Grok 4.5 Pricing: xAI API vs Grok Build vs Cursor

Last updated: July 13, 2026

Grok 4.5 is SpaceXAI’s coding and agentic model, released on 8 July 2026 and trained jointly with the code editor Cursor [1, 6]. It reaches buyers through three routes: the direct xAI API, the Grok Build command-line agent, and Cursor [1, 3, 6]. Those routes differ in price, usage limits, data handling, contract, and version control, so the access path, not “Grok versus another model,” is the first purchase decision.

Best for

Teams that want a 500K-context model on the direct API and run their own agent loop. Existing Cursor users who value the editor and agent workflow as much as the model. Buyers willing to measure cost per accepted result rather than compare token prices alone.

Not for (yet)

Teams that need a public numeric Cursor allowance before budgeting. Buyers who need confirmed European entitlement, data residency, a pinned model release, or a model-specific SLA before evaluating. Anyone who assumes a bring-your-own-key setup bypasses Cursor’s infrastructure.

Key facts at a glance

Item	Public documentation, as of 12 July 2026	Evidence boundary
Launch	SpaceXAI and Cursor announced Grok 4.5 on 8 July 2026 [1, 6]	Vendor announcement
Direct xAI API price	500K context; $2.00 input, $0.50 cached input, $6.00 output per 1M tokens [2, 3]	Volatile; recheck before publish
Priority route	Priority Processing bills 2x standard token rates when priority is delivered [2]	Different from Cursor’s fast variant
Cursor	Included on paid plans with qualitative usage language; fast variant listed at $4 input / $18 output [6]	No numeric Grok allowance on the public page
Long context	xAI says requests above 200K use a different rate [3]	The rate is not shown on the model page
Efficiency claim	About 15,954 output tokens per SWE-Bench Pro task, roughly 4.2x fewer than xAI’s Opus 4.8 figure [1]	Vendor comparison; not independently matched
FSR evidence level	Document-first only	No hands-on test or benchmark

CONTENTS

One brand, three control planes
The $2 / $6 rate is not a workload budget
What the benchmarks establish, and what they do not
Three data paths, one privacy question
Europe and regions: four different facts
The procurement checklist
Which surface fits which team
Alternatives by buyer constraint
FAQ
How we checked this
Sources and change log

One brand, three control planes

The first buying decision is direct xAI API versus Grok Build versus Cursor. The routes share a brand, but they do not give a buyer the same price, the same usage visibility, the same data path, or the same contract.

On the direct API, the price is straightforward: $2.00 per million input tokens, $0.50 for cached input, and $6.00 for output, with a 500,000-token context window [2, 3]. The model page lists the aliases grok-4.5-latest and grok-build-latest [3]. xAI’s enterprise terms name X.AI LLC as the contracting party for the API and related business services [11].

Grok Build is harder to pin down as a commercial object, and the ambiguity is worth stating plainly because it can distort a price comparison. xAI’s launch page says Grok 4.5 is the default model in Grok Build [1]. The same pricing catalog that lists grok-4.5 also lists a separate Code API model named grok-build-0.1, with a 256K context window and rates of $1 input, $0.20 cached input, and $2 output per million tokens [2]. The Grok 4.5 model page then lists grok-build-latest as an alias [3]. These pages may describe different layers of one product, but the public documentation does not give a buyer a clean mapping among the Grok Build agent, grok-4.5, grok-build-latest, and grok-build-0.1. A team should not assume Grok Build inherits the raw grok-4.5 price and context, and should not assume it bills at the cheaper grok-build-0.1 line, without confirming which identifier its invoice uses.

GROK BUILD: FOUR IDENTIFIERS, NO PUBLIC MAP

Launch page: Grok 4.5 is the default model in Grok Build.
Model page: grok-4.5, 500K context, aliased as grok-4.5-latest and grok-build-latest.
Pricing catalog: a separate Code API model grok-build-0.1, 256K context, $1 / $0.20 / $2.
Buyer action: confirm in writing which identifier your Grok Build usage bills against before you model its cost. The reviewed public pages do not resolve this.

