Plain AI Daily

Grok 4.5 Explained: Retired, and Grok 4.7 Has Now Arrived Behind the Paid Modes

By 21 min read

Grok 4.5 has been retired. SpaceXSI replaced it with Grok 4.6 across every plan including Free in August 2026, then announced Grok 4.7 on September 21. Its pricing page has since been rebuilt: the plans are now Free, Plus $8, Premium $30, Super $100 and Ultra $200, and the model row is gone.

Grok 4.5 is no longer the model you get. SpaceXSI retired it from every consumer Grok plan between August 12 and August 15, 2026 and put Grok 4.6 in its place, ticked for all seven tiers including Free. Re-read on October 2, 2026, neither SpaceXSI's pricing page nor grok.com's mode picker names Grok 4.5 anywhere. What has changed since is the model above it: SpaceXSI announced Grok 4.7 on September 21, 2026, and grok.com's picker now names Grok 4.7 behind the paid Expert and Heavy modes while the pricing page still lists Grok 4.6 for all seven plans. The free default, Fast, names no model on either surface. The rest of this page is the record of Grok 4.5 itself -- the July 8, 2026 launch by SpaceXSI (the company behind Grok, which brands itself SpaceXSI rather than xAI across its developer docs, its models page and grok.com itself, re-read October 12, 2026), the free-on-every-plan verdict that survived the swap, the prices, SpaceXSI's own benchmarks and the EU timeline -- because those are still the questions people type its name to ask.

Key Takeaways

  • Release date: July 8, 2026. Grok 4.5 shipped publicly that day, roughly six weeks after its original late-May target.
  • It was free on every plan, and it has since been replaced twice over. SpaceXSI's pricing page listed Grok 4.5 on every consumer plan, Free included, from at least July 13 until August 12, 2026. By August 15 that row read Grok 4.6, still ticked for Free, and on the newest capture we can read (October 10, 2026) it still does.
  • Grok 4.7 exists, and it is the first swap that may not be free. SpaceXSI announced it on September 21, 2026. On grok.com today the Expert mode reads "Thinks hard · Grok 4.7" and Heavy reads "Team of Experts · Grok 4.7", and both are marked as requiring an upgrade. The pricing page has no Grok 4.7 row at all.
  • "You pay for reasoning depth, not the model" held for two swaps and is now in doubt. Free gets Fast. Auto, Expert and Heavy require a subscription, as does video generation. That was a clean verdict while every mode ran the same model; now that Expert and Heavy name Grok 4.7 and Fast names nothing, the two Grok surfaces no longer agree.
  • SpaceXSI publishes five consumer prices, re-read live on November 4, 2026: Free at $0, Plus $8, Premium $30, Super $100 and Ultra $200, with Teams and Enterprise alongside them. The SuperGrok names this page once used return zero on the page.
  • Developers get it too. It is live on the SpaceXSI API as grok-4.5, metered at $2 per million input tokens and $6 per million output tokens, and in Cursor on all plans.
  • No cut-off is published for Grok 4.5, and as of October 12, 2026 none is published for Grok 4.6 either. The only cut-off SpaceXSI's models page now states is Grok 4.7's: "The knowledge cut-off date of Grok 4.7 is May 2026." The February 1, 2026 figure this page used to carry is no longer on that page. Live web and X search cover recent events; the model on its own does not.
  • Grok 4.5 is retired from the plans but not from the API. SpaceXSI's models page still lists grok-4.5 (aliases grok-4.5-latest and grok-build-latest, 500,000-token limit) at the same $2 and $6 per million tokens as Grok 4.6, read September 9, 2026.
  • The EU was excluded at launch, and only the API block has lifted. SpaceXSI's launch post said Grok 4.5 was "not yet available in the EU in any SpaceXSI products or the API console"; its changelog reopened the API console on July 17 and has said nothing about the consumer Grok apps.
  • SpaceXSI's own benchmarks put it third. Its launch charts show Claude Fable ahead of Grok 4.5 on four of the five evals SpaceXSI chose to publish -- but it is far cheaper and, by its own measure, about 4x more token-efficient.

