charge one thing either — the price moves with how you route the request and which service tier you ask for.

Two things stand out before any arithmetic. AWS resells the model at xAI’s own list price if you accept Global routing — there is no reseller margin visible in that row. Ask for your requests to stay inside a region or a geography and you pay 10% more for the identical model. That is the price of data residency, stated plainly, and it is not a number either vendor puts in a headline.

What one reply actually costs on a WordPress site

A chatbot reply is not an abstract token count. On this site it is a system prompt, a block of retrieved documentation, and a visitor’s question — and the last of those is the smallest part by a wide margin. Measured from this install’s own transcripts on 2 September:

The samples behind those means are small — six retrieval contexts, ten replies, fourteen visitor messages — and they are stated rather than smoothed. Tokens are estimated at four characters each. What the profile establishes is a shape rather than a precise figure, and the shape is decisive: 95.7% of the tokens in a chatbot exchange are input. You are paying almost entirely to send the model your own documentation, over and over.

Run that profile through the seven prices and the spread becomes concrete.

Two of those prices are not for sale

The cheapest row on the chart is one a support chatbot can never use, and the most expensive is one it can never trigger. Both deserve to be understood rather than skipped.

Flex is asynchronous. It is Bedrock’s discount lane for work that can wait, and a visitor staring at a typing indicator cannot. This is the same structural finding we reached about OpenAI’s batch API yesterday, arrived at from a different vendor: every provider now sells tokens at roughly half price on a lane that cannot serve a request someone is waiting on. The discount is real and the chatbot is simply not eligible for it. Where it does land is offline work — embedding your knowledge base, generating content — which is a much smaller share of most sites’ bills than the headline discount suggests.

The long-context rate needs a prompt this workload will never assemble. xAI’s card is explicit about the mechanism, and it is harsher than a tiered rate: once a request’s prompt reaches 200,000 tokens, the higher rate applies to every token in that request, not only the ones above the line. Cross the threshold by one token and the whole request doubles in price.

A retrieval chatbot sits at 1.9% of that threshold. The cliff is the single largest number on the rate card and it is irrelevant to this use case — which is worth saying out loud, because it is very relevant to the agent and coding workloads Grok 4.6 was actually built for. If you are pointing Grok at a 400-page codebase, that row is your real price. If you are answering questions about a plugin’s documentation, it is noise.

One honest gap: xAI’s card publishes the long-context rule, and the Bedrock model card for xai.grok-4.6 publishes a single Standard rate with no threshold mentioned. Whether AWS applies the same doubling is not something either page states. If your prompts run long, confirm it with AWS rather than assuming either behaviour.

Caching helps less than the rate card implies

The cached-input rate is the most attractive number on xAI’s card: $0.50 against $2.00, a 75% discount. It applies to a stable prefix the provider has seen before — and on a retrieval chatbot, most of the prompt is not stable.

The system prompt is identical on every single reply: 870 tokens, 23.1% of the input. The retrieval context is 76% of it and changes with the question, because different questions pull different chunks out of the knowledge base. Only an exact repeat question would present the same prefix twice.

So the reliable saving is the system prompt alone: 870 tokens moving from $2.00 to $0.50 per million is $1.30 per 1,000 replies, or 15.3% of the bill. Worth having. Not the 75% the rate card advertises, and any calculator that applies the cached rate to your whole prompt is quietly overstating the saving by a factor of four. Note also that this assumes the cache is actually hit — prefix caching depends on how your plugin assembles the request and how long the provider retains the prefix, neither of which is guaranteed by the rate card.

On a chatbot, the output price barely matters

This is where the Grok ladder gets interesting, and where most pricing comparisons mislead. Model comparisons lead with output rates, because that is where vendors differentiate most dramatically — Grok 4.6 charges $6.00 per million against Grok 4.3’s $2.50, a 2.4× difference. On this workload it is nearly invisible.

Output tokens are 8–12% of the bill across the entire ladder. Dropping from Grok 4.6 to Grok 4.3 saves 40%, and essentially all of that saving comes from the input rate falling from $2.00 to $1.25. The 2.4× output difference — the number a spec comparison would put in bold — contributes about four percentage points.

The practical rule: on a retrieval chatbot, pick your model on the input rate. Five models on that card collapse into three distinct prices for this workload, and the thing separating them is what it costs to send your own documentation, not what it costs to get an answer back.

