The Gist: SpaceXAI made Grok 4.5 public on July 9, 2026, its new flagship model priced at $2.00 per million input tokens and $6.00 per million output tokens — roughly half what Anthropic charges for Claude Opus 4.8 ($5.00 and $25.00) and what OpenAI charges for GPT-5.6 Sol ($5.00 and $30.00). Independent testing from Artificial Analysis places Grok 4.5 at number 4 on its Intelligence Index with a score of 54, behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8, while it matches GPT-5.5 on coding agents (a score of 76) at a cost of $2.49 per task. The model ships with a 500,000-token context window and low, medium, and high reasoning modes — but it is blocked in the European Union under the EU AI Act's systemic-risk rules, putting a major flagship out of reach for an entire market at launch. SpaceXAI has framed this as a staggered rollout rather than a permanent ban, with EU availability signaled for around mid-July 2026.
Key Takeaways
- Half the token price is the whole story. Grok 4.5 lists at $2.00 per million input tokens and $6.00 per million output tokens, against $5.00 and $25.00 for Claude Opus 4.8 and $5.00 and $30.00 for GPT-5.6 Sol. The aggressive-pricing bet, not a benchmark crown, is the headline.
- Fourth, not first. On the independent Artificial Analysis Intelligence Index it scores 54, ranking number 4 behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8. Elon Musk's "Opus-class" description is a vendor claim, not an independent finding.
- Coding is its strongest suit. The Artificial Analysis Coding Agent Index puts Grok 4.5 at 76, roughly level with GPT-5.5, at $2.49 per task versus $11.80 for Claude Fable 5.
- There is a real accuracy caveat. Artificial Analysis reports 52 percent accuracy and a 54 percent hallucination rate on its AA-Omniscience test — a reason to keep the model away from unverified factual work.
- Europe is locked out — for now. Grok 4.5 is blocked in the EU under the AI Act's systemic-risk provisions, so an entire regulated market cannot access the flagship yet. SpaceXAI has signaled a staggered EU rollout expected around mid-July 2026, not a permanent ban.
What SpaceXAI Actually Shipped
SpaceXAI — the company formerly known as xAI, which folded into SpaceX earlier in 2026 and completed its corporate rebrand in July — announced Grok 4.5 on July 8, 2026 and made it public the following day. We covered the name change in xAI Is Officially SpaceXAI Now; the products kept the Grok brand, and Grok 4.5 is the first flagship to ship under the new corporate identity. It replaces Grok 4.3 as the company's top model, though both Grok 4.3 and Grok 4.20 remain available for teams that do not want to move.
The hard specifications come from the model card. According to SpaceXAI's Grok 4.5 documentation, the model is exposed through the API as grok-4.5, accepts text and image inputs and returns text, supports function calling and structured outputs, and offers three reasoning levels — low, medium, and high, with high as the default. It carries a 500,000-token context window, and the company lists rate limits of 150 requests per second and 50 million tokens per minute across the us-east-1 and us-west-2 regions. The full SpaceXAI model list now shows Grok 4.5 sitting above the older Grok models.
SpaceXAI says the model was trained with the help of Cursor and on Nvidia GB300 GPUs, and Musk described it publicly as "Opus-class, much faster." That phrasing matters: "Opus-class" is a comparison to Anthropic's top model, and it is the vendor's own framing rather than a third-party measurement. As the independent numbers below show, the two do not sit in the same tier on the public leaderboards. The neutral read is that SpaceXAI shipped a capable, cheap, coding-strong flagship — and made a specific marketing claim on top of it that the data does not yet support.
