Grok 4.5: Features, pricing and what changed

Grok 4.5 was the model behind SpaceXAI’s July 2026 rollout. It reached the xAI API on 8 July 2026, was formally announced on 16 July, opened to EU API users on 17 July, and arrived across Grok.com, X, iOS and Android on 22 July. It is built for coding, agentic work, research and complex knowledge tasks.

This guide covers what Grok 4.5 does, what it costs in pounds and through the API, how much weight its benchmark results deserve, and the questions worth settling before you put it anywhere near client work.

Update, 24 September 2026: Grok 4.6 launched on 12 August and Grok 4.7 followed on 21 September. Read our Grok 4.7 guide for current access and pricing, and our Grok 4.6 review for the cost lessons from that release. This guide retains the Grok 4.5 launch specifications, prices, benchmark snapshots and governance discussion originally checked on 26 July and 10 August. For a new project, test Grok 4.7 first and compare the cost of completing your own tasks. Historical figures below are not a current market price list or leaderboard.

Worth clearing up before we start, because search results confuse the two constantly: Grok is the model and assistant built by SpaceXAI, the company formerly called xAI. Groq, spelled with a Q, is a separate American company that makes inference hardware. They are unconnected.

The answer in brief

• Grok 4.6 and then Grok 4.7 have succeeded 4.5. The 12 August release is available through the API, Grok Build and supported partners. Keep a working 4.5 integration only while its availability and results meet your requirements; test a replacement before changing production traffic.
• 500,000-token context window, text and image input, text output, knowledge cut-off of 1 February 2026.
• Reasoning cannot be switched off and defaults to the highest of three settings, which has a direct effect on latency and on the bill.
• £30 a month for SuperGrok on the UK App Store, with a free tier to start. X Premium+ is £31 a month.
• $2 and $6 per million tokens on the API below a 200,000-token prompt. Cross that line and both rates double, across the whole request.
• The context window went backwards. Grok 4.3 handles a million tokens and costs less.
• Independent scoring sits at 56 on the Artificial Analysis Intelligence Index, up two points from the 54 it held through late July. The rank and speed figures on the same page moved substantially within a single day, which tells you something about how much weight leaderboard positions can carry.
• No Grok 4.5 model or system card has been published.

What is Grok 4.5?

Grok 4.5 is an earlier general-purpose reasoning model from SpaceXAI. It takes text and images, returns text, works with up to 500,000 tokens of context, and can search the web and X, run code and call functions you define.

At launch, SpaceXAI described Grok 4.5 as its strongest model for coding, agentic tasks and knowledge work. Its API identifier is grok-4.5. Grok Build has since gained Grok 4.6 and Grok 4.7 Fast, so a moving alias or product default should not be treated as a permanent reference to 4.5. Check the developer release notes before relying on an alias in production.

The word “latest” needs a category attached, and this is where a lot of published coverage falls over. Grok Imagine and Grok Voice are separate image, video and speech products with their own release cycles, not successors for general reasoning. Grok Build is a coding agent rather than a model, and grok-build-0.1 remains a separately priced endpoint. Grok 4.3 and the Grok 4.20 family are still on the price list and still useful, particularly where context size matters.

There is a corporate wrinkle too. SpaceX acquired xAI on 2 February 2026, and the company now trades as SpaceXAI. Its consumer terms, enterprise terms and privacy policy all name SpaceXAI LLC (checked 24 September 2026), while the API, console and SDK keep the xAI name and regulators’ files from earlier in 2026 name X.AI LLC. This article uses SpaceXAI for the company today, xAI for product names and earlier events, and whatever name a regulator used in its own records.

Grok model timeline

Two of those releases are easy to miss. Neither Grok 4.20 nor Grok 4.3 got an announcement post on x.ai/news; both appear only in the developer release notes, the pricing table and third-party cloud catalogues. Anyone working from the company blog alone skips a whole generation, which explains a good deal of the confusion in circulation.

