Grok Review 2026: Features, Pricing & Verdict

Pricing after the 12 October unification

Grok’s pricing used to be the most confusing in the category — two parallel tracks, five overlapping products. On 12 October 2026 SpaceXAI collapsed Grok, Cursor and X Premium into one ladder called The X Subscription.

Tier	Monthly	Annual	What you get
Free	$0	—	Limited Grok chat, limited agent requests, basic X posting
Plus	$8	$80/yr	Grok Chat, Voice, Imagine, Build, the Grok API, Cursor with Grok models only, X Premium checkmark, reduced ads. No Grok Bot.
Premium	$30	$300/yr	4x the Plus usage pool, Grok Bot, faster chat and faster image and video generation, cloud agents and automations, ad-free X, preferred handle
Super	$100	$1,000/yr	15x the Plus pool, plus priority model access at peak times
Ultra	$200	$2,000/yr	50x the Plus pool, plus earliest access to new products and models
Teams	$30, $100 or $200 per seat	—	Seat price matches the tier that seat is on; 4–50x the Plus pool with team management, SSO and spend controls

Prices verified 16 October 2026 against cursor.com/pricing and cursor.com/docs/account/pricing, which is where SpaceXAI publishes the unified ladder; the pricing tracker carries the dated change log.

The annual prices are documented, not advertised. Every annual figure above comes from the billing documentation, which appeared on 15 October 2026 — each one exactly two months free. None appear on cursor.com/pricing, which shows monthly rates only and sends you to checkout with annual billing switched off, so the discount is real but you have to go looking for it. Individual tiers only, not Teams or Enterprise.

The legacy plans continue, closed to new signups. SuperGrok at $30, X Premium at $8, X Premium+ at $40 and the Heavy tier all remain for existing subscribers, with SpaceXAI saying it will “eventually” migrate everyone and that there is no timeline. Nothing forces you to switch.

Two things the table does not say. Usage is published only as multiples of Plus, and the size of a Plus allowance is published nowhere, so no buyer can tell in advance whether a tier covers their month. It refreshes weekly rather than monthly, and one pool is shared across the Grok apps, Grok Bot, the Cursor editor, the CLI and cloud agents. And inside Cursor that pool covers Grok models only: Claude and GPT are selectable on Premium and above but billed separately at API list price from the first request, out of prepaid credit.

Recommendation. Plus at $8 if you use X, nothing here if you don’t. Super and Ultra are headroom purchases with no published capacity — measure a month on a lower tier before climbing.

The 200,000-token cliff, and why the flagship is the wrong Grok for long prompts

This is the part of Grok’s rate card that costs people money, and no xAI page draws the conclusion.

Every Grok model on the API has two prices: one below 200,000 prompt tokens and one at or above it. The higher rate is exactly 2x on input, cached input and output alike, and per xAI’s own footnote a request that reaches the threshold is billed at the higher rate for all tokens in the request — not just the ones past the line. One token over and the whole bill doubles.

API rates verified 16 October 2026 against docs.x.ai/developers/pricing, per million tokens:

Model	Context	Input	Cached input	Output	Same, at ≥200k prompt tokens
Grok 4.7 (flagship)	500k	$2.00	$0.50	$6.00	$4.00 / $1.00 / $12.00
Grok 4.6	500k	$2.00	$0.50	$6.00	$4.00 / $1.00 / $12.00
Grok 4.5	500k	$2.00	$0.30	$6.00	$4.00 / $0.60 / $12.00
Grok 4.3	1M	$1.25	$0.20	$2.50	$2.50 / $0.40 / $5.00
Grok Build 0.1	256k	$1.00	$0.20	$2.00	$2.00 / $0.40 / $4.00

Read the context column next to the price column. The 500,000-token flagship costs more than the 1,000,000-token mid-tier, at every prompt length, and the gap widens above the cliff. Take a 250,000-token prompt returning 2,000 tokens:

• Grok 4.7, over the line: $1.00 input + $0.024 output = $1.02
• Grok 4.3, over the line: $0.625 input + $0.010 output = $0.64 — 38% cheaper, with 500,000 tokens of headroom still unused
• Grok 4.7, trimmed to 199,000 tokens: $0.398 input + $0.012 output = $0.41 — cutting 20% of the context cuts 60% of the bill

Three rules fall out. If your prompts live near 200,000 tokens, trimming under the line beats any model swap. If they are genuinely long, 4.3 is the model, not 4.7 — the flagship charges a premium for a shorter ceiling. And if you need a frontier model on a 500,000-token prompt, Grok has no cheap answer, which is where Claude and Gemini — both serving their full 1M window at standard rates — become the honest cross-vendor comparison.