Cursor is the third control plane. Its launch post says paid individual and team subscriptions include significant Grok 4.5 usage and lists a fast variant at $4 input and $18 output per million tokens [6]. That $4 / $18 figure is Cursor’s, and it is a different mechanism from xAI’s Priority Processing. Cursor’s public plan language describes the allowance qualitatively and then allows on-demand billing after included usage is consumed [6]. It does not publish a Grok-specific numeric allowance, how retries consume it, or whether the fast route draws from the same pool.

Our full Cursor review breaks down how the advertised $20 plan trends toward roughly $60 in daily agent use, and how bring-your-own-key requests still route through Cursor’s own infrastructure.

The access-path matrix

Surface	Model or SKU actually referenced	Contracting party	Base price	Numeric quota visible	Long context	Data route	Serving region	Version pinning
Direct xAI API	grok-4.5 (aliases grok-4.5-latest, grok-build-latest) [3]	X.AI LLC [11]	$2 / $0.50 / $6 per 1M [2]	Yes, per token	500K; different rate above 200K [3]	xAI API terms; no training without permission [4]	us-east-1, us-west-2 [3]	Alias exists; fixed dated ID not confirmed [3]
Grok Build	Grok 4.5 default; grok-build-0.1 also listed separately [1, 2]	X.AI LLC (assumed) [11]	Not clearly mapped in public docs [1, 2, 3]	Not documented in this review	Depends on the billed identifier	Not separately documented in this review	Not separately documented	Not documented
Cursor	Grok 4.5 via Cursor; provider routing applies [6, 7]	Anysphere, plus model-provider dependencies [7, 8]	Plan-bundled; fast variant $4 / $18 [6]	No public numeric Grok allowance [6]	Follows API; plan accounting unclear	Cursor backend; Privacy Mode governs training and retention [7]	Depends on Cursor and its providers	Managed by Cursor

The reader who approves only the model name has not finished the review. The surface, plan, privacy mode, provider route, region, and terms still need to be named.

The $2 / $6 rate is not a workload budget

The standard API rate is competitive, and the efficiency claim has support. xAI reports 15,954 output tokens per SWE-Bench Pro task, about 4.2 times fewer than its Opus 4.8 comparison [1]. Independent testing by Artificial Analysis also found Grok 4.5 in the Grok Build harness used fewer tokens than several competing coding-agent systems [9]. That evidence supports an efficiency case inside the harnesses that were measured. It does not produce a production budget on its own.

Grok 4.5 pricing and speed at launch: — xAI’s launch pitch for Grok 4.5: speed, token efficiency, $2 / $6 pricing, and where to access it. Prices and availability are volatile vendor claims and should be rechecked; the model was not yet available in the EU at launch.

Four cost layers sit above the sticker price. Priority Processing doubles input, cached, reasoning, and output rates when priority service is delivered [2]. Server-side tools add per-invocation charges [2]. Requests above 200K context use a rate the model page references but does not show [3]. Grok 4.5 does not appear in xAI’s batch-discount table [2]. On top of those, failed runs, retries, human repair, and Cursor’s on-demand overage are workload-specific and are not in any rate card.

THE METRIC THAT MATTERS IS NOT PRICE PER TOKEN

cost per accepted output = (model tokens + tools + priority + context premium + retries + repair + plan overage) ÷ accepted outputs

This does not show that Grok 4.5 is expensive. It shows why the public rate card cannot answer the buyer’s real question alone. On the direct API, log per-request cost, tool calls, retry count, and accepted outputs. In Cursor, capture included usage, fast-route accounting, on-demand charges, and failure consumption from the dashboard before projecting a monthly figure.

What the benchmarks establish, and what they do not

“Opus-class,” the phrase Elon Musk used, has no published equivalence definition, so it is positioning rather than a measurement [1]. The vendor’s own chart is more specific and more mixed. On the five engineering benchmarks xAI showed at launch, its chart puts Anthropic’s Fable model ahead on four (DeepSWE 1.0, DeepSWE 1.1, Terminal-Bench 2.1, and SWE-Bench Pro) and Grok 4.5 ahead on one, SWE Marathon [1]. A reader who takes away only “Grok wins” or only “Grok loses” has misread a split result.

