What Is Grok 4.5? The V9 Flagship That Sold Token Economy Instead of Top Scores

Grok 4.5 is a 1.5-trillion-parameter mixture-of-experts reasoning model from the vendor — the company that now ships and markets under the SpaceXAI brand — released on July 8, 2026. It takes text and images in, returns text, carries a 500,000-token context window, and lists at $2.00 per million input tokens and $6.00 per million output. It was the model that put "tokens per finished task" on the front of a launch slide, and it is the model the lab has since built two newer flagships on top of.

That last clause is the whole reason this page needs writing. Search "Grok 4.5" and you get launch coverage from July: the Cursor partnership, the "Opus-class" quote, the 80-tokens-per-second claim. What you do not get is a straight answer to the question a developer actually has in late September, which is whether Grok 4.5 is still worth wiring into anything when Grok 4.6 and Grok 4.7 exist at the same list price. So this page is a reference entry, not a news piece: what the model is, where it sits, what it measurably does well, what it measurably does badly, and who should pick something else. Nothing here is framed as a launch, because it is not one.

On sourcing, since it matters more than usual on a page like this: every figure below is labelled by where it came from. Numbers that xAI published about its own model are marked vendor-reported. Numbers from Artificial Analysis are independent measurements, read directly from the model page on September 23, 2026. Where a number could not be confirmed on a primary source, the page says so instead of estimating.

Where Grok 4.5 sits in its own family — and it is not the top

The single most common mistake in Grok 4.5 coverage is treating it as the flagship. It was, for about five weeks. Since then xAI has shipped Grok 4.6 on August 12, 2026 and Grok 4.7 in early September, both of which are still priced at $2.00 and $6.00 per million tokens. Grok 4.5 is now the middle of the stack: more capable than the cheap tier, two generations behind the front, and — this is the part that matters — the only model in the current Grok line whose cached-input rate is cheaper than its successors'. Grok 4.5 caches at $0.30 per million tokens; Grok 4.6 and Grok 4.7 both cache at $0.50, according to xAI's own pricing table.

Here is the lineup as the vendor's pricing documentation currently lists it, all figures per million tokens and all vendor-published:

• Grok 4.20 (three variants: multi-agent, reasoning, non-reasoning) — 1M context, $1.25 input / $2.50 output, $0.20 cached, shipped March 31, 2026

• Grok 4.3 — 1M context, $1.25 input / $2.50 output, $0.20 cached, plus a batch discount; the cheapest way to get a million tokens of Grok context

• Grok 4.5 — 500K context, $2.00 input / $6.00 output, $0.30 cached, released July 8, 2026

• Grok 4.6 — 500K context, $2.00 input / $6.00 output, $0.50 cached, released August 12, 2026

• Grok 4.7 — 500K context, $2.00 input / $6.00 output, $0.50 cached, released early September 2026

• Grok Build 0.1 — 256K context, $1.00 input / $2.00 output, a coding specialist rather than a general model

Two things fall out of that list. First, Grok 4.5 did not get a price cut when it stopped being the flagship — it kept the flagship price, which means Grok 4.6 is now strictly better value for the same money unless you specifically want the cheaper cache rate or you are pinned to the 4.5 snapshot. Second, Grok 4.5 has half the context of Grok 4.3 and the 4.20 family at twice the input price. The V9 foundation is the reason: 1.5 trillion parameters, roughly three times the size of Grok 4.3, and a bigger base model is what bought the capability at the cost of the window.

If what you actually need is long context on a budget, Grok 4.3 is the honest answer inside this family and Grok 4.5 is not. That trade — capability against window — is the defining decision in the Grok line, and it is worth stating plainly rather than pretending the newer number is better at everything.