Two multipliers stack on top. Priority Processing is a flat 2x on all token types. The US regional endpoint is 1.1x and is available for Grok 4.7 and 4.6 only, which puts US-resident 4.7 at $2.20/$0.55/$6.60 below the cliff and $4.40/$1.10/$13.20 above it — the same 10% premium that applies to the Amazon Bedrock listing.

The lineup, October 2026

Grok 4.7 — the flagship since 21 September 2026, and what xAI’s docs call “the most capable model we’ve built.” Knowledge cutoff May 2026, 500,000-token context. It shipped on an unchanged $2/$6 rate card, but it is a new and larger base model trained on a mix weighted toward problems that take many hours, and every benchmark in its column is measured at xHigh reasoning effort where 4.6’s column says High. Flat token price, more tokens per task — so measure cost per completed task, not per token. Scores to know: CursorBench 4.0 46.3% against Fable 5.1’s 51.8%; EEBench 64.0%, best in table; Harvey Legal 19.6%, three times the next-best; Terminal-Bench 4.0 38.0% against Fable 5.1’s 57.9%, which is the row that matters for unattended agents and the one where the cheap rate does not pay.

Grok 4.6 and 4.5 — both still sold at the same $2/$6. 4.6 was the August post-training refresh that took the Artificial Analysis Intelligence Index from 56 to 61; 4.5 was the first flagship under the SpaceXAI name and has the cheapest cached input of the 500k models at $0.30. Select either only to pin behaviour you have validated.

Grok 4.3 — the cost-efficient workhorse and, per the arithmetic above, the right default for most production traffic. 1M-token context, native video input, built-in reasoning, $1.25/$2.50. It is also the redirect target for every model xAI retired on 15 May 2026, and the only Grok line with a 20% Batch API discount.

Grok Build 0.1 — the purpose-built coding model, 256,000-token context, $1/$2. Plus subscribers and above reach it through OpenCode with no separate API billing, which is a real cost win against Claude Code or Codex if you are already paying for X.

Grok Imagine — the image and video line, now the product name across xAI’s docs; the “Aurora” branding the models launched under no longer appears there. Images run $0.02 per image on grok-imagine-image, $0.04 on grok-imagine-image-2.0 (the recommended one) and $0.05 on grok-imagine-image-quality. Video is $0.020/sec on the 1.5-lite tier, $0.050/sec on the base model and $0.080/sec on grok-imagine-video-1.5. Against Runway and Veo 3 the trade is consistent: Grok Imagine wins on speed and price, the other two on cinematic quality and prompt adherence.

One dated migration to put in the calendar. grok-imagine-image-quality is deprecated and retires on 2 November 2026, redirecting to grok-imagine-image-2.0 at quality: low — a cent per image cheaper, and a silent output-quality change rather than an error. It is also the chained redirect target for grok-imagine-image-pro, retired in May, so a 2024-era integration can be two hops from a model that stops existing in two weeks. If you generate images on the Grok API, pin grok-imagine-image-2.0 explicitly now. Our model retirement tracker carries the row.

Retired, and still in a lot of old code. grok-4-0709, grok-4-fast-*, grok-4-1-fast-*, grok-3 and grok-code-fast-1 all went on 15 May 2026. Nothing breaks — xAI keeps the slugs resolving and silently redirects them — but billing switches to the replacement’s rates, so a grok-4-1-fast integration that budgeted $0.20/$0.50 has been paying Grok 4.3’s $1.25/$2.50 since May without a single error in the logs. Grep your code for the old slugs. Grok 4 Heavy no longer appears in xAI’s documentation at all, and the Heavy subscription is closed to new signups.

Where you can buy it

Four surfaces: inside the X app, where most casual users meet it; grok.com, the standalone web app; the xAI API at docs.x.ai, rate card above; and Amazon Bedrock since 19 August 2026, which matters mainly to enterprises that could not easily onboard xAI as a new vendor — note the feature surface differs by Bedrock endpoint.

