What is xAI Grok 4.6?
Grok 4.6 is the model SpaceXAI released on 12 August 2026 and documents as its frontier model for coding, agentic tasks and knowledge work, built on Grok 4.5 with a particular focus on long-running agents and more ambitious interactive and visual work. Its own release notes describe a longer supplemental training run than the previous generation, model-generated data curated for reasoning and advanced technical concepts, high-quality engineering data, and an improved optimizer and training recipe, followed by SFT trajectories regenerated with Grok 4.5 across reasoning efforts, agent harnesses and domains such as STEM, software engineering and knowledge work, and then reinforcement learning over knowledge work, general coding and domain-specific environments for kernel optimization, web development and computer-aided design. Three later facts belong in this paragraph rather than in a footnote. It is no longer the newest model in its family, because Grok 4.7 shipped on 21 September 2026 on a new, larger base model and the current selection guidance in the official model table routes code and general work to that model. It is not on the deprecation list either, whose two entries are a voice-transcription model reaching end of life on 2 October 2026 and an image-quality model retiring on 2 November 2026. And the model page lists no alias for it at all, which is a difference from the previous generation rather than a gap to fill in. The identifier is grok-4.6, the documented context window is 500,000 tokens, inputs are text and image with text as the only output, function calling and structured outputs are supported, reasoning is enabled with no documented way to switch it off, and the four supported reasoning efforts are low, medium, high and extra-high, defaulting to high. Batch API is documented as not supported, the rate limits are 150 requests per second and 50,000,000 tokens per minute, and the regions named are us-east-1, us-west-2 and us-central-1. Where the previous generation had to be written about as a caveat, this one is the opposite: extra-high is available on Grok 4.6 and later, and it is the earlier model whose request for it is treated as high. Two labels point at this model and both are correct in their place — the developer’s pages write the company name SpaceXAI and the model name with no hyphen, while this site writes xAI Grok 4.6, which is what the navigation and the heading show.
What xAI Grok 4.6 is worth using for
Three lines on the model page that none of the seventeen competing pages copied
The model page for this generation prints a short list, and three items on it appear on none of the seventeen third-party pages this research could read. The first is the region list. The documented regions are us-east-1, us-west-2 and us-central-1, and this is the only one of the three recent generations whose page names us-central-1. The second is the rate limit: 150 requests per second and 50,000,000 tokens per minute, printed on the same page as the specification rather than in a support article. The third is Batch API, documented as not supported, which is a statement about absence and therefore easy to omit and equally easy to misread as an unlisted discount. Two further lines are worth reading with care. The model page states no maximum output for this model at all, and that silence is not a very large number: a later model’s entry in the release notes carries an explicit no-limit sentence and this model’s page does not, so the honest reading is that the model page does not say, while the release notes for this generation do. And the regions field describes where inference lands, which is a different thing from the optional US regional endpoint, a separate product that the documentation says currently serves only two of the family’s models.
A vendor model page lists three things that no third-party article about the model repeats: a third availability region the earlier generation lacked, a rate limit printed beside the specification, and a batch-processing feature marked as not supported. Write a section that presents all three as specifications rather than as selling points, then add the two readings that need care: the page states no maximum output, which is silence rather than a large number, and a region list is not the same thing as a separate regional endpoint product.
Extra-high reasoning is available from this generation on, and this box still has no level control
Four official statements describe reasoning for this model and they agree with each other, which makes this generation different from the one before it. The model page lists four supported reasoning efforts — low, medium, high and extra-high — with high as the default. The capabilities guide states that extra-high is available on Grok 4.6 and later, and that on models which do not support it, such as Grok 4.5, a request carrying extra-high is treated as high. The same guide states that reasoning cannot be disabled, and its per-model table lists the same four values for Grok 4.6. The release notes for this generation quote the same four with the same default. Read together, the fourth level is not a caveat here but a feature introduced with this generation, and the model that has to be written about with a downgrade warning is the earlier one. On this site the answer runs the other way again. The live catalog marks reasoning as unsupported for Grok 4.6, so the composer’s thinking control renders disabled rather than switched off, and none of the four documented values can be selected or sent. Two claims about two different things then sit side by side: the official pages describe the model’s capability, and this catalog entry describes this box’s controls. Only the second one governs what you can press.