Grok 4.5 leads SWE Marathon but trails Anthropic's Fable on the other four of xAI's five launch benchmarks. — Grok 4.5 across xAI’s five launch benchmarks. On the vendor’s own figures it leads one (SWE Marathon) and trails Anthropic’s Fable on four. Benchmark figures are vendor-published and not independently verified.

Evidence	Source	What it measures	What it establishes	What it does not
Five-benchmark chart	xAI launch page [1]	Provider-run agent harnesses	Grok 4.5 is competitive on engineering tasks	Neutral, matched-harness parity
Token-efficiency figure	xAI launch page [1]	Output tokens per SWE-Bench Pro task	Lower token use inside xAI’s test	Cost per accepted task in production
Coding-agent index	Artificial Analysis [9]	Grok Build as a system	Efficiency of the Grok Build harness	Raw-model behavior outside that harness
CursorBench	Cursor disclosure [6]	Cursor-specific benchmark	That the figure is unreliable and was excluded	That other benchmarks are contaminated

Two caveats close the section. Cursor disclosed that an earlier snapshot of its own codebase was accidentally included in training, that this gave Grok 4.5 an advantage on CursorBench, that the impact was unclear, and that it therefore excluded that benchmark [6]. That makes the CursorBench figure unusable; it is not evidence that the other benchmarks are contaminated. Separately, Artificial Analysis scored the model at 54 on its Intelligence Index and placed it fourth at launch [9]. Rank on that index moves as new models are added, so a launch position is a dated snapshot, not a fixed property.

Three data paths, one privacy question

The data boundary changes with the access path, and Cursor’s own documentation is what establishes the difference.

On the direct API, xAI states that it does not train on API inputs or outputs without explicit permission, that ordinary API data is kept 30 days for abuse auditing and then deleted, and that an enterprise Zero Data Retention option processes requests without persisting them [4]. It states that the platform is SOC 2 Type 2 compliant and that a business associate agreement is available for regulated health data [4]. These are vendor statements, not independent audits, but they are documented.

Through Cursor, the picture differs. Cursor states that even when a customer supplies their own API key, requests still pass through Cursor’s backend, where final prompt construction happens [7]. With Privacy Mode on, Cursor does not use customer data for training and maintains zero-data-retention agreements with providers, with an exception if abuse detectors trigger [7]. With Privacy Mode off, Cursor states it may use codebase data, prompts, editor actions, and code snippets to improve and train models, and may share prompts and limited telemetry with the selected provider [7]. Code indexing and temporary caching add further processing stages [7]. Grok Build’s end-to-end data path is not documented in the same detail in the pages reviewed here, so it should be confirmed separately rather than assumed identical to the raw API.

DID YOUR CURSOR CODE TRAIN GROK 4.5? THE DOCUMENTS DO NOT SAY

Cursor says Grok 4.5 was trained on trillions of tokens of Cursor data [6], and its general rules make Privacy Mode the switch for training use [7]. It is tempting to conclude that a specific developer’s code trained the model if their Privacy Mode was off. No public statement maps the training corpus to accounts, dates, or privacy settings, so the exact boundary is unresolved. The provenance concern is real; the specific claim is not supported.

For European buyers, one more document matters. xAI published a training-content summary for Grok 4.5 in the format the EU AI Act uses for general-purpose models. It reports more than 10 trillion text tokens plus image, audio, and video data, drawn from public, licensed, third-party, synthetic, and user and contractor sources, and it states that interactions with Grok may be used for training subject to privacy settings and opt-outs [5]. It is a training-content summary, not a full technical or system card, so it does not describe architecture, evaluation design, or safety testing [5].

xAI's stated Grok 4.5 training recipe (GB300 GPUs, data curation, reinforcement learning, hours-long agentic rollouts) and its Grok Build workflow in Excel, PowerPoint, and Word. — What xAI’s launch page says about training Grok 4.5 and using it inside Grok Build. Training scale and capability statements are vendor claims.