The Grok 4.5 release itself was staged rather than single-dated. The release notes record API availability on 8 July and EU console access on 17 July, the launch article carries 16 July, and the consumer rollout is dated 22 July. Citing all four beats picking a favourite.

What changed after Grok 4.5?

Our August check preceded the release of Grok 4.6 on 12 August, and Grok 4.7 followed on 21 September. Both keep the $2 and $6 standard rates of Grok 4.5 but charge $0.50 rather than $0.30 per million cached input tokens (SpaceXAI release notes, checked 24 September 2026). The earlier advice to wait for Grok 4.6 is now out of date. The useful comparison is between the model already running your workflow and the newest option: accuracy, latency, tool use and total cost per finished task. A future release rumour is not a reason to delay that assessment.

Grok 4.5 features and specifications

The context window became smaller

Grok 4.3 carries a million-token window. Grok 4.5 gives you half of that, and Grok 4.3 is cheaper too, at $1.25 and $2.50 per million tokens against $2 and $6. A newer version number does not mean a bigger specification, and on this one the older model wins outright.

Context size is a weak buying criterion in any case. A large window is no assurance that the model will locate, interpret or cite the right passage inside it. Structured source packs, clear labelling and a retrieval test you can repeat will do more for output quality than another half-million tokens of headroom.

Reasoning effort, and why it costs you

Three settings are available: low, medium and high. The reasoning documentation states that the setting defaults to high when unspecified, and that reasoning cannot be disabled at all.

That is a billing fact dressed as a technical one. Reasoning tokens are generated whether or not your workload needs them, and the default is the most expensive option. Low effort suits classification and extraction, medium covers most research and planning, and high earns its cost on difficult code or multi-stage investigations. Settle the question with a fixed evaluation set rather than an assumption that more thinking is always better.

Live information and tool use

Native X search alongside web search is Grok’s clearest structural advantage, for the straightforward reason that no competitor owns a major social platform. For gathering current sources, tracking reaction to a campaign or spotting a narrative before it hardens, it does something the alternatives cannot.

It does not turn social posts into market research. Platform demographics, coordinated activity, spam and a small number of very loud accounts all distort what looks like consensus. Treat the output as one qualitative signal and read it beside search behaviour, analytics, support tickets and actual customer conversations.

Tools bring their own exposure. A retrieved page can carry instructions aimed at the model rather than the reader. A function call can change data that is awkward to change back. Code execution can produce an analysis that is confident and wrong. Isolate untrusted content, keep permissions narrow, log what the model does, and require a human to confirm anything that publishes, spends, contacts a customer or edits a record.

Grok 4.5 pricing and access

The 22 July announcement confirms Grok 4.5 powers Grok on the web, on X and in the mobile apps. Developers reach it through the xAI API. The launch post adds that at launch it was the default model in Grok Build and available in Cursor on all plans. The developer guide also listed gateways including OpenRouter, Vercel, Cloudflare, Snowflake and Databricks Mosaic at launch; that page now redirects to the Grok 4.7 guide. The launch post also describes it as having been trained alongside Cursor, which is a closer relationship than gateway availability.

Grok’s consumer products handle spreadsheets, presentations, PDFs, voice and generated media. Those are product features built around the model. They are not evidence that the base API emits anything other than text.

What it costs in the UK

SpaceXAI’s plan page lists a free tier and SuperGrok at $30 per month. It no longer names Grok 4.5: as of 24 September 2026 the only model it lists is Grok 4.6, ticked for every plan including Free. In sterling, the published figures are:

One caveat carries real weight. Apple lists purchase amounts without labelling billing periods, so while the £10, £30, £100 and £300 tiers line up plausibly with monthly and annual equivalents, that mapping is inferred rather than stated. Checkout remains authoritative for period, tax, promotions and eligibility.

API pricing

From the official pricing documentation, in US dollars per million tokens (checked 24 September 2026):

The threshold behaves in a way that catches people out. Once a prompt reaches 200,000 tokens, the higher rate applies to every token in the request rather than only the excess. Sending 190,000 tokens costs about $0.38. Sending 210,000 costs about $0.84. Prompt design is a commercial decision here, not only a technical one.