Four official sources agree that a new reasoning level is available on a model and later, and that on an earlier model a request for it is silently treated as the next level down. A hosted chat interface for the same model exposes no reasoning control at all. Write two paragraphs: first the four official statements and what they jointly establish about which model introduced the level, then how a service can document a capability the model has while offering no way to request it.
This model and its successor carry identical input and output rates, and only the cache separates this one from the generation before
The pricing table on the model page prints six figures and one rule. At or below 200,000 prompt tokens the standard rate is $2.00 per million input tokens, $0.50 per million cached input tokens and $6.00 per million output tokens. From 200,000 prompt tokens upward those same three figures become $4.00, $1.00 and $12.00, and the official sentence states that a request whose prompt reaches that line is billed at the higher rate for all tokens in the request rather than only for the ones past it. Two comparisons follow from reading the tables together rather than the model page alone. The first is that this model and the next one are priced identically: the official model table prints the same $2.00 input and $6.00 output for both, so nothing in the price sheet records the difference the benchmark table measures. The second is counter-intuitive and worth stating plainly: cached input here is $0.50 per million, while the previous generation is documented at $0.30, so the older model is the cheaper one whenever a large stable prefix is resent. That is the clearest single illustration that a price table does not encode capability, and it is also the reason a workload that resends the same context should be costed rather than assumed. Six of the seventeen competing pages this research read mentioned the long-context rule at all, and one mentioned that this model shares its rates with its successor.
Two models in the same family are listed at identical input and output rates, and the older of the two has the cheaper cached-input rate. A third, newer model is documented at the same headline rates again. Explain what the price table does and does not tell a reader about which model to choose, describe how a request that resends a large stable prefix is billed differently from one that sends fresh text, and state what changes once that prefix sits in the cache.
What the official pages document for this model, and what this box actually opens
Two sets of statements describe this model and they answer different questions, so they are listed apart. On the official side, the model page documents a 500,000-token context window; text and image as inputs with text as output; function calling and structured outputs; reasoning enabled with no documented off switch and four effort values; Batch API not supported; 150 requests per second and 50,000,000 tokens per minute; and the regions us-east-1, us-west-2 and us-central-1. It states no maximum output for this model, and the release notes for the same generation do state one, so the two are kept as separate attributions rather than merged. On this site’s side, the page behind this chat box opens text chat with streaming replies, a web-search switch billed per search, and image and PDF attachments with a 24 MB per-file limit, against a reference price published in points. The gaps are the useful part. No context figure appears anywhere in the interface, because the catalog entry publishes no context length, so the official 500,000-token window describes the API rather than this page. No reasoning control exists, because the catalog marks reasoning unsupported for this model, so none of the four documented levels can be chosen here. And every reply is requested with at most 4,096 output tokens, because the catalog entry declares no maximum output and the composer caps whatever it reads at the same number. One more difference is a matter of naming rather than capability: the model page lists no alias for this identifier, so there is no shorthand to address it by, where the previous generation had two.
Write a product-page section that keeps two vocabularies apart: what a vendor’s API documentation states about a model, including a context window, input and output modalities, reasoning levels, a batch feature that is not supported and a rate limit, and what a hosted chat box actually opens, including a text field, an attachment control, a search switch and a per-reply output ceiling of its own. List every difference explicitly and name the two ceilings the host sets rather than inherits.