The Pricing Play: Half the Token Price of Its Rivals
Price is where Grok 4.5 makes its loudest argument. SpaceXAI lists it at $2.00 per million input tokens, $0.50 per million cached input tokens, and $6.00 per million output tokens. Set that next to the two models it is most often compared with and the gap is stark: Claude Opus 4.8 runs $5.00 and $25.00 per million tokens, and GPT-5.6 Sol runs $5.00 and $30.00. On output tokens — the expensive side of most real workloads — Grok 4.5 is roughly a quarter of the price of GPT-5.6 Sol.
| Model | Input (per million tokens) | Cached input (per million tokens) | Output (per million tokens) |
|---|---|---|---|
| Grok 4.5 | $2.00 | $0.50 | $6.00 |
| Claude Opus 4.8 | $5.00 | — | $25.00 |
| GPT-5.6 Sol | $5.00 | $0.50 | $30.00 |
All figures are in US dollars per million tokens. Grok 4.5 pricing is from SpaceXAI's documentation; the Opus 4.8 and GPT-5.6 Sol figures are their respective published list prices. If input, output, and cached pricing feel like a foreign language, our explainer on how input, output, and cached token pricing works breaks down why the output column tends to dominate a real bill.
Headline token prices only tell part of the story, so the more useful number is cost per completed task. On that measure, Artificial Analysis pegs Grok 4.5 at $2.49 to run its standard agentic evaluation, against $11.80 for Claude Fable 5 — the current index leader. A model that costs a fifth as much to finish the same job changes the math for anyone running high-volume, cost-sensitive pipelines, even if it is not the single most capable option on the board.
Where Grok 4.5 Actually Lands
Strip away the marketing and the independent picture is consistent: Grok 4.5 is a strong mid-pack flagship that punches hardest on code and price. On the Artificial Analysis Intelligence Index, published on July 8, 2026, it scores 54 and ranks number 4 — behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8. That is a respectable result, but it is not the top of the table, and it is a full tier below the "Opus-class" label the vendor attached to it.
Coding is the bright spot. Artificial Analysis gives Grok 4.5 a Coding Agent Index score of 76, which puts it roughly level with GPT-5.5 — a genuinely useful position for a model that costs far less to run. For agentic coding work, where a model chews through large numbers of tokens across many steps, pairing GPT-5.5-level coding output with Grok 4.5-level pricing is the combination that will get engineering teams to run a bake-off.
The caveat is accuracy. On the AA-Omniscience test, Artificial Analysis records a score of 26 for Grok 4.5, with 52 percent accuracy and a 54 percent hallucination rate. In plain terms, when the model is asked hard factual questions it gets a little over half right and confidently invents answers more than half the time it is wrong. That is a well-known trade-off for fast, cheap models, and it is the strongest argument for keeping Grok 4.5 inside workflows that verify their own outputs — code that runs, retrieval that cites sources — rather than open-ended factual research. For context on where the frontier sits, our reviews of Claude Fable 5 and Claude Opus 4.8 cover the models ranked above it.
One honest gap: as of July 11, 2026, Grok 4.5 has not yet appeared on the independent LMArena Elo leaderboard or on the independent SWE-bench Verified results tracked by third parties. Those numbers may arrive in the coming weeks, and when they do they will either reinforce or complicate the Artificial Analysis picture. Until then, the Intelligence Index and Coding Agent Index are the most complete independent reads available, and both point to the same conclusion: fast, cheap, coding-strong, not frontier-leading.
Blocked in the EU: A Flagship a Whole Market Cannot Touch
The most consequential fact about Grok 4.5 has nothing to do with benchmarks. At launch, the model is not available in the European Union, because it falls under the EU AI Act's rules for general-purpose AI models that carry what the regulation calls "systemic risk." This is a factual, regulatory outcome, and it is worth stating plainly and without spin: a major flagship model shipped this week, and roughly 450 million people in a single regulated market cannot legally access it through official channels. That block is real today, but it is expected to be temporary: SpaceXAI’s launch communications point to EU availability arriving around mid-July 2026, once Grok 4.5 clears the AI Act’s model-evaluation and systemic-risk requirements. In other words, this is a staggered rollout, not a permanent exclusion.
The AI Act, whose full text sits on EUR-Lex, sets extra obligations for the most capable general-purpose models — model evaluations, systemic-risk assessments, incident reporting, and cybersecurity measures — and the European Commission's regulatory framework for AI lays out how those tiers work in practice. Independent trackers such as the AI Act explorer map the same provisions for anyone who wants to read the primary rules rather than a summary. We have followed the Act's shifting timeline closely, including the moment the EU pushed its high-risk rules to December 2027.