Europe and regions: four different facts

European availability is easy to state wrongly, because four separate facts get merged. At launch, xAI’s materials said Grok 4.5 was not yet available in the European Union through its products or the API console, and its documentation said availability was expected later in the month [1, 3]. The model page lists only United States serving regions [3]. Its AI Act training summary lists a European Union market placement date of 14 July 2026 and names an authorised representative in Estonia [5].

KEEP THESE APART

Product entitlement: not in EU surfaces at launch, expected later in the month [1, 3].
Market placement (AI Act): 14 July 2026 [5].
Serving region: only US regions listed [3].
Data residency: no European residency commitment located in the reviewed public sources.

The public sources do not state the reason for the launch-window gap, and this analysis does not assign one. Market placement is not the same as a European serving region, product entitlement, or a residency commitment. European buyers should confirm current status in each surface directly, and recheck after 14 July.

The procurement checklist

Several items matter more to a serious buyer than the headline benchmark, and none is a legal conclusion. They are recorded as documented facts and open questions.

Item	What the documents show	What to confirm before standardizing
Version pinning	grok-4.5-latest alias exists; xAI has retired an older model and redirected its identifier [3]	A fixed, dated Grok 4.5 identifier and a change-notice policy
Evaluation rights	Terms differ by access path [11]	Whether your contract permits publishing your own benchmarks
Service commitment	No public model-specific SLA located [4, 11]	Uptime, support, and service credits in the order form
Billing continuity	Prepaid-credit and limit exhaustion can interrupt access, separate from rate limits	The stop behavior and top-up policy for your account
Corporate structure	A June 2026 securities filing describes a proposed, conditional SpaceX and Anysphere merger, near sixty billion dollars, expected to close in Q3 2026 [12]	The latest filing; treat downstream roadmap or contract change as a scenario, not a fact

The billing-continuity row rests on xAI’s billing documentation as summarized in secondary review rather than a page opened for this analysis, so it is marked for direct confirmation.

Which surface fits which team

Future Stack Reviews has not tested Grok 4.5, so the guidance below is document-based and organized by the decision each team actually faces.

Choose the direct API when control matters more than a bundled workflow: you want the 500K context, you run your own agent loop, and you can meter per-request cost, tool calls, and retries. Confirm the fixed model identifier and the above-200K rate first [2, 3].
Choose Grok Build only after you confirm which identifier it bills against, because the public pages do not map the agent, grok-4.5, and grok-build-0.1 to a single invoice [1, 2, 3].
Choose Cursor when the editor and agent workflow are the product you are buying, not just the model. Before you budget, capture the plan’s included Grok usage, fast-route accounting, and on-demand terms from the dashboard, and set Privacy Mode deliberately [6, 7].
Wait, or run a direct-API pilot only, if you need confirmed European entitlement, data residency, a pinned release, or a model-specific SLA, none of which is settled in public documents today [3, 4, 5].

Alternatives by buyer constraint

Future Stack Reviews has not tested these alternatives either; the table maps documented options to the constraint that usually drives the choice, with pricing attributed where a vendor has published it. Ranking and inclusion are not influenced by any commercial arrangement.

If your binding constraint is	Reasonable options to weigh	Note
Lowest token price at high volume	OpenAI GPT-5.6 Luna, reported at $1 input / $6 output; open-weight coding models for self-hosting	Grok 4.5’s $2 / $6 is competitive but not the floor
Peak coding-benchmark standing	Anthropic’s Fable model, which leads four of five rows on xAI’s own chart [1]	Independent, matched-harness testing still advised
Confirmed EU residency now	EU-based providers that document European processing	Verify residency and transfer terms in the contract, not marketing
A single vendor across chat, agent, and API	OpenAI GPT-5.6 tiers (Sol, Terra, Luna) or Anthropic’s current lineup	Compare per-path terms the same way as for Grok 4.5
The IDE workflow itself	Cursor with any supported model, or a competing agentic editor	The editor, not the model, is the primary purchase here

Future Stack Reviews has tested two of these routes hands-on: our Claude Code review and OpenAI Codex review apply the same cost-per-accepted-result lens to the agentic coding tools these models power.