Official xAI Grok 4.6 specifications
- Developer
- SpaceXAI
- Model ID
- grok-4.6
- Official context window
- 500,000 tokens
- Official maximum output
- Not stated on the model page
- Official inputs
- Text and image
- Output
- Text
xAI Grok 4.6 and xAI Grok 4.7 side by side
Every value below comes from a single page — the Grok 4.7 announcement on x.ai — and from a single table within it. That table prints a Grok 4.7 xHigh column and a Grok 4.6 High column across two price lines and seven benchmarks, and the page’s own footnote names exactly those seven, so one official table accounts for both columns of every row and all nine are reproduced here in the order the official table prints them. Read the direction before the size of the gaps: the left column is this page’s model and the right column is its successor, and the successor scores higher on every benchmark row. Two of the nine rows are the opposite kind of fact and are the reason this section is worth reading at all: both models are listed at $2 per million input tokens and $6 per million output tokens, so the successor’s advantage is not visible in the price at all. A third column pair for GPT-5.6 Sol Max and Fable 5.1 Max exists on the same table and is not reproduced, and the GDPval figure that appears in a separate chart on the same page is deliberately left out, because mixing a chart reading into a table of table readings would blur where each number comes from. Two facts travel with the figures rather than being dropped. The footnote states that the DeepSWE v1.1 value marked with an asterisk for the successor is a high-effort score, so that one row compares different effort settings. And every figure is a vendor evaluation of the models, not a measurement of this website. No row predicts the quality or the speed of the conversation you are about to have.
| Official evaluation | xAI Grok 4.6 | xAI Grok 4.7 |
|---|---|---|
| Input token price ($ per 1M) | $2 | $2 |
| Output token price ($ per 1M) | $6 | $6 |
| Software engineering · CursorBench 4.0 | 40.4% | 46.3% |
| Software engineering · DeepSWE v1.1 | 65.2% | 71.0% * |
| Electrical engineering · EEBench | 53.0% | 64.0% |
| Multi-hour office work · AA Briefcase v1.1 | 1,546 | 1,657 |
| Multi-hour terminal work · Terminal-Bench 4.0 | 20.3% | 37.6% |
| Legal work · Harvey Legal Agent Benchmark | 15.8% | 19.6% |
| Clinical reasoning · HealthBench Professional | 48.5% | 56.7% |
How to use xAI Grok 4.6 online
Choose the model
This page opens with xAI Grok 4.6 selected. The model navigation on the left and the selector inside the chat box stay in sync, switching models keeps the current conversation, and starting a new chat is how you get a clean context. Two naming facts are worth fixing before you send anything. Two labels point at this model and both are correct in their place: the developer’s pages write the company name SpaceXAI and the model name with no hyphen, while this site’s display label, which is what the navigation and the heading show, reads xAI Grok 4.6. And the model page lists no alias for this identifier at all, which is a real difference from the previous generation, whose page lists two — so there is no shorthand form to fall back on here.
Decide what you expect from reasoning, because this box has no reasoning control
The official pages document four reasoning efforts for this model — low, medium, high and extra-high — with high as the default, and unlike the previous generation there is no downgrade warning attached to any of them: the capabilities guide states that extra-high is available on this model and later. The same guide states that reasoning cannot be disabled. What this page gives you is neither of those. The catalog entry marks reasoning as unsupported for this model, so the thinking control renders disabled and no effort value can be chosen or sent. There is therefore no lever here for reasoning at all, and the only way to shape how much thinking a task gets is the prompt itself. Note that this is not the same statement as the two official ones: the model reasons, the API accepts four values, and this box cannot ask for any of them.
Sign in, send, and check the cost against the official pages
Sending a message needs an account with available points. The composer reads a reference price from the live model catalog and a finished request reports what it actually cost. One ceiling is set by this site rather than by the model, so plan around it: replies are requested with at most 4,096 output tokens, and the catalog entry declares no maximum output, so the composer supplies the limit. Because the rates quoted on this page are the rates the official pages print, re-check them there before relying on a figure for planning — and treat any third-party summary that presents a relay, resale or promotional figure as out of scope, because no official page publishes one. Nothing here converts points into dollars, and nothing on the official pages puts an expiry on the rates.