One detail worth flagging for anyone comparing the two newest models: Grok 4.6 and Grok 4.5 are identically priced on input and output, but their cache rates differ — $0.50 against $0.30. The newer model has the worse cache discount. It is a small effect at 23% cacheable prefix, but it runs against the direction you would assume.

That $2.00 / $6.00 rate card is not unique to xAI. Alibaba’s Qwen3.8 Max publishes the identical numbers and therefore prices this exact workload at the identical $8.54 per thousand replies — though what that means for your bill depends entirely on which price you are moving from.

How that compares to what this site actually runs

mxchat.ai does not run Grok. It runs GPT-5.6 Luna, at $0.20 input and $1.20 output per million, which prices the same measured reply at $0.956 per 1,000. Grok 4.6 on the same traffic is 8.9× more expensive.

That is not an argument that Grok is overpriced. It is an argument about fit. Grok 4.6 is built for long-running agents, coding, and 500K-token context — capability a support bot answering “how do I change my plan” does not draw on. You are buying agent-grade reasoning and then using it to look up a documentation page. When a chatbot’s cost is 96% the price of shipping your own text to a model, the cheapest model that reads that text competently wins, and the frontier tier is spending money on headroom the workload never reaches.

Where Grok earns its rate on a WordPress site is the narrower case: a bot that has to reason over a large body of content in a single pass, drive multi-step tools, or work through a long conversation without losing the thread. If that is your bot, the 500K context and the reasoning effort settings are the reason to pay, and the routing choice below becomes the thing to get right.

What the pricing pages do not tell you

We ran the live Google results for grok api pricing while writing this. Most of page one is accurate as far as it goes — xAI’s own docs hold the top slots, and several third-party trackers state Grok 4.6 at $2.00 and $6.00 correctly. The commoner problem is not wrong figures but stale subjects: several well-ranked pages are still describing Grok 4.3 or Grok 4 as the current flagship, at rates that were correct for those models and are not the rate you will be billed for 4.6.

The consistent gap is different, and it is the reason this post exists: none of them prices the routing decision. The Bedrock In-Region premium is a live 10% on every token, it applies to the exact same model, and it is invisible unless you read the AWS model card next to the xAI card. If you are choosing Bedrock for compliance reasons, that is a cost you should be choosing deliberately rather than discovering on an invoice.

Common questions

Is Grok cheaper on Bedrock or direct from xAI? Identical, if you accept Global cross-Region routing — both are $2.00 input and $6.00 output per million. Bedrock is 10% more expensive if you need In-Region or Geo routing, and 1.75× the standard rate if you select the Priority tier.

What does Grok 4.6 cost per chat message? On a retrieval chatbot with a profile like this site’s, about $0.0085 per reply, or $8.54 per 1,000. A bot handling 100 conversations a day at three replies each would run roughly $2.56 a day.

Does the 200K long-context rate affect a WordPress chatbot? No. A typical retrieval prompt is under 4,000 tokens, about 1.9% of the threshold. It matters for agent and code workloads, where the doubling applies to every token in the request once you cross the line.

Which Grok model should a support bot use? The cheapest one on input that answers competently — grok-build-0.1 at $4.10 per 1,000 replies or grok-4.3 at $5.12, rather than 4.6 at $8.54. Output quality differences show up in the bill at only 8–12% weight on this workload.

Can I use the Flex discount for my chatbot? No. Flex is asynchronous, so it cannot serve a reply a visitor is waiting for. It is useful for offline jobs like embedding a knowledge base or bulk content generation.

The short version

One model, seven prices, and the two that bracket the range are both unavailable to a live chatbot. Among the prices you can actually pay, the spread is still 1.9× — from $8.54 direct or on Global routing to $16.44 on Bedrock’s Priority tier — and the single decision that moves the number most is not which Grok you pick but how much of your own documentation you send it on every request.

If you are choosing a model for a WordPress chatbot, price the input rate against your measured prompt size before you compare anything else. DeepSeek’s off-peak window, Gemini Flash’s expiring introductory rate and Anthropic’s cancelled increase all move the same lever. MxChat supports xAI alongside OpenAI, Anthropic, Google and DeepSeek, so switching providers is a settings change rather than a migration — the documentation covers the model configuration, and the plugin licences are listed on the shop page. If you are still choosing a plugin, our comparison of WordPress chatbot plugins is the place to start.