Grok 4.5

Grok 4.5 is SpaceXAI's flagship reasoning model, publicly available since July 9, 2026. It costs $2 per million input tokens and $6 per million output tokens — roughly half the sticker price of Claude Opus 4.8 and GPT-5.6 Sol — with a 500,000-token context window, text and image input, function calling, and structured outputs. Elon Musk describes it as "Opus-class, much faster." We rate it 8.7 out of 10 after early hands-on testing on the xAI API.

Quick Verdict

We ran Grok 4.5 on the xAI API two days after its public release, across four representative tasks — non-trivial coding, multi-step reasoning, structured JSON output, and strict instruction-following. It handled all four correctly on the first attempt, and it felt genuinely fast for a reasoning model, returning answers in under five seconds each. The headline is the price-to-capability ratio: at $2 per million input tokens and $6 per million output tokens, it is about half the price of Opus 4.8 and GPT-5.6 Sol, and independent testing from Artificial Analysis puts its cost per task near the bottom of the frontier tier. It is not the smartest model of July 2026 — it sits fourth on the Artificial Analysis Intelligence Index — and its measured hallucination rate is a real caveat. This is an early verdict on a model that is two days old.

• Score: 8.7 out of 10 — the strongest value play at the frontier tier in July 2026.
• Pro: $2 input and $6 output per million tokens — around half the price of Opus 4.8 and GPT-5.6 Sol.
• Pro: Fast, correct output in our early hands-on tests, with reliable JSON mode and exact instruction-following.
• Con: Ranks fourth on the Artificial Analysis Intelligence Index, behind Claude Fable 5, GPT-5.5, and Opus 4.8.
• Con: Blocked in the European Union, and its independently measured hallucination rate is high.

What Is Grok 4.5?

Grok 4.5 is the flagship large language model from SpaceXAI — the renamed AI unit that emerged after xAI was folded into SpaceX, with the "SpaceXAI" branding taking effect on July 6, 2026. The product keeps the Grok name. Grok 4.5 was announced on July 8 and made publicly available on July 9, 2026, replacing Grok 4.3 as the flagship while Grok 4.3 and Grok 4.20 remain available.

The API model identifier is grok-4.5, with the aliases grok-4.5-latest and grok-build-latest. It is served through an OpenAI-compatible endpoint at api.x.ai, which means most existing SDK code runs against it with only a base-URL and model-name change. According to xAI's model documentation, Grok 4.5 offers a 500,000-token context window, accepts text and image input and returns text, and supports function calling and structured outputs. xAI says the model was trained jointly with the team at Cursor on GB300-class GPUs.

The positioning comes straight from Elon Musk, who described Grok 4.5 as "Opus-class, much faster" and, in a separate framing, "roughly comparable to Opus 4.7, but much faster." That framing sits alongside the vendor material on x.ai, but it is a vendor claim, not an independent verdict — but our own early testing and the independent numbers from Artificial Analysis both point to a model that competes near the top on capability while undercutting the field badly on price. For the background on the corporate reshuffle, we covered the SpaceXAI rebrand and the earlier dissolution of xAI into SpaceX when they happened.

One thing to be clear about up front: several widely repeated claims about Grok 4.5 are not confirmed by primary sources. Reported parameter counts (a "1.5 trillion parameter" or "V9" figure circulating from a leak), claims that the grok-4.5 API tier ships with always-on real-time X data, and pricing for server-side tools are all unconfirmed against xAI's own documentation as of publication. We do not treat any of them as fact in this review, and neither should you.

Key Features in 2026

Grok 4.5's feature set is a deliberate frontier-tier package with one obvious gap. Everything below is drawn from xAI's model documentation at docs.x.ai, cross-checked against our own API calls.