Whatever one thinks of the policy, the market effect is concrete. European developers who want a frontier-adjacent model at Grok 4.5 pricing will reach for an available alternative — an Anthropic, OpenAI, Google, or Mistral model that has cleared the EU's requirements — rather than route around the block. For SpaceXAI, the choice not to offer Grok 4.5 in the EU at launch trades access to one of the world's largest developer markets for a faster, lighter-touch rollout everywhere else. For buyers outside the EU, it is a non-issue; for teams with European users or data-residency obligations, it is the single most important line in the launch.
Two Flagships, One Week, Opposite Bets
Timing turned Grok 4.5 into half of a story it did not write alone. On the same day it went public, July 9, 2026, OpenAI moved its GPT-5.6 family — Sol, Terra, and Luna — to general availability, according to OpenAI's announcement and its developer documentation. Two frontier labs shipped flagship models in the same week, and they made almost opposite bets.
OpenAI's bet is capability at a premium. GPT-5.6 Sol sits at the top of the current benchmarks and is priced to match, at $5.00 and $30.00 per million tokens, with a positioning aimed at the hardest coding, agentic, and scientific work. We wrote about the unusual way OpenAI staged that rollout in OpenAI Launched GPT-5.6 Sol — Then Let Washington Decide Who Gets It. SpaceXAI's bet is the mirror image: accept a number 4 ranking and lead on price, betting that "good enough, much cheaper, and strong at code" wins a large slice of real production traffic.
Both bets can be right at once, because they target different buyers. The team building a high-stakes research agent that cannot afford a wrong answer will pay for Sol; the team running millions of routine coding or automation calls will look hard at Grok 4.5's $2.49 per task. What the same-week launches really show is that the frontier has split into two competitions — one for the top of the leaderboard, and one for the best capability you can buy per dollar — and Grok 4.5 is aimed squarely at the second.
What It Means for Builders
For anyone choosing a model this month, Grok 4.5 is a clear candidate for one job and a clear pass for another. It is a strong fit for high-volume, cost-sensitive workloads that lean on code and automation, where its GPT-5.5-level Coding Agent score and $2.49 cost per task line up against far pricier rivals. It is a weaker fit for open-ended factual work, given the 52 percent accuracy and 54 percent hallucination figures, and it is a non-starter for teams that must serve EU users through official, compliant channels.
The sober summary is that Grok 4.5 does not change who leads the frontier — Claude Fable 5, GPT-5.5, and Claude Opus 4.8 still sit above it on the independent index — but it does change the price of admission to strong, coding-capable inference. In a market where the top models keep getting more expensive, a credible flagship at half the token price and a fifth of the cost per task is a real competitive event, even from number 4. The asterisk is geographic: for a large, wealthy market, that event is happening behind glass.
Frequently Asked Questions
What is Grok 4.5?
Grok 4.5 is SpaceXAI's flagship AI model, made public on July 9, 2026. It is priced at $2.00 per million input tokens and $6.00 per million output tokens, carries a 500,000-token context window, and ranks number 4 with a score of 54 on the independent Artificial Analysis Intelligence Index. It replaces Grok 4.3 as the company's top model.
How much does Grok 4.5 cost?
SpaceXAI lists Grok 4.5 at $2.00 per million input tokens, $0.50 per million cached input tokens, and $6.00 per million output tokens. That is roughly half the price of Claude Opus 4.8 ($5.00 and $25.00 per million tokens) and GPT-5.6 Sol ($5.00 and $30.00 per million tokens). Artificial Analysis measures its cost at $2.49 per completed task.
Is Grok 4.5 better than Claude Opus 4.8?
On the independent Artificial Analysis Intelligence Index, Claude Opus 4.8 ranks above Grok 4.5, which sits at number 4 with a score of 54. Grok 4.5 is the cheaper model and matches GPT-5.5 on coding agents. The right choice depends on your priority: top accuracy points to Opus 4.8, while price and coding value point to Grok 4.5.