Grok 4.5 Release Date: When It Actually Shipped

Grok 4.5 was released on July 8, 2026. That is the date SpaceXSI published the model and made it available to developers the same day; it landed in the consumer Grok apps immediately after. The confusion around this date is understandable, because the model was talked about publicly for weeks before anyone could touch it -- and it arrived about six weeks later than originally signalled.

DateWhat happened
Late May 2026Original expected launch window. Missed.
June 28, 2026Musk confirms Grok 4.5 exists, running in private beta at SpaceX and Tesla only. No public access.
July 8, 2026Public release. SpaceXSI's announcement says it is available "today in Grok Build, in Cursor on all plans, and from the SpaceXSI console" -- and not in the EU.
July 9, 2026Rollout reaches the consumer Grok apps.
By July 13, 2026SpaceXSI's pricing page lists Grok 4.5 as included on every plan, Free included. SpaceXSI never announced this as its own event, so the exact date the free tier flipped is not published -- July 13 is simply the earliest date we can prove it from the page itself.
July 17, 2026EU access opens on the API console, per SpaceXSI's release notes.

If you read a page saying Grok 4.5 has "no release date" or that you "cannot use it on any plan," that page is describing the private-beta period that ended on July 8.

Can You Still Use Grok 4.5? Availability by Plan

SpaceXSI has rebuilt its pricing page, and the plan names this site published are gone. Read live on October 29, 2026 -- the first live read, after the host turned out to answer 200 through a different HTTP client -- x.ai/pricing carries a new consumer lineup with two tiers that did not exist here before, including an $8 entry plan. The names SuperGrok, SuperGrok Lite, SuperGrok Plus and SuperGrok Heavy return zero on the whole page.

PlanPrice on the pageUsage, in SpaceXSI's own words
Free$0"Limited"
Plus$8/mo"Basic"
Premium$30/mo"4x Plus"
Super$100/mo"15x Plus"
Ultra$200/mo"50x Plus"
Teams$30+/user/mo"4-50x Plus"
EnterpriseCustom, contact sales"Pooled"

And the model row this page used to read has been removed. The matrix's feature rows now run Usage, Grok Chat, Image generation (Imagine), Video generation, Grok API access, Grok Build, Grok Bot, Cursor Agent and the team and enterprise rows. The strings Grok 4.5, Grok 4.6 and Grok 4.7 each return zero on the page, on a body read from the nav to the site-map footer. Until August 12, 2026 that row said Grok 4.5 and was ticked for every tier; from August 15 it said Grok 4.6; it is now not there at all.

One honest limit on what that read supports. It does not mean SpaceXSI has stopped saying which model a plan runs -- only that this page no longer does it, which is a statement about the page and not about the company. Re-read live on November 2, 2026, the strings Grok 4.5, Grok 4.6, Grok 4.7, Grok 5 and SuperGrok all still return zero, in the HTML as well as the stripped text.

The matrix itself is now decoded, which the earlier read could not do. Each cell is its own SVG rather than a table cell, so the ticks had to be walked positionally against the seven column headers. Done on November 2, 2026, here is what the free plan actually gets:

Row on SpaceXSI's matrixFreePlus $8Premium $30
UsageLimitedBasic4x Plus
Grok ChatLimitedYesYes
Image generation (Imagine)LimitedYesYes
Grok BuildLimitedYesYes
Video generationNoNoYes
Grok BotNoNoYes
Grok API accessNoYesYes
Cursor AgentYesYesYes
Premium checkmarkNoYesYes
Ad-free scrollingNoSome surfacesMost surfaces

So the free plan is throttled rather than stripped: chat, images and Build are all there and marked "Limited". The two things it does not have at all are video generation and Grok Bot, and video generation is missing from the $8 Plus tier too, which makes $30 Premium the first tier that generates video. X's own Premium tiers are still not on SpaceXSI's pricing page at all, and X's help pages are not publicly readable, so this page does not tell you what Grok on X costs. If that is your route in, the check below works regardless.

The 10-second check: open grok.com or the Grok app and look at the mode picker. The four modes are Auto, Fast, Expert and Heavy, with Fast set as the default and the only one available without a subscription. Read again on October 2, 2026, SpaceXSI's own page configuration now names Grok 4.7 behind Expert and Heavy, and still names no model at all behind Fast or Auto. Grok 4.5 does not appear on that page any more, and nor does Grok 4.6.