The threshold is not unique to the flagship either: grok-4.3 doubles to $2.50 and $5.00 past the same 200,000-token line, and grok-build-0.1 to $2.00 and $4.00. Tools are billed on top: web search, X search and code execution at $5 per 1,000 calls, file attachments at $10 per 1,000, collections search at $2.50 per 1,000.

For a sense of scale, a single request with 50,000 uncached input tokens and 5,000 output tokens comes to roughly $0.13 before tool fees. Agentic workflows make many calls and several retries, so track cost per finished piece of work, reviewer time included, rather than per token.

Grok 4.5 sits among the lower-priced frontier-class APIs without being cheapest on every measure. Gemini 3.6 Flash lists a lower input price at $1.50; Grok has the lower output price at $6.00 against $7.50. Any claim built on the word “cheapest” needs a defined workload behind it.

Grok 4.5 benchmarks: what the numbers actually show

SpaceXAI publishes strong software-engineering results. These are vendor-run figures, taken from its own launch post.

The company also reports around 80 tokens per second and roughly 4.2 times fewer output tokens than Claude Opus 4.8 at maximum effort on SWE-Bench Pro, at 15,954 against 67,020. To its credit, SpaceXAI wins outright on only one of its own five charts, which is more candid than launch material usually manages.

Independent measurement adds two things worth having.

Artificial Analysis, which runs its own evaluations, scores Grok 4.5 at 56 on its Intelligence Index at our 10 August re-check, up from the 54 it held across repeated checks in late July, a two-point drift that says more about evaluation suites being re-run than about the model. The rank and throughput figures on the same page are far less stable again: three readings taken on 26 July returned ranks between fourth and thirteenth, against comparison sets ranging from 170 to 190 models, with output speeds between 9.8 and 91.3 tokens per second. Nothing about the model changed in those hours. The composite score is the most durable number on the page; a rank or a tokens-per-second figure quoted without a timestamp is close to meaningless.

The public Terminal-Bench 2.1 leaderboard records a Cursor CLI run of Grok 4.5 at 79.3% ± 1.5%, fourth of seventeen entries, submitted on 9 July 2026. SpaceXAI’s own figure is 83.3. The four-point gap sits outside the stated confidence interval, and neither party documents the cause clearly enough to call one number wrong. Different harness versions, reasoning settings and run conditions could all account for it, which is precisely the point: a model score is inseparable from the agent, tools, effort setting, token budget and scoring method that produced it.

For procurement, build 30 to 100 private tasks that resemble your actual work. Blind the outputs where you can, and score factual accuracy, completion quality, latency, cost, reviewer time and how badly the failures fail. A marketing evaluation might include campaign summaries, source-grounded briefs, structured data, landing-page code, spreadsheet analysis and customer-safe copy. Our AI strategy and consultancy team runs this kind of structured comparison for exactly this decision.

What do we know about factual reliability?

No broad independent reliability study of Grok 4.5 had been published at the cut-off. The Artificial Analysis index includes knowledge and non-hallucination components, but a composite of 56 is not a promise that any particular citation, calculation or campaign claim will hold.

Studies of earlier Grok versions are useful for deciding what to test and useless as evidence about this model. Put grounding, citation accuracy, source-date checking, expressed uncertainty and deliberately adversarial search results into your private evaluation. For anything you intend to publish, open every material citation and recalculate the important numbers somewhere other than the model.

Practical uses for marketing and business teams

Current-topic research and social listening. Web and X search together will gather sources, surface the questions people are actually asking and track reaction to a launch. Ask for primary sources, journalism, opinion and user-generated content to be separated, insist on dates and URLs, then check the ones that matter. Treat X data as a qualitative signal and read it alongside search behaviour, analytics and customer interviews.

SEO and answer-engine content. Grok will turn a verified source pack into a brief covering intent, entities, headings, comparison dimensions, FAQs and answer blocks, and it is genuinely good at flagging a statistic with nothing behind it. It cannot produce defensible keyword volume without a real data source, and it cannot promise a ranking. Pair it with Search Console, Keyword Planner or a licensed platform, then add the first-hand experience no model has. That combination underpins our answer engine optimisation work.