What xAI Grok 4.6 costs, and why this page carries no end date
The first thing to establish about this model’s price is what is not attached to it. No official page puts an end date on any of its rates, and the one dated price sentence in the surrounding documentation belongs to a different member of the family. Both the model page and the announcement were checked on 4 October 2026. At or below 200,000 prompt tokens the standard rate is $2.00 per million input tokens, $0.50 per million cached input tokens and $6.00 per million output tokens. From 200,000 prompt tokens upward the same three figures become $4.00, $1.00 and $12.00, per million tokens each. The stated rule is that a request whose prompt reaches 200,000 tokens is billed at the higher rate for all tokens in the request, so crossing the line reprices the entire call rather than only the part above it. Two details follow from the way the tables are laid out. Cached input is priced at $0.50 per million tokens below the line, a quarter of the uncached input rate — and, read against the previous generation, it is the more expensive of the two, since that model is documented at $0.30. Batch processing is not a discount tier here at all: the model page records Batch API as not supported, and the platform pricing table applies its batch discount to four other models with a sentence stating that models not listed have no batch discount. The other charges on that page are for platform features rather than for this model: web search at $5 per thousand calls, X search at $5 per thousand posts fetched and $10 per thousand user profiles fetched, code execution at $5 per thousand calls, file attachments at $5 per thousand calls, collections search at $2.50 per thousand calls, file storage at $0.025 per GiB per day, collections storage at $0.10 per GiB per day, priority processing at twice the standard token rate, and the US regional endpoint at 1.1 times the token rate for the two models it serves, a list that does include this model — unlike the previous generation, which it does not. Note where the official documentation is silent: the pricing page prints its 1.1-times worked example for the newer model only, so no dollar figure is attached to this model’s US endpoint here. Where this model sits in its own generation is arithmetic rather than interpretation. The official model table puts Grok 4.6 and Grok 4.7 at identical input, cached-input and output rates, so the price sheet records no difference between them at all; the only price movement in the family belongs to the previous generation, whose cache is cheaper. None of these figures is this site’s checkout price. The workspace bills points, at 229 points per million input tokens, 687 points per million output tokens and 57.2 points per million cached read tokens, and there is no conversion between the two bases.
View points and plans ↗xAI Grok 4.6 FAQ
What does xAI Grok 4.6 cost, and will the price change?
The rates the official pages print are $2.00 per million input tokens, $0.50 per million cached input tokens and $6.00 per million output tokens at or below 200,000 prompt tokens, and $4.00, $1.00 and $12.00 above that line, where the higher rate applies to all tokens in the request. On whether it will change: this page attaches no end date, because the official sources give none — the model page prints rates and rules only, and the one dated price sentence in the surrounding documentation names a different model. Batch processing is not available for this model at all, so there is no batch discount to quote. One comparison is worth carrying away: this model and its successor are listed at identical input and output rates, while the previous generation has the cheaper cache at $0.30. None of these figures is this site’s checkout price: the workspace bills points, at 229 points per million input tokens, 687 points per million output tokens and 57.2 points per million cached read tokens, with no conversion between the two bases.
How does Grok 4.6 compare with Grok 4.7 on SpaceXAI’s own table?
Across the nine rows that one official table prints for both models — two price lines and seven benchmarks — the successor scores higher on every benchmark. The two price rows are identical at $2 per million input tokens and $6 per million output tokens, which is the clearest single reading in the table: the newer model is not more expensive at the standard tier. The largest gaps sit on Terminal-Bench 4.0 and EEBench, and the smallest on CursorBench 4.0. Two notes travel with the numbers. The footnote marks the successor’s DeepSWE v1.1 value with an asterisk indicating a high-effort score, so that row compares different effort settings. And a GDPval figure appears in a separate chart on the same page with the successor ahead; it is not reproduced here, because it is a chart reading rather than a row of the table. All of these are vendor evaluations of the models, not measurements of this website.
What do the official pages document for this model?
The model page records a 500,000-token context window, text and image inputs, text output, function calling, structured outputs, reasoning enabled with no documented way to disable it, four reasoning efforts (low, medium, high and extra-high) with a default of high, Batch API not supported, rate limits of 150 requests per second and 50,000,000 tokens per minute, and the regions us-east-1, us-west-2 and us-central-1. Three of those are easy to misread. The model page states no maximum output for this model, which is an absence rather than a large number — though the release notes for the same generation do carry an explicit no-limit sentence, so the two statements are attributed separately here rather than merged. The regions field describes where inference lands, not the optional US regional endpoint, which the documentation says currently serves two models of the family. And no knowledge cutoff is published for this model, so any cutoff figure quoted for it is not from the official pages. This page runs against none of those numbers directly: replies here are requested with at most 4,096 output tokens, and no context figure is displayed.