The spec sheet

These are the specifications xAI publishes on its own model page, not third-party estimates:

• Model identifier — grok-4.5, with the aliases grok-4.5-latest and grok-build-latest

• Modality — text and image in, text out. Images are accepted as input; the model does not generate them

• Context window — 500,000 tokens, shared between prompt and response rather than 500K in plus a separate output allowance

• Licence — proprietary. There are no open weights and no downloadable checkpoint; you reach Grok 4.5 through an API or through the vendor's own products

• List price — $2.00 per million input tokens, $6.00 per million output, $0.30 cached input

• Long-context pricing — at 200,000 prompt tokens and above, the rate steps to $4.00 input / $12.00 output / $0.60 cached, and the higher rate applies to every token in the request, not just the tokens past the threshold. This is the single most expensive detail on the page to miss

• Reasoning control — four levels, low, medium, high and xhigh, defaulting to high. Reasoning cannot be switched off entirely, and reasoning tokens bill as output tokens

• Tool and format support — function calling and structured outputs are supported natively. Server-side search tools are billed on top of token usage; we could not confirm the current per-call rates on the vendor's model page, only that they exist

• Rate limits — 150 requests per second and 50 million tokens per minute

• Batch API — not supported on Grok 4.5, unlike Grok 4.3

• Serving regions — us-east-1 and us-west-2 at launch. European availability was explicitly not available on day one and was described by the vendor as expected in mid-July 2026; we did not find a later primary-source statement confirming the exact date it opened, so treat the EU timeline as unconfirmed here

One discrepancy worth recording rather than smoothing over: OrcaRouter's own catalogue entry for grok/grok-4.5 shows a cache-read rate of $0.50 per million, which matches what xAI charges for Grok 4.6 and Grok 4.7 rather than the $0.30 the vendor's pricing table shows for Grok 4.5. The vendor's documentation is the authority on the price; our page appears to be carrying the successor's cache rate, and it has been flagged rather than quietly repeated here.

What Grok 4.5 is measurably good at

The headline claim at launch was token economy, and it is the one claim on the launch page that has a mechanism behind it rather than just a score. On SWE-bench Pro, xAI reports that Grok 4.5 averages 15,954 output tokens per resolved task against 67,020 for Claude Opus 4.8 running at maximum effort — roughly 4.2 times fewer tokens to finish the same work. Fewer output tokens per resolved task is the difference between a model that is cheaper and a model that is cheaper and faster in wall-clock terms, because output tokens are what you wait for. Vendor-reported, on the vendor's own harness, and unreproduced by anyone independent — but it is a structural claim, not a cherry-picked percentage.

The vendor's benchmark table, published with the model and therefore vendor-reported throughout, reads like this against the same models:

• DeepSWE 1.0 — Grok 4.5 62.0%, behind Claude Fable 5 at 66.1% and GPT-5.5 at 64.31%, ahead of Claude Opus 4.8 at 55.75%

• SWE Marathon, resolution rate — Grok 4.5 29.0%, ahead of Claude Opus 4.8 at 26.0% and Claude Fable 5 at 24.0%. This is the one coding benchmark where Grok 4.5 led the field

• Terminal-Bench 2.1 — Grok 4.5 83.3%, effectively level with Claude Fable 5 at 84.3% and GPT-5.5 at 83.4%, ahead of Claude Opus 4.8 at 78.9%

• DeepSWE 1.1 — Grok 4.5 53%, behind Claude Fable 5 at 70% and GPT-5.5 at 67%

• SWE-bench Pro — Grok 4.5 64.7%, behind Claude Fable 5 at 80.4% and Claude Opus 4.8 at 69.2%

The honest reading is that Grok 4.5 wins on long-horizon task resolution and on the economics of finishing a task, and loses on the harder single-shot coding benchmarks. It is a good model for agent loops and a mediocre one for one-shot hard problems — and that distinction is exactly what the token-efficiency number predicts it should be.

The independent picture is thinner, and it is where this page has to be careful. Artificial Analysis currently lists Grok 4.5 (high) at an Intelligence Index score of 39, ranked #52 of 212 models, on index version v4.3.2 — above the median of 25 for comparable models, but not near the front. If you have seen a score of 54 or 56 quoted for this model in July coverage, do not compare it to 39: Artificial Analysis has revised the index since launch and the two numbers come from different revisions. This is exactly the kind of stale-figure-comparison that makes benchmark tables untrustworthy, and it is why this page quotes the index version alongside the score.