• 500,000-token context window — roughly 1,000 pages of text in a single prompt. Large, though smaller than Grok 4.3's stated 1 million tokens and GPT-5.6's 1.05 million.
• Text and image input, text output — Grok 4.5 accepts multimodal input (text plus images) but returns text only. It does not generate images natively; that job stays with the separate Grok Imagine model.
• Function calling — supported, so the model can drive tools and agent loops through the standard OpenAI-compatible tool interface.
• Structured outputs — supported. In our tests JSON mode returned valid, schema-adherent objects every time.
• Reasoning effort control — the model exposes low, medium, and high reasoning effort, with high as the default. Our calls showed dedicated reasoning tokens in the usage payload, confirming an internal chain-of-thought pass.
• High rate limits — 150 requests per second and 50 million tokens per minute per the documentation, which is generous headroom for production agent fleets.
• Two regions — us-east-1 and us-west-2. There is no EU region, which ties into the availability restriction discussed below.
• OpenAI-compatible API — the endpoint mirrors the OpenAI Chat Completions shape, so migration from another provider is largely a configuration change.
• Prompt caching — cached input is billed at $0.50 per million tokens, and our usage payloads showed cached tokens being counted, which compounds savings on agent loops with stable system prompts.

If you are new to how these API knobs translate into a monthly bill, our explainer on input, output, and cached token pricing walks through the math, and our primer on agentic coding models versus chatbots explains why function calling and reasoning effort matter for real work.

Pricing in 2026

Grok 4.5 costs $2.00 per million input tokens, $0.50 per million cached input tokens, and $6.00 per million output tokens, according to xAI's model documentation. That is roughly half the sticker price of Claude Opus 4.8 at $5 input and $25 output, and GPT-5.6 Sol at $5 input and $30 output per million tokens.

The comparison that matters most is against the models Grok 4.5 is actually pitched against. On output tokens — where reasoning models spend the bulk of the bill because internal reasoning tokens are billed as output — Grok 4.5 is four to five times cheaper than Opus 4.8 and GPT-5.6 Sol. Note that the older Grok 4.3 is still cheaper on paper at $1.25 input and $2.50 output; Grok 4.5 charges a premium over its predecessor for a meaningful capability step up, which is a reasonable trade if you need the extra intelligence.

The cost story is even stronger once you factor in real workloads. Because pricing is per token and a reasoning model's output includes its chain-of-thought, the practical cost per completed task is what counts. Independent measurement from Artificial Analysis puts Grok 4.5 at about $2.49 per task on its evaluation suite, against $11.80 for the far pricier Claude Fable 5 — a roughly five-fold difference. To keep prose honest we quote pricing in plain units: two dollars per million input tokens, six dollars per million output tokens. There is no free API tier for the flagship; the cheapest route to programmatic access is the metered API itself.

Hands-On Testing

We accessed Grok 4.5 through the xAI API using the grok-4.5 model identifier and the OpenAI-compatible Chat Completions endpoint documented at docs.x.ai, two days after its public release. This is early hands-on work: a small, four-task battery run once, not a matured verdict. We are reporting exactly what we saw.

Non-trivial coding. We asked for a TypeScript function to merge overlapping and adjacent intervals, calling out the tricky edge cases — empty input, fully nested intervals, and intervals that only touch at an endpoint. Grok 4.5 returned a clean, correct implementation in about four seconds. It sorted a copy rather than mutating the caller's array, and it used a single comparison to catch both overlapping and touching intervals:

if (cur[0] <= last[1]) last[1] = Math.max(last[1], cur[1]);

That <= is the detail junior implementations get wrong — a strict less-than would fail to merge [1,3] and [3,5]. The model also appended an accurate complexity note. This was correct on the first try, with no hand-holding.

Multi-step reasoning. We gave it a two-train closing-distance word problem with a staggered departure, asking for the meeting clock time and the distance from the origin station. Grok 4.5 correctly accounted for the 30-minute head start, computed the meeting at 12:06 and 186 kilometers from station A, and — the part we liked — verified its own answer with an independent second calculation. It finished in roughly four seconds.

Structured output. We requested strict JSON extraction from a sentence, with a fixed set of keys and types. The model returned valid JSON on every run, adhered to the requested schema, and correctly excluded the EU from the list of available regions because the source sentence stated the model was blocked there. That kind of careful reading of an implicit constraint is exactly what you want from a model driving downstream automation.