FAQ

How much does Grok 4.5 cost?

The direct xAI API lists $2.00 input, $0.50 cached, and $6.00 output per million tokens, with 500K context, as of 12 July 2026 [2, 3]. Priority Processing doubles those rates when applied, requests above 200K use a different unpublished rate, and tools bill separately [2, 3].

Is the $4 / $18 “fast” price the xAI API rate?

No. The $4 / $18 figure is from Cursor’s launch materials for a fast variant in Cursor [6]. On the xAI API, the faster route is Priority Processing, a 2x multiplier that yields $4 input, $1 cached, and $12 output [2]. They are different products.

Is Grok Build cheaper than the Grok 4.5 API?

Unclear from public documents. xAI says Grok 4.5 is the default in Grok Build [1], but the pricing catalog also lists a separate grok-build-0.1 model at 256K context and $1 / $0.20 / $2 [2]. Confirm which identifier your usage bills against before assuming a price.

Did my Cursor code train Grok 4.5?

The public documents do not answer this for any specific user. Cursor says Grok 4.5 was trained on Cursor data and that Privacy Mode governs training use [6, 7], but no statement maps the corpus to accounts, dates, or settings, so the exact boundary is unresolved.

Can I use Grok 4.5 in the European Union?

At launch it was not in EU surfaces, with availability expected later in the month [1, 3]. xAI’s AI Act summary lists an EU placement date of 14 July 2026 [5]. Placement is not the same as a serving region or residency commitment, so confirm current status directly.

Does xAI train on my API data?

xAI says it does not train on API inputs or outputs without explicit permission, keeps ordinary API data 30 days for abuse auditing, and offers enterprise Zero Data Retention [4]. Through Cursor the boundary differs, because requests pass through Cursor’s backend and depend on Privacy Mode [7].

How we checked this

This is a document-first analysis. Future Stack Reviews did not install, purchase, run, benchmark, or observe Grok 4.5 in any account, and no statement above is a hands-on result. Every price, performance, and availability figure is attributed to the party that published it and dated to the 12 July 2026 access date, because a vendor stating a number is different from an independent party verifying it. This is a source-bound procurement analysis, not legal advice. Volatile items, the price, the above-200K rate, European entitlement, Cursor’s numeric quota, the Grok Build identifier mapping, and the merger’s status, are marked for same-day recheck before publication.

Sources and change log

All sources accessed 12 July 2026. Vendor pages are attributed as vendor statements.

xAI, Introducing Grok 4.5. https://x.ai/news/grok-4-5
xAI Docs, Pricing. https://docs.x.ai/developers/pricing
xAI Docs, grok-4.5 model page. https://docs.x.ai/developers/models/grok-4.5
xAI Docs, API security FAQ (data and privacy). https://docs.x.ai/developers/faq/security
xAI, Public Summary of Training Content for Grok 4.5, version 1, dated 8 July 2026. Published via media.x.ai.
Cursor (Anysphere), Introducing Grok 4.5. https://cursor.com/blog/grok-4-5
Cursor (Anysphere), Data Use and Privacy Overview. https://cursor.com/data-use
Cursor (Anysphere), Privacy Policy. https://cursor.com/privacy
Artificial Analysis, Grok 4.5 launch assessment (Intelligence Index score and coding-agent testing).
xAI, Grok Build changelog. https://x.ai/build/changelog
xAI, Enterprise Terms of Service. https://x.ai/legal/terms-of-service-enterprise
Reported June 2026 securities filing describing the proposed SpaceX and Anysphere transaction (secondary reporting; primary filing to be reconfirmed).

Change log. Rebuilt 13 July 2026 from evidence accessed 12 July 2026, restructured from an earlier internal draft around a source-mapped access-path matrix. Rechecks required before any publication: European entitlement on or after 14 July, the exact above-200K rate, Cursor’s numeric Grok quota, the Grok Build identifier and billing mapping, the billing-stop behavior, and the latest merger filing.