ModeModel named in the picker (October 2, 2026)Free?
FastNone namedYes -- this is the default
Auto ("chooses Fast or Expert")None namedNo -- needs a subscription
Expert ("thinks hard")Grok 4.7No -- needs a subscription
Heavy ("team of experts")Grok 4.7No -- needs a subscription

This is read from the configuration grok.com serves a logged-out visitor, so "needs a subscription" means "not offered to this session." A signed-in free account may be shown something different, and SpaceXSI publishes no plan-by-plan mode table to check it against.

What Grok 4.5 Costs

As a consumer, nothing -- and on the API, $2 per million input tokens and $6 per million output tokens. The consumer answer is still $0: the pricing page's single model row, now Grok 4.6, is ticked for the free plan in the same place Grok 4.5 used to hold, and Fast mode remains the free default. What is no longer certain is whether free and paid run the same model, because the app's picker attaches Grok 4.7 to the paid modes only. The token price only matters if you are a developer, but it is the clearest signal of how SpaceXSI is positioning this model, so it is worth seeing next to the competition.

ModelInput (per 1M tokens)Output (per 1M tokens)
Grok 4.5$2.00$6.00
Grok 4.3 (previous)$1.25$2.50
Anthropic Opus 5 (current Opus, July 24, 2026)$5.00$25.00

Grok 4.5 is priced at roughly a quarter of Anthropic's Opus tier on output while claiming comparable quality. That Opus price has not moved: Claude Opus 5, released July 24, 2026, kept the same $5/$25 as the Opus models before it. It is, however, more expensive than the Grok 4.3 it succeeds -- so it is an upgrade you pay for, not a free bump. Per SpaceXSI's model docs it carries a 500,000-token context window and supports configurable reasoning effort (low, medium, or high, defaulting to high).

What's New in Grok 4.5

The headline change is speed and efficiency, not size. SpaceXSI's own launch announcement says Grok 4.5 was trained alongside Cursor, the AI coding editor the company acquired, "on datasets spanning knowledge in coding, science, engineering, and math," across "tens of thousands of NVIDIA GB300 GPUs." Its own docs describe it as the model to use "for everything else, including code," and pitch it as the most intelligent and fastest option in the lineup.

The numbers SpaceXSI publishes are all about doing the same work with less. It says Grok 4.5 is served at 80 tokens per second and reaches "roughly 2x the token efficiency of comparable leading models, solving tasks in under half the number of steps." On its own token-efficiency chart, Grok 4.5 finishes an average SWE Bench Pro task using 15,954 output tokens against Opus 4.8's 67,020 -- about 4.2x fewer. Fewer tokens for the same answer is what makes a $2/$6 model cheap in practice rather than just on the price list. It is also now the default model in Grok Build, SpaceXSI's app-building tool.

What it does not know: SpaceXSI no longer publishes a cut-off for it, or for its replacement. Re-read October 12, 2026, SpaceXSI's models page states one cut-off and one only: "The knowledge cut-off date of Grok 4.7 is May 2026." The string "February 1, 2026" does not appear on it, so the figure this page carried for Grok 4.6 can no longer be sourced there, and none has ever been published for Grok 4.5. Either way, SpaceXSI states plainly that "Grok has no knowledge of current events or data beyond what was present in its training data" unless search tools are switched on. In the Grok app that matters less than it sounds, because real-time web and X search fill the gap -- but it is the reason Grok can confidently miss something that happened last month if you ask it cold.

For everyday users, the honest translation: this is a model tuned for coding and technical work that happens to also power your chatbot. If you use Grok to draft emails and settle arguments, you will notice speed more than genius.

Is It Actually Good?

It is credible but not a landslide -- and SpaceXSI's own charts say so. The launch announcement claims Grok 4.5 "exceeds comparable leading models," then publishes five benchmark charts in which it wins one. On the other four, Anthropic's Fable is ahead of it, and on three of those so is GPT-5.5. These are SpaceXSI's numbers, on SpaceXSI's page, chosen by SpaceXSI.