Instruction-following over a longer input. We fed an 18-section synthetic report and asked for the three highest-throughput subsystems, in an exact three-line format, sorted descending, with no commentary. Grok 4.5 returned precisely three lines, in the right order, in the requested format, and added nothing else. It was also the fastest task in our set, under two seconds.

Two behaviors stood out across the battery. First, speed: every response came back in roughly two to four seconds, which is fast for a reasoning model and is consistent with the "much faster" half of Musk's positioning. Second, the usage payloads exposed dedicated reasoning tokens and counted cached tokens, confirming both the internal chain-of-thought and working prompt caching. We did not stress-test the full 500,000-token context in this early pass, and we would want a much larger sample before drawing conclusions about factual reliability — a point the independent benchmarks below make pressing.

Independent Benchmarks

According to Artificial Analysis, which published independent Grok 4.5 results on July 8, 2026, the model scores 54 on the Intelligence Index — fourth place, behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8. It scores 76 on the Coding Agent Index, roughly level with GPT-5.5, at an estimated cost of $2.49 per task.

The coding number is the more flattering of the two. A Coding Agent Index of 76 puts Grok 4.5 in the same neighborhood as GPT-5.5 on agentic coding, which — paired with the price — is a strong practical position for developer workloads. The Intelligence Index placement is more sobering: fourth at 54 means Grok 4.5 is a very good model, not the best available, and it trails the current leaders by a clear margin.

The caveat we weight most heavily is factual reliability. On the AA-Omniscience benchmark, Artificial Analysis reports an overall score of 26, an accuracy of 52 percent, and a hallucination rate of 54 percent. A hallucination rate above half on that evaluation is high, and it lines up with the standard advice for any Grok model: pair it with retrieval, citations, or human review for anything factual, and do not trust it blind on edge cases.

Two data gaps deserve to be stated plainly, because their absence is itself information. As of July 11, 2026, Grok 4.5 is not yet listed on LMArena, and it has no independently verified SWE-bench Verified score. Those are two of the most-watched public leaderboards, and until Grok 4.5 appears on them, any head-to-head coding or preference ranking against Opus 4.8, GPT-5.5, or Claude Fable 5 rests on a single independent source plus vendor framing. We will update this review when that data lands. For the deeper background on why cost-per-task has become the metric that moves budgets, Artificial Analysis publishes its methodology openly at artificialanalysis.ai.

Pros and Cons

Pros

• About half the token price of Claude Opus 4.8 and GPT-5.6 Sol at $2 input and $6 output per million tokens.
• Very low cost per task — roughly $2.49 on the Artificial Analysis suite, near the bottom of the frontier tier.
• Fast, correct output in our early hands-on tests, with responses under five seconds on every task.
• Reliable structured outputs — valid, schema-adherent JSON on every run in our testing.
• OpenAI-compatible API means most existing SDK code runs with only a base-URL and model-name change.
• Strong agentic coding position — a Coding Agent Index of 76 from Artificial Analysis, roughly level with GPT-5.5.
• Generous rate limits at 150 requests per second and 50 million tokens per minute per xAI documentation.

Cons

• Ranks fourth on the Artificial Analysis Intelligence Index — a very good model, not the smartest of July 2026.
• High measured hallucination rate — 54 percent on AA-Omniscience, so it needs retrieval or human review for factual work.
• Blocked in the European Union under the EU AI Act's systemic-risk provisions.
• No independently verified SWE-bench Verified score and no LMArena listing as of July 11, 2026.
• Smaller 500,000-token context window than Grok 4.3's stated 1 million and GPT-5.6's 1.05 million.
• No native image generation — multimodal input only, with text output.