Grok has also been on the Department of War’s GenAI.mil portal since 31 August 2026. That matters to commercial buyers only as a caution: agencies buy Grok for cents under a GSA arrangement, a number with no bearing on commercial rates.

What Grok genuinely does well

Real-time X integration. Grok’s most defensible advantage, and nothing else can do it. Ask what the reaction to an event is and Grok searches X live, pulls trending posts and synthesises. If your work touches public sentiment, breaking news or anything happening on social media, Grok is uniquely valuable.

This stays intact in the consumer apps, where X Search is part of the subscription. On the API it got more expensive on 21 September 2026: X Search moved from $5 per 1,000 calls to $5 per 1,000 posts fetched plus $10 per 1,000 profiles, with parent and quoted posts counted and no documented way to cap result volume. If you are building on the API for X data, the denominator is the whole story.

Image generation. Grok Imagine is fast, photorealistic and often better at rendering text in images than the alternatives. Not Midjourney-grade for artistic work, but for “generate a realistic image of X” it is quick and usually equal. Included in the subscription from Premium up.

Less filtering on gray-zone topics. Grok engages with questions ChatGPT and Claude routinely hedge on — political analysis, contested opinions, direct assessments. For analysts who need an unfiltered read this is valuable; for anyone who wants guardrails it is the opposite.

Extended reasoning, research and voice. Reasoning effort is configurable up to an xhigh setting; DeepSearch runs 5–20 minutes of web and X research into a cited report, comparable to Gemini’s and ChatGPT’s; voice is on par with ChatGPT’s, with customisable personalities. One API note: grok-voice-2.0 speech-to-speech is marked deprecated in xAI’s docs at the same price as grok-voice-3.0 — move to 3.0.

Where Grok falls short

A 500,000-token flagship in a 1M-token market. Grok 4.7’s window is half of what Claude’s and Gemini’s current models serve, and the half you do get above 200,000 tokens costs double. For document-analysis work this has gone from a Grok strength to a Grok weakness.

Terminal-Bench. 38.0% against Fable 5.1’s 57.9% is a wide gap on the benchmark that predicts whether an agent finishes a task unattended, and a cheap rate card does not rescue a run that needs three attempts.

Writing quality behind Claude. Grok is good but not great at long-form prose, and the voice reads slightly contrarian even on neutral topics. For anything destined for a published byline, Claude is the better drafting tool.

The Musk association, and the filtering that comes with it. Not a technical limitation but a market reality: the branding is tied to Elon Musk’s persona, and the “anti-woke” positioning means some professional buyers will not touch it regardless of capability. The less-filtered outputs that are a feature on contested topics also read, on neutral ones, as pushing a perspective — a different one from mainstream AI, but a perspective. Related: the image line has drawn public controversy for generating what other models refuse, which is a live consideration for brand-sensitive work.

Unpublished capacity. The ladder is clean now, but you cannot price it. “4x the Plus pool” is unactionable when the Plus pool has no published size, and the weekly refresh means a heavy week cannot borrow from a light one.

Smaller productivity ecosystem, thinner niche expertise. No equivalent of Projects, Artifacts or Custom GPTs. Strong on mainstream knowledge and X-sourced information, thinner on specialised academic, medical and technical ground where Claude’s training goes deeper.

Grok vs the alternatives

If you need	Buy	Where Grok lands
Live social sentiment	Grok	No real competition
Professional writing	Claude	Last of four; the voice is its weakest mode
Unattended coding agents	Fable 5.1 / Claude Code	Capable, but Terminal-Bench says it needs supervision
Long prompts on a budget	Grok 4.3	The in-house answer, and cheaper than the flagship
Long prompts at the frontier	Claude or Gemini	Grok has no cheap 500k option
Cheapest competent tokens	Claude Haiku 5.5 / DeepSeek	Grok Build 0.1 at $1/$2 is close
Image generation	Midjourney, then ChatGPT/Gemini	Fast and good, not Midjourney-grade for artistic work
Research with citations	Gemini ≈ Perplexity	DeepSearch is good, not best

Who should and shouldn’t buy it

Buy it if you are on X daily (Plus at $8 is good value on its own); if you do social-media, journalism or PR work that needs the live X feed; if you run long-prompt production traffic, where Grok 4.3 is one of the better rate cards in the category; if you are already paying for X and want Grok Build through OpenCode with no separate API billing; or if you want less filtering on contested topics.