Benchmark (SpaceXSI's own charts)Grok 4.5Best score shownWhere Grok 4.5 places
SWE Marathon (resolution rate)29.0%Grok 4.51st
Terminal Bench 2.183.3%Fable (max) 84.3%3rd
DeepSWE 1.062.0%Fable (max) 66.1%3rd
SWE Bench Pro64.7%Fable (max) 80.4%3rd
DeepSWE 1.153%Fable (max) 70%4th

Source: SpaceXSI, "Introducing Grok 4.5," July 8, 2026. Competitor figures are SpaceXSI's, drawn from the other developers' published system cards and leaderboards.

That is a meaningfully different situation from the pre-launch hype, when the pitch was that internal tests put it above Opus-class models. Musk's own framing was more restrained -- Grok 4.5 as "roughly comparable to Opus 4.7, but much faster", as reported at launch. What shipped is a fast, cheap, near-frontier model, which is a real achievement and also not the crown. Independent third-party testing remains thin, so treat any "beats GPT-5.6" or "beats Claude" headline with suspicion. Anthropic's current flagship is Claude Fable 5 -- the model beating Grok 4.5 in four of SpaceXSI's own five charts.

Read that table the right way, though. Every one of those benchmarks measures software engineering, because that is what Grok 4.5 was built for. None of them tells you whether it writes a better birthday message than Gemini. For everyday chat, the gap between any of these models is far smaller than the gap between the benchmarks suggests.

Grok 4.5 in the EU: The Block Has Partly Lifted

The EU block lifted on July 17, 2026 -- for developers. SpaceXSI's launch post carried a one-line note at the bottom: "Grok 4.5 is not yet available in the EU in any SpaceXSI products or the API console. EU availability is expected in mid-July." No further explanation was given. Nine days later its release notes added a matching one-liner: "Grok 4.5 is now available in the API console for EU users."

That entry covers the API console specifically. SpaceXSI has published nothing about the consumer Grok apps or grok.com in Europe, so we cannot tell you that your EU Grok app now serves 4.5 -- only that the region-wide freeze the company described at launch is no longer absolute.

The practical advice for EU readers is the same as everywhere else, and it is cheap to follow: do not upgrade your plan hoping that money unlocks the model. Open the mode picker on the plan you already have and look. Everything in it is now Grok 4.6, on every plan, and Grok 4.5 is no longer offered anywhere.

What changed (July 22, 2026): this section previously said the EU was excluded "with access expected mid-July." SpaceXSI's changelog now records EU API access as live since July 17, so the section was rewritten and the timeline row corrected.

Should You Pay for It?

Do not pay to get the newest Grok yet, but this advice is weaker than it was. For Grok 4.5 and Grok 4.6 it was clean: the model sat on the free plan, and the paid tiers bought the same Grok allowed to think longer (Expert and Heavy), plus video generation and higher limits. Grok 4.7 breaks the pattern on one of the two surfaces -- grok.com names it behind the paid modes only, and the pricing page no longer names any model at all. Re-read live on November 4, 2026, Grok 4.5, Grok 4.6, Grok 4.7 and Grok 5 each return zero on x.ai/pricing, in the HTML as well as the text, so there is no longer a ticked row to read either way. The honest answer is that free Grok is still excellent and that SpaceXSI has stopped telling you on its pricing page which model you are getting. Try it for $0 before you decide.

Pay if you keep hitting the ceiling. The things a subscription genuinely unlocks are concrete: Expert and Heavy reasoning modes for hard multi-step problems, video generation, unrestricted real-time web and X search rather than the free tier's "Limited" version, and higher rate limits across the board. If free Grok keeps cutting you off mid-task, the $30/month Premium tier buys headroom rather than intelligence, and the $8 Plus tier is the cheapest step up.

Skip it if Grok is your third chatbot. Paying $30 for a secondary assistant is hard to justify when Gemini, ChatGPT and Claude all have capable free tiers and their $20 plans cover more ground. Keep free Grok installed for its X integration and spend the money elsewhere -- our guide to which AI chatbot you should pay for lays out that call.

