Grok 4.5 review: 9/9 on our executed benchmark, at 67% more per 1,000 tasks than Grok 4.3

We ran Grok 4.5 through our own executed coding benchmark on 2026-07-29. It scored 9/9 at a measured $2.93 per 1,000 tasks, 6.6 s average, 289 reasoning tokens per call, priced at list on 2026-07-29. We had already run Grok 4.3 on the same harness: 8/9 at $1.75, missing flatten, priced at list on 2026-07-17. So the newer flagship genuinely fixed the miss, and it charges 67% more per 1,000 tasks to do it. That is the upgrade-or-not answer. Everything else on this page — the Artificial Analysis Intelligence Index of 54, the SWE-Bench Pro resolve rates, the token-efficiency arithmetic — is third-party, and labelled as such throughout.

Every review of Grok 4.5 that hit the first page of search this week repeats the same two facts: an Intelligence Index of 54, and coding roughly on par with GPT-5.5 at half the cost. Both are supported by the data. Neither tells you whether it is the right model to put in production. This review starts from those numbers and then does the thing the AI Overview will not: it normalizes them by what a task actually costs to finish.

It also does something no aggregation of those numbers can. We ran Grok 4.5 on our own executed benchmark, and we had already run Grok 4.3 on the same harness — so we can answer the question a Grok 4.5 review is actually opened for: is the upgrade worth it, and by how much. That section is the only first-party measurement on this page. Everything else is attributed.

What Grok 4.5 actually is

Grok 4.5 is xAI's newest frontier model, launched in early July 2026 and independently benchmarked by Artificial Analysis. It ships on the x.ai domain and is served as x-ai/grok-4.5 on OpenAI-compatible gateways. (After the SpaceX–xAI tie-up, some mid-2026 coverage also brands the maker SpaceXAI — it is the same company.)

On the independent Artificial Analysis Intelligence Index, a vendor-neutral aggregate, Grok 4.5 scores 54. That is roughly a 16-point jump over Grok 4.3, which is the company's largest generation-over-generation gain on that index to date. If you are coming from the previous release, our Grok 4.3 review is a useful baseline for how much changed.

Under the hood it is a reasoning model with text and image input, function calling and structured outputs, a 500k-token context window, and a reasoning_effort parameter with low, medium and high settings (high is the default). Independent measurement clocks output at about 92.9 tokens per second, faster than the frontier average; some vendor-side write-ups cite roughly 80 tokens per second, so treat the exact figure as measurement-dependent.

Where it lands on the frontier

Rank depends on how you count. Artificial Analysis's model page lists Grok 4.5 at #8 of 188 when every model configuration is counted separately. Their editorial write-up frames it as roughly 4th among distinct frontier models, behind Fable 5, GPT-5.5 and Opus 4.8. Both framings are fair; the honest summary is that Grok 4.5 sits at the lower edge of the top tier on raw intelligence, not at the top of it.

Where it does lead is agentic tool use. Grok 4.5 posts the top score on several agentic benchmarks, including tau3-Banking at 33 percent (above GPT-5.5 xhigh at 31 percent) and the lead on AutomationBench-AA. If your workload is tool-calling and multi-step orchestration rather than one-shot reasoning, that ranking matters more than the aggregate index.

The coding claim, scoped honestly

The community line is that Grok 4.5 is on par with GPT-5.5-xhigh in coding at about half the cost. That is true, but only for a specific benchmark: the Artificial Analysis Coding Agent Index. It is not true across the board. On raw SWE-Bench Pro resolve rate, third-party testing (eesel, the-decoder) puts Grok 4.5 at roughly 64.7 percent — ahead of GPT-5.5 at 58.6 percent, but behind Opus 4.8 at 69.2 percent and Fable 5 at 80.4 percent. These SWE-Bench numbers come from review blogs rather than the primary benchmark, so read them as directional.

So the accurate one-liner is: Grok 4.5 is a strong, mid-frontier coder that beats GPT-5.5 on task resolution and matches it on the agentic index, but it is not the highest-quality coder available. If you want the full methodology behind these price-per-task coding comparisons, see our LLM coding cost benchmark, and for the head-to-head positioning read Grok vs GPT-5.

What we measured: 9/9 at $2.93 per 1,000 tasks

Everything above this line is somebody else's number. This section is ours.

We ran Grok 4.5 on our executed 9-task Python benchmark on 2026-07-29. It scored 9/9, at a measured $2.93 per 1,000 tasks priced at list on 2026-07-29, 6.6 s average per task, and 289 reasoning tokens per call. Nine tasks, one scored attempt each, temperature 0, a 4,000-