Content and paid-media variants. Headlines, ad angles, email variants, social adaptations, a webinar turned into an article. Supply approved examples, a voice guide, a prohibited-claims list and the audience. Before anything goes live, check originality, trademarks, regulated claims, platform policy and landing-page consistency. Volume for its own sake has never been a durable search strategy and is a worse one now.

Analysis, reporting and office workflows. Code execution and the Microsoft add-ins can clean data, explain an anomaly, draft a report or turn an outline into a first-pass deck. Keep the source data immutable, validate totals somewhere you trust, and make the model state its filters, definitions and assumptions. Quiet assumption changes are the failure mode here, and they are hard to spot in a polished summary.

Coding and marketing automation. Grok Build and API tool calling speed up landing-page prototypes, schema generation, analytics scripts, feed transformations and internal tooling. Treat what comes back like a pull request from a capable new starter: review it, test it, scan dependencies and secrets, and keep a named human accountable for deployment.

Grok 4.5 compared with ChatGPT, Claude and Gemini

No model wins for every organisation. The table uses each vendor’s own documentation and headline API rates per million tokens, checked on 26 July 2026. It is not a quality ranking.

Sources: OpenAI, Anthropic, Google, Meta. We have covered three of these rivals in depth: GPT-5.6 Sol, Claude Opus 5 and Claude Fable 5.

Read the context column with care, because the vendors are not counting the same thing. OpenAI quotes 1.05 million as a combined input and output figure. Google publishes separate ceilings, at 1,048,576 in and 65,536 out. Anthropic quotes a million in with a 128,000 output cap. SpaceXAI quotes 500,000 and publishes no output limit at all. Google also doubles its input price above 200,000 tokens, as does Grok above its own threshold, so a flat headline rate flatters whichever model you happen to look at first.

The comparison above is a July/August snapshot. Since then, Claude Fable 5.1 and GPT-6 Astra have arrived, while GPT-5.6 Sol’s API pricing has been reduced for a promotion. Use those newer guides when choosing a model today; keep the original table for understanding the 4.5 launch.

Beyond the numbers, procurement should weigh data location, retention terms, security evidence, rate limits, support and integration effort. Those differences usually dwarf the token price.

Privacy, safety and UK governance

Consumer and business data are governed differently

SpaceXAI’s consumer FAQ states that it “may use your content and interactions with Grok” to train its models, and gives a specific opt-out path: Settings, then Data Controls, then “Improve the model” in the mobile app, or Settings, then Data, then “Improve the Model” on Grok.com. Private Chat is excluded from training. The same page confirms that “A limited number of our authorized personnel may review your conversations with Grok” for purposes including model improvement and security investigation, and that automated classifiers monitor usage.

Business accounts sit on different terms. The FAQ says plainly: “We do not use content from our business and enterprise customers to improve our models.” The Enterprise Terms go further and make it contractual, stating that SpaceXAI will not use user content to train foundation models, large language models or other AI systems. Enterprise content is generally deleted within 30 days, subject to legal and safety exceptions, and Zero Data Retention shortens that further while moving the record-keeping burden onto you.

The practical rule is short. Never run client work through a consumer account. Grok.com, X, the mobile apps, the API and a third-party gateway do not share terms, and the difference between them is contractual rather than cosmetic.

Residency and security need contractual confirmation

SpaceXAI advertises data residency as an Enterprise feature on its pricing page, described as controlling where your data lives, but publishes no UK or EU location matrix for Grok 4.5. Its subprocessor list runs to more than twenty entries covering infrastructure, payments, support and monitoring. Almost all are recorded as US, with one exception: an entry for its own subsidiary located in the UK. The list carries no last-updated date, which matters because the data-processing addendum incorporates it as amended from time to time.