Don’t make it your primary AI if you don’t use X (the best pricing is X-bundled and the general assistants are better); if you write for a byline; if you run unattended agents, where the Terminal-Bench gap is real; if you need a frontier model on very long prompts, where 500k with a 200k cliff is the wrong shape; if your business has conservative content-moderation requirements; or if the Musk association is disqualifying, because it does not wash off.

FAQ

Can I still buy SuperGrok at $30? No. SuperGrok, X Premium, X Premium+ and the Heavy tier all closed to new signups on 12 October 2026. Existing subscribers keep them indefinitely — SpaceXAI says it will “eventually” migrate everyone and has published no date. New buyers pick from Free / Plus $8 / Premium $30 / Super $100 / Ultra $200.

Is $8 really the whole price? For Grok chat, voice, image and video generation, Grok Build, the API and Cursor with Grok models, yes. Three things sit outside it: Grok Bot needs Premium at $30, third-party models in Cursor are billed separately at API list price from the first request, and API usage beyond the included pool is prepaid credit at the rates in the table above.

Which Grok model should I call on the API? Grok 4.3 for most things — $1.25/$2.50, a 1M-token window and a Batch API discount. Grok 4.7 when the task genuinely needs the frontier and the prompt is comfortably under 200,000 tokens. Grok Build 0.1 for agentic coding at $1/$2. Grok 4.6 and 4.5 only to pin behaviour you have already validated.

What is the catch on the 200,000-token price? That it applies to the whole request, not to the tokens above the line. A 200,001-token prompt costs twice what a 199,999-token prompt costs, on every token including cached ones. On a long-running agent session the context crosses once and then stays across, so every subsequent turn is repriced. There is no flag and no opt-out.

Does Grok still have a 2-million-token context window? No, and it never does now — that figure is not on xAI’s docs under any model. The flagship Grok 4.7 is 500,000 tokens; Grok 4.3 and the 4.20 series are 1M; Grok Build 0.1 is 256,000. If you were choosing Grok for context length, 4.3 is the model and Claude or Gemini are the cross-vendor comparisons.

The verdict

Since 12 October the consumer question mostly dissolves: at $8, Plus is cheap enough that “is Grok worth a second subscription?” stops being a real calculation for anyone already on X. For non-X users the value drops just as sharply, because Grok’s differentiators only matter to a specific audience and Claude and ChatGPT cover general professional work better.

The developer question is sharper, and it runs the other way from the marketing. Grok’s best API value is not its flagship — it is Grok 4.3, which is cheaper per token and carries twice the context ceiling. Buy the flagship for frontier reasoning on short prompts, buy 4.3 for everything long, and whatever you buy, keep the prompt under 200,000 tokens — that single line is worth more than any model choice on the page.

What changed

• 16 Oct 2026: Annual billing is documented at $80/$300/$1,000/$2,000 a year — two months free, docs only, not on the public pricing page. Rewrote the API section around the 200,000-token 2x cliff; corrected the context window from a claimed 2M to the documented 500k on Grok 4.7; flagged grok-imagine-image-quality’s 2 November retirement and the 15 May retirements still silently rebilling old code.
• 13 Oct 2026: SpaceXAI unified Grok, Cursor and X Premium into one ladder on 12 October — Free / Plus $8 / Premium $30 / Super $100 / Ultra $200, with SuperGrok, X Premium+ and Heavy closed to new signups. Usage is published only as multiples of Plus.
• 21 Sep 2026: Grok 4.7 shipped on an unchanged $2/$6 rate card, but with every benchmark measured at xHigh effort rather than 4.6’s High.
• 18 Sep 2026: X Search API billing changed denominator from $5 per 1,000 calls to $5 per 1,000 posts fetched plus $10 per 1,000 profiles.
• 12 Aug 2026: Grok 4.6 became the default model in Cursor, pricing unchanged at $2/$6 per million.
• 8 Jul 2026: Grok 4.5 launched as the first flagship under the SpaceXAI name, at $2/$6 per million.