What Grok 4.5 is measurably bad at

Artificial Analysis reports output speed of 55.1 tokens per second, ranked #126 of 212 models and below the median of 74 — a long way from the 80 tokens per second xAI quoted at launch and further still from the ~340 tokens per second that the fastest current models post. It also reports time to first token of 10.94 seconds against a median of 3.86 seconds. Those two numbers together describe the real cost of a reasoning model that defaults to high and cannot be turned off: you wait, and then it talks slowly. Leave the reasoning level at the default and Grok 4.5 will feel like the slowest model you have on the key.

Three more measured negatives, in order of how much they should affect a decision:

• Artificial Analysis has flagged Grok 4.5 as deprecated and says it continues benchmarking only the default 10,000-token-input workload for it. That is a third party telling you it has stopped fully measuring the model you are considering, and it recommends Grok 4.6 instead. Deprecated on a leaderboard does not mean unavailable on an API — it means the independent evidence base is frozen

• It loses the two benchmarks that most closely resemble hard, single-shot coding: DeepSWE 1.1 and SWE-bench Pro, both to Claude Fable 5 and to GPT-5.5, per xAI's own table

• It has half the context of Grok 4.3 at double the input price, and no batch API at all. If your workload is long-document processing or offline batch inference, Grok 4.5 is the wrong model in its own family

On hallucination and calibration, this page could not confirm a current independent figure. July coverage circulated a hallucination rate around 54%, but we could not verify that number on the Artificial Analysis model page we read, and it does not appear on the vendor's page either. Rather than repeat a figure we cannot source, we are recording it as not measured here.

Who should pick Grok 4.5, and who should pick something else

Pick Grok 4.5 if you are running agentic loops where the cost per completed task matters more than the cost per token, and you have already measured that the V9 token-economy behaviour holds on your workload. It is also the right pick if you need a Grok-family snapshot with a $0.30 cached-input rate on a large, stable, repeatedly-hit prefix — that cache rate is genuinely cheaper than what Grok 4.6 and Grok 4.7 charge, and on a cache-heavy workload that difference can outweigh two generations of capability.

Pick Grok 4.6 or Grok 4.7 instead for almost everything else. Same list price, same 500K window, newer training, better independent scores, and Artificial Analysis's own recommendation. There is no version of "I want the best Grok" where the answer is 4.5 today.

Pick Grok 4.3 if context is the constraint: 1M tokens at $1.25 and $2.50, with a batch discount that Grok 4.5 does not offer. Long-document RAG and whole-repository analysis are still its territory.

Pick something outside the family if you are optimising for peak score rather than price per task — Claude Fable 5 leads every coding benchmark in xAI's own comparison table, and GPT-5.5 sits ahead of Grok 4.5 on DeepSWE 1.0 and 1.1. If you want open weights and can accept a smaller model, that is a different purchase entirely and no Grok model competes for it.

Calling Grok 4.5 without a second contract

Grok 4.5 is live on OrcaRouter at xAI's list price, with the provider's rate passed through at zero markup — so the $2.00 and $6.00 above are what you pay, and a vendor price change lands on our side the same day it lands on theirs. That matters more than usual for this particular model, because the decision in front of most readers is not "Grok 4.5 or nothing" but "Grok 4.5 or Grok 4.6 or Grok 4.3", and comparing three models from one family is exactly the thing a single endpoint over 200-plus models makes cheap. One key, no second contract, and an A/B test that is a string change rather than an integration.

The failover argument is unusually concrete here. Grok 4.5 is still routable but no longer fully benchmarked by the independent trackers and no longer the vendor's flagship — a model you might want in a fallback chain rather than as your only path. Automatic failover across providers is the mechanism that lets you keep it for what it is good at without betting a production route on a snapshot the ecosystem has started to move past.

What this page will not do is pretend the choice is obvious. Grok 4.5 is a good model with a specific strength, sitting in a family where the same money now buys a newer one. If you are starting fresh today, the honest recommendation is Grok 4.6. If you are already running Grok 4.5 in an agent loop and it is finishing tasks at the token counts you budgeted for, the honest recommendation is to change nothing and re-measure in a quarter.

Grok 4.5 is routable on OrcaRouter today, and so are Grok 4.6, Grok 4.7 and Grok 4.3. OrcaRouter passes the provider's list price through at zero markup, so a vendor price change lands on our side the same day it lands on theirs, and moving a workload between two Grok models is a string change rather than a second contract.