None of that tells you where a particular customer’s prompts will be processed. Ask for the processing locations, the transfer mechanism, the retention setting, audit evidence, incident terms and the subprocessor notification procedure, in the contract. A pricing page is not a data-processing agreement.

Safety documentation remains incomplete

No dedicated model or system card for Grok 4.5 could be found at the cut-off. The most recent detailed card is for Grok 4.20, dated April 2026, and its findings cannot be read across to a different model.

That is not evidence the model is unsafe. It does mean a risk team has less public material to work with than it would for several competitors, and that your own testing carries more of the load. Test hallucination, prompt injection, harmful output, data leakage, bias and refusal behaviour against your own workflows before anything customer-facing goes live. Our AI content governance work covers that ground.

Regulatory status

Three matters are open, and none has produced a finding.

Ofcom opened an Online Safety Act investigation into X on 12 January 2026, concerning Grok being used to create and share illegal sexualised imagery. In a scope update on 3 February it set out a jurisdictional limit in its own words: “we are currently unable to investigate the creation of illegal images by the standalone Grok service in this case”, and “We are not investigating xAI at this time.” Standalone chatbots that are neither user-to-user nor search services fall largely outside the Act. Ofcom added that it is examining whether to open a separate investigation into age-assurance duties.

The ICO went further than Ofcom could. On 3 February 2026 it opened formal investigations into both X Internet Unlimited Company and X.AI LLC over personal-data processing and safeguards. It is the only UK regulator to have named the model developer directly, which is worth knowing when you are completing a vendor-risk questionnaire.

The European Commission opened Digital Services Act proceedings on 26 January 2026.

Canada’s Privacy Commissioner reached a conclusion where the others have not, issuing PIPEDA Findings 2026-004 in June 2026 concerning X Corp and X.AI LLC. That office has no power to fine. Treat each regulator’s jurisdiction and conclusion separately rather than collapsing them into a single global verdict.

Who should use Grok 4.5, and who should not

Strong candidate for teams that need live X and web information inside a workflow, cost-sensitive text generation at volume, coding assistance, or tool-driven automation. It is most compelling where the X data advantage matters more than having the largest available context window.

Probably the wrong choice for organisations that need a published UK or EU processing-location matrix, a dedicated safety card for the exact model in use, or more than 500,000 tokens of context, where Grok 4.3 gives twice the window for less money. It is also a poor fit for any workflow that cannot absorb an occasional factual or tool error without human review. Regulated organisations should take current contractual, security and legal advice rather than relying on a general comparison.

A practical adoption checklist

1. Choose one bounded workflow with a measurable outcome.
2. Classify the data, then pick the consumer, business or API route accordingly. Never the consumer route for client work.
3. Build a representative test set, including the difficult and adversarial cases.
4. Run at least one alternative model on identical instructions and evidence.
5. Measure accuracy, reviewer time, latency, token cost and tool fees.
6. Keep prompts below 200,000 tokens where practical, and test prompt caching.
7. Set reasoning effort deliberately, because the default is the most expensive one.
8. Apply least-privilege tool access and require confirmation for consequential actions.
9. Require primary-source checks and named human approval before publication.
10. Record model version, settings and incidents where lawful, and diarise a review of prices, terms and regulatory status.

The balanced verdict

Grok 4.5 earns a place on the shortlist for coding, cost-sensitive generation, source-assisted research and anything that benefits from live X data. The price is genuinely competitive and the X integration is something no rival can replicate.

The limitations are just as concrete. A smaller context window than the model it replaced, at a higher price. Rates that double at a threshold you can cross without noticing. No published safety documentation. A set of open regulatory questions that a cautious board is entitled to weigh.

None of that settles the decision either way. It means the decision should be made against your own requirements and your own test results rather than a launch chart. The organisations that get value from AI are rarely the ones that picked whatever sat at the top of a leaderboard. They are the ones who ran three candidates against their own work, know what each costs at their own volumes, and can say which of their workflows survives a wrong answer.

If you want that comparison run properly, talk to AIWIZ Digital Marketing. We will tell you what we would actually do, including the times the answer is to stay where you are.