Is Grok 4.5 available in the European Union?
Not at launch. Grok 4.5 is blocked in the European Union under the EU AI Act, which treats it as a general-purpose AI model with systemic risk, and SpaceXAI has not made it available to EU users through official channels yet. This is a staggered rollout rather than a permanent ban: SpaceXAI has signaled EU availability is expected around mid-July 2026, once the model clears the Act’s evaluation and systemic-risk requirements. Until then, teams serving European users need a compliant alternative.
Why is Grok 4.5 blocked in the EU?
The EU AI Act places additional obligations — such as model evaluations, systemic-risk assessments, and incident reporting — on the most capable general-purpose AI models. Grok 4.5 falls into that category, and it is not being offered in the EU at launch. This is a regulatory outcome, not a comment on the model's quality.
How does Grok 4.5 compare to GPT-5.6?
Both went public in the same week, on July 9, 2026, but they take opposite approaches. GPT-5.6 Sol is a premium-priced flagship ($5.00 and $30.00 per million tokens) that leads current benchmarks, while Grok 4.5 is aggressively priced ($2.00 and $6.00 per million tokens) and ranks number 4. GPT-5.6 Sol bets on capability; Grok 4.5 bets on price.
What is Grok 4.5's context window?
Grok 4.5 has a 500,000-token context window, according to SpaceXAI's documentation. It accepts text and image inputs and returns text, supports function calling and structured outputs, and offers low, medium, and high reasoning modes, with high as the default.
Is Grok 4.5 good at coding?
Yes, coding is its strongest area. Artificial Analysis gives Grok 4.5 a Coding Agent Index score of 76, roughly level with GPT-5.5, at a cost of $2.49 per task. For high-volume agentic coding work, that pairing of GPT-5.5-level output and low pricing is its main selling point.
Does Grok 4.5 hallucinate?
On the AA-Omniscience test from Artificial Analysis, Grok 4.5 scores 26, with 52 percent accuracy and a 54 percent hallucination rate. That means it invents answers on hard factual questions more often than not when it is wrong, so it is best used inside workflows that verify their own outputs rather than for open-ended factual research.
Did Grok 4.5 replace Grok 4.3?
Yes. Grok 4.5 replaces Grok 4.3 as SpaceXAI's flagship model. However, both Grok 4.3 and Grok 4.20 remain available through the API for teams that prefer to stay on an older model rather than migrate immediately.
Is Grok 4.5 really "Opus-class"?
"Opus-class, much faster" is Elon Musk's own description, so it is a vendor claim rather than an independent measurement. Independent data from Artificial Analysis places Grok 4.5 at number 4 on the Intelligence Index, a tier below Claude Opus 4.8. It has not been independently confirmed as Opus-class.
What reasoning modes does Grok 4.5 support?
Grok 4.5 supports three reasoning levels — low, medium, and high — with high as the default, according to SpaceXAI's documentation. The model is exposed as grok-4.5 in the API, with rate limits of 150 requests per second and 50 million tokens per minute across the us-east-1 and us-west-2 regions.
Sources
- SpaceXAI — Grok 4.5 model documentation (pricing, context, reasoning, rate limits)
- SpaceXAI — Models overview
- Artificial Analysis — Intelligence Index, Coding Agent Index, and cost-per-task methodology
- Artificial Analysis — Model leaderboards
- OpenAI — GPT-5.6 (Sol, Terra, Luna) announcement
- OpenAI — Model and pricing documentation
- EUR-Lex — Regulation (EU) 2024/1689 (the AI Act), full text
- European Commission — Regulatory framework for AI
- The AI Act Explorer — Provisions on general-purpose AI and systemic risk
Published by Anthony Martinez, ThePlanetTools.ai. This article is editorial analysis of publicly reported information. ThePlanetTools.ai has no affiliation with SpaceXAI, OpenAI, or Anthropic. Benchmark figures are attributed to Artificial Analysis; pricing and specifications are from SpaceXAI's official documentation.