Developers, get it if you already live in Cursor or on the API -- at $2/$6 per million tokens with claimed 2x token efficiency, Grok 4.5 is one of the cheapest ways to run a near-frontier model, and that is the use case it was explicitly built and trained for.

The bottom line: the practical answer to "do I get Grok 4.5?" is that you get its replacement, for the same $0. Grok 4.5 was a genuinely strong, unusually cheap model that SpaceXSI put on the free plan, and Grok 4.6 took its place across every consumer plan by August 15, 2026, Free included. Since September 21, 2026 there is a Grok 4.7 as well, and the two SpaceXSI surfaces disagree about who gets it: the pricing page still says Grok 4.6 for everyone, the app picker says Grok 4.7 for Expert and Heavy only. Open the app, leave it on Fast, and you are on the free model either way -- and the question worth answering, for free, is still whether you use it enough to want Expert mode.

Corrections and Updates

November 4, 2026: the October 29 rebuild correction reached this page's body and missed four other surfaces. An FAQ answer, a Key Takeaway, the "Should You Pay" opener and the closing paragraph were all still selling SuperGrok at $30 and SuperGrok Plus at $100, and the opener still said "the pricing page still has a single Grok 4.6 row ticked for Free". Re-read live today, x.ai/pricing (strip 8,271) lists Free $0, Plus $8, Premium $30, Super $100, Ultra $200 and returns zero for SuperGrok, Heavy, Expert, Grok 4.5, Grok 4.6, Grok 4.7 and Grok 5 in both the HTML and the stripped text. All four surfaces now carry the live lineup. This site's own Grok 4.6 explainer had it right since October 29; the FAQ and Key Takeaway are what AI answer engines quote, and they were the ones left wrong.

October 29, 2026: this page's plan table named four SuperGrok tiers and said the pricing page "still ticks Grok 4.6 for all seven plans." SpaceXSI has rebuilt that page and neither is true. Read live for the first time today -- x.ai answers 200 through Node's fetch, where every previous read of it on this site came from a Wayback capture -- x.ai/pricing lists Free $0, Plus $8, Premium $30, Super $100, Ultra $200, Teams $30+/user and Enterprise. SuperGrok, Heavy and Expert each return zero on the whole page, and so do Grok 4.5, Grok 4.6 and Grok 4.7: the model row is gone from the matrix, replaced by a Usage row reading Limited / Basic / 4x Plus / 15x Plus / 50x Plus / 4-50x Plus / Pooled. The summary and the plan table now carry the new lineup. The per-plan "what you do not get" column is withdrawn, not rewritten, because the new matrix was not decoded cell by cell. Grok 4.5's own retirement, dates, benchmarks and EU timeline are unaffected.

October 12, 2026: this page called the company SpaceXAI. Its own surfaces call it SpaceXSI. Re-read today, "SpaceXSI" appears 25 times on the release notes, 7 times on the models page and 15 times on grok.com, including in grok.com's own page description and schema ("Grok is an AI assistant built by SpaceXSI"). "SpaceXAI" appears zero times on all three. Every use of the name on this page was changed, including in the earlier correction entries below.

October 12, 2026: this page said February 1, 2026 is the published cut-off and that it belongs to Grok 4.6. The models page no longer carries that date at all; the only cut-off it states is Grok 4.7's May 2026. Recorded as a deletion at the source rather than a new figure: nothing on SpaceXSI's surfaces says Grok 4.6's cut-off changed, only that the number is no longer published there. One Key Takeaway and the knowledge section were corrected.

October 2, 2026: this page said the Grok you open today runs Grok 4.6, and SpaceXSI released a Grok 4.7 eleven days before this check. The title, summary and closing all said Grok 4.6 was SpaceXSI's current model with no successor. SpaceXSI announced Grok 4.7 on September 21, 2026 -- "SpaceXSI's most powerful model for coding and knowledge work" -- and on grok.com, read October 2, the Expert mode now reads "Thinks hard · Grok 4.7" and Heavy reads "Team of Experts · Grok 4.7", both marked as requiring an upgrade. Two things did not change and are worth stating so this correction is not read as more than it is: the pricing page's single model row, on the newest capture we can read (September 28, 2026), still says Grok 4.6 and is still ticked for all seven plans including Free, with no Grok 4.7 row; and Grok 4.7's own announcement names its availability as "Cursor and Grok Build ... the Grok API, third-party coding harnesses, and model routers and cloud platforms" and does not mention the consumer Grok apps at all. So the consumer picture is a disagreement between two SpaceXSI surfaces, and both readings are now on this page.