Best Use Cases

• High-volume agent workflows. When token cost dominates unit economics — support agents, RAG pipelines, lead qualification — the four-to-five-fold price gap versus Opus 4.8 is decisive at scale.
• Agentic coding. A Coding Agent Index of 76 plus low cost per task makes Grok 4.5 a strong pick for bug hunting, refactors, and scaffold generation, especially in tool-driven loops.
• Cost-sensitive reasoning. Multi-step logic and analysis where you want a reasoning model but cannot justify frontier-leader prices on every call.
• Structured data extraction. JSON mode was reliable in our tests, which suits classification, entity extraction, and schema-constrained automation.
• Drop-in provider swaps. Teams already on an OpenAI-compatible stack can A/B Grok 4.5 against their current model with minimal code change.
• Latency-sensitive products. Sub-five-second responses in our testing make it viable for interactive assistants where a slower reasoning model would frustrate users.

One workload to approach with care: factual, safety-critical, or regulated tasks. Given the 54 percent hallucination rate on AA-Omniscience, we would keep medical, legal, and compliance workflows on a model with a stronger factual-precision record, or gate Grok 4.5 behind a verified retrieval layer. And because the model is unavailable in the EU, any product serving European users needs a different primary model or a compliant fallback.

Grok 4.5 vs. Alternatives

Grok 4.5's competitive case is almost entirely about price-to-capability. Here is how it lines up against the models buyers actually cross-shop in July 2026.

Pick Claude Opus 4.8 when maximum reasoning accuracy justifies paying four to five times more per output token — Musk's own "Opus-class" framing concedes that Opus is the reference point, not the one being beaten. Pick GPT-5.5 if you live in the OpenAI ecosystem or want a longer independent track record. Pick Claude Fable 5 if you need the current Intelligence Index leader and cost is not the constraint. Consider Gemini 3.1 Pro if context length and Google integration matter most. Stay on Grok 4.3 only if its lower sticker price outweighs the capability gap for your workload. Choose Grok 4.5 when you want frontier-adjacent output at roughly half the price of the leaders, and can manage its factual-reliability caveat.

EU Availability and Access

Grok 4.5 is blocked in the European Union. SpaceXAI has not made the model available to EU users, a restriction attributed to the systemic-risk obligations the EU AI Act places on the most capable general-purpose models. The model is served only from the us-east-1 and us-west-2 regions, with no EU region.

For teams outside the EU, access is straightforward: the metered xAI API through the OpenAI-compatible endpoint, plus the consumer Grok app and SuperGrok subscription tiers. For teams inside the EU, Grok 4.5 is not a legally available option today, and you should plan around a compliant primary model. The EU AI Act's obligations for general-purpose AI with systemic risk are documented on the EU AI Act reference site and the European Commission's regulatory framework pages. This is the single biggest availability caveat for the model, and it is not a temporary rollout delay — it is a regulatory boundary.

Frequently Asked Questions

What is Grok 4.5?

Grok 4.5 is SpaceXAI's flagship large language model, announced July 8, 2026 and publicly available July 9, 2026. It replaces Grok 4.3 as the flagship while Grok 4.3 and Grok 4.20 remain available. It offers a 500,000-token context window, text and image input with text output, function calling, and structured outputs, and is served through an OpenAI-compatible API.

How much does Grok 4.5 cost?

According to xAI documentation, Grok 4.5 costs $2 per million input tokens, $0.50 per million cached input tokens, and $6 per million output tokens. That is roughly half the sticker price of Claude Opus 4.8 at $5 input and $25 output, and GPT-5.6 Sol at $5 input and $30 output per million tokens. There is no free API tier for the flagship.

Is Grok 4.5 better than Claude Opus 4.8?

Not on raw intelligence. Artificial Analysis ranks Grok 4.5 fourth on its Intelligence Index at 54, behind Claude Opus 4.8. Elon Musk describes Grok 4.5 as "Opus-class, much faster," which frames Opus as the reference point rather than the model being beaten. Grok 4.5's advantage is price and speed, not top-end accuracy.

What is Grok 4.5's context window?

Grok 4.5 has a 500,000-token context window per xAI documentation — roughly 1,000 pages of text. That is smaller than Grok 4.3's stated 1 million tokens and GPT-5.6's 1.05 million tokens, but still large enough for long-document analysis and multi-file coding tasks.

Why is Grok 4.5 blocked in the EU?