Is extra-high reasoning available on this model?
Yes, and this is where the generation boundary sits. Four official sources agree: the model page lists four supported reasoning efforts with extra-high among them and high as the default; the capabilities guide states that extra-high is available on Grok 4.6 and later and that on models which do not support it, such as Grok 4.5, a request carrying extra-high is treated as high; the same guide’s per-model table lists the same four values; and the release notes for this generation quote the same four with the same default. So the caveat that has to be documented for the earlier model does not apply here. What does apply is this site: the catalog marks reasoning as unsupported for this model, so the thinking control renders disabled and no level can be chosen or sent. The model can reason, the API accepts four values, and this box can ask for none of them.
Is Grok 4.6 still the newest model in its family?
No, and the official pages carry the dates either way. Grok 4.6 shipped on 12 August 2026; Grok 4.7 shipped on 21 September 2026 on a new, larger base model, and the current selection guidance in the official model table routes code and general work to it. Grok 4.5 dates to 16 July 2026. Two things point the other way and both are also official. Grok 4.6 remains listed and priced in the model table, and it is also named, alongside Grok 4.7, as one of the two models the US regional endpoint currently serves. And the deprecation list does not mention it anywhere — its two entries are a voice-transcription model reaching end of life on 2 October 2026 and an image-quality model retiring on 2 November 2026. The launch announcement’s own frontier claim is dated 12 August 2026, so read today it records a moment rather than a ranking.
Does the model page list any alias for Grok 4.6?
No, and this is a genuine difference from the previous generation rather than an omission from this page. The Grok 4.5 model page lists two aliases; the Grok 4.6 model page lists none. The general rule the documentation gives is that a bare model name resolves to the latest stable version, a name ending in -latest resolves to the latest version, and a name ending in a date resolves to that fixed version, but no alias is printed for this identifier specifically. So when a third-party page shows a shorthand form for Grok 4.6, that shorthand is not something the official model page publishes. Here on this site, no alias is exposed either — the model is addressed by the catalog identifier alone.
Is Grok 4.6 free, and are relay or resale prices official?
Neither, in the sense the question usually means. The launch announcement does contain a sentence offering doubled included usage inside one document product and one editor for a first week, but the sentence carries no date range anywhere on that page, and this site does not publish eligibility claims whose window cannot be checked — so no free-access or doubled-usage claim appears on this page. On relay and resale prices: no official page publishes one. The rates quoted in this article are the ones printed by the developer’s own model and pricing pages, and everything else — relay services, mirror sites, unregistered facsimiles — is outside the official record. Where this site is concerned, billing is in points against the catalog reference price, and there is no conversion between points and dollars.
What does this page expose compared with the official API?
Less than the official pages document, and the difference is listed rather than glossed over. Available here: text chat with streaming replies, image and PDF attachments with a 24 MB per-file limit, and a web-search switch billed per search. Not available here: any context figure in the interface, because the catalog entry publishes no context length; any reasoning control at all, because the catalog marks reasoning unsupported for this model, so none of the four documented effort values can be chosen; any alias, because the model page lists none for this identifier; function calling and structured outputs; the endpoint type; the three regions; the rate limits, which are not exposed as any kind of quota display; and Batch processing, which the official model page does not support in any case. One ceiling is set by this site rather than by the model: replies are requested with at most 4,096 output tokens. Pricing is in points and is not a conversion of the official dollar rates.
Sources and model information
Information checked
Start an xAI Grok 4.6 conversation
The chat box above runs this model, and the guide on this page is written to be checked rather than trusted: the specifications come from the official model page, the comparison comes from the single official table that prints both columns, and the places where official pages disagree with each other or with this workspace are named rather than smoothed over. Send the task the way you would brief a person — outcome, constraints, evidence, stopping condition — and re-check the figures against the pages linked below, since this model’s rates carry no end date and its successor has already been listed at the same headline input and output rates.
Start a chat with xAI Grok 4.6 ↑