September 9, 2026: the February 1, 2026 cut-off is Grok 4.6's, not Grok 4.5's. This page cited SpaceXSI's models page for "Grok 4.5's knowledge cut-off." Re-read on September 9, 2026, that page states the cut-off for Grok 4.6 only and states none for Grok 4.5. The date itself is unchanged and the practical answer is unchanged, because Grok 4.6 is the model on every plan. Re-verified in the same read: grok-4.5 is still live on the API, listed alongside grok-4.6 at $2 and $6 per million tokens, so its retirement is from the consumer plans only.

September 2, 2026 (re-verified): still Grok 4.6, and still free.

August 28, 2026: Grok 4.5 has been retired from SpaceXSI's consumer plans, and this page said otherwise for thirteen days.

August 12, 2026: Grok 4.5 is no longer SpaceXSI's newest model. SpaceXSI released Grok 4.6 that day, and sentences calling 4.5 "the newest model" were corrected to say it was the model the app ran. (Superseded: Grok 4.6 replaced Grok 4.5 on every consumer plan, Free included, by August 15, 2026. See the August 28 entry above.)

July 28, 2026: this page previously told you Grok 4.5 runs on a "new V9 foundation at roughly 1.5 trillion parameters, about triple the size." That was wrong and came from a crypto news blog; SpaceXSI's launch post publishes no parameter count, architecture name or size comparison, so the claim was removed rather than re-sourced.

July 26, 2026: this page previously said no free Grok tier included Grok 4.5 and that paid consumer plans were still receiving it in stages.

Keep up without the jargon

Frequently Asked Questions

What is the Grok 4.5 release date?

July 8, 2026. SpaceXSI (the company behind Grok, formerly xAI) announced it that day and it went live immediately for developers on the API and in Cursor. It reached consumer Grok apps the following day, though not on every plan at once.

Can I use Grok 4.5 right now?

Not by name any more. SpaceXSI replaced the Grok 4.5 row with Grok 4.6 in August 2026, ticked on every plan including Free. Since September 21, 2026 there is also a Grok 4.7, which grok.com names behind the paid Expert and Heavy modes. The pricing page has not caught up.

Is Grok 4.5 free?

It was, and its successor is. SpaceXSI still ticks Grok 4.6 for the Free plan in the row Grok 4.5 used to occupy, checked on the October 10, 2026 capture. The newer rule is less clean: grok.com now names Grok 4.7 only behind the paid Expert and Heavy modes, so paying may now buy a newer model.

Is Grok 4.5 available in the EU?

The question has moved on, because Grok 4.5 is retired everywhere. While it was current, SpaceXSI's changelog opened the API console to EU users on July 17, 2026, after a July 8 launch without EU access, and published nothing about the consumer Grok apps in Europe. Open your own mode picker to see which model it names for you.

Is Grok 4.5 better than Claude or ChatGPT?

Not on SpaceXSI's own numbers. Its launch post published five coding benchmarks, and Claude's Fable model scores higher on four of them. Grok 4.5 wins on price and speed instead: $2 per million input tokens, 80 tokens per second, and about 4x fewer tokens per task.

Does Grok 4.5 know about recent events?

Only through search. SpaceXSI's models page no longer publishes a cut-off for Grok 4.5. Either way, SpaceXSI says Grok has no knowledge beyond its training data unless web or X search is switched on, so what the model was trained on matters less than whether search is enabled.

How much does Grok 4.5 cost?

Nothing on the free plan, then or now. The consumer tiers on x.ai/pricing, re-read live November 4, 2026, are Free $0, Plus $8, Premium $30, Super $100 and Ultra $200. The SuperGrok names are gone. Developers pay $2 per million input tokens and $6 per million output.

Get the plain-English AI brief

One email. What changed in AI and what it means for you.