Grok 4.5 is not available to European Union users, a restriction attributed to the systemic-risk obligations the EU AI Act places on the most capable general-purpose models. The model is served only from the us-east-1 and us-west-2 regions, with no EU region. This is a regulatory boundary, not a temporary rollout delay.

Does Grok 4.5 hallucinate?

Yes, at a rate worth planning around. Artificial Analysis reports a 54 percent hallucination rate and 52 percent accuracy on its AA-Omniscience benchmark. For factual, safety-critical, or regulated work, pair Grok 4.5 with retrieval, citations, or human review rather than trusting it blind on edge-case facts.

Who makes Grok 4.5?

Grok 4.5 is made by SpaceXAI, the renamed AI unit that emerged after xAI was folded into SpaceX, with the SpaceXAI branding taking effect July 6, 2026. The product keeps the Grok name. xAI says the model was trained jointly with the team at Cursor on GB300-class GPUs.

How do I access the Grok 4.5 API?

Grok 4.5 is served through an OpenAI-compatible endpoint at api.x.ai using the model identifier grok-4.5, with the aliases grok-4.5-latest and grok-build-latest. Because the API mirrors the OpenAI Chat Completions shape, migrating from another provider is largely a base-URL and model-name change. Access is limited to non-EU users.

Is Grok 4.5 on LMArena or SWE-bench Verified?

No. As of July 11, 2026, Grok 4.5 is not yet listed on LMArena and has no independently verified SWE-bench Verified score. Its independent standing rests mainly on Artificial Analysis results published July 8, plus vendor framing. We will update this review when the missing leaderboard data lands.

Does Grok 4.5 generate images?

No. Grok 4.5 accepts text and image input but returns text only — it does not generate images natively. Image generation is handled by the separate Grok Imagine model, not the Grok 4.5 chat and API model.

How fast is Grok 4.5?

In our early hands-on testing on the xAI API, every response came back in roughly two to four seconds across four tasks, which is fast for a reasoning model. That is consistent with the "much faster" half of Elon Musk's positioning, though we tested a small sample two days after release rather than at production scale.

Should I switch from Grok 4.3 to Grok 4.5?

Switch if you need the capability step up and can absorb the higher token price — Grok 4.5 charges $2 input and $6 output per million tokens versus Grok 4.3 at $1.25 and $2.50. Stay on Grok 4.3 if its lower sticker price outweighs the intelligence gain for your workload, or if you need its larger stated 1 million token context.

Final Verdict

Grok 4.5 is the most aggressive value play at the frontier tier in July 2026. It is not the smartest model available — Artificial Analysis puts it fourth on the Intelligence Index — and its measured hallucination rate is a genuine caveat for factual work. But at roughly half the token price of Opus 4.8 and GPT-5.6 Sol, with a cost per task near the bottom of the field and a strong agentic coding score, it changes the math for any workload where token cost dominates the bill. In our early hands-on tests it was fast and correct across coding, reasoning, structured output, and instruction-following, which matches the "Opus-class, much faster" framing on speed even if the independent intelligence ranking does not fully back the "Opus-class" half.

Our recommendation is a hybrid: route high-volume agent traffic, agentic coding, and cost-sensitive reasoning to Grok 4.5, and keep factual, safety-critical, and EU-facing workloads on a model with a stronger reliability record and legal availability. That is the same split we would run ourselves. We rate Grok 4.5 8.7 out of 10 — the score is driven by value and speed, and held back from higher by the fourth-place intelligence ranking, the hallucination rate, the EU block, and the thin independent data on a model that is only two days old. We will revisit this verdict once Grok 4.5 appears on LMArena and picks up an independent SWE-bench Verified score.

Sources

• xAI model documentation — grok-4.5 (specifications, pricing, rate limits, regions)
• SpaceXAI / x.ai (vendor, product positioning)
• Artificial Analysis (independent Intelligence Index, Coding Agent Index, cost per task, AA-Omniscience — published July 8, 2026)
• EU AI Act and the European Commission AI regulatory framework (systemic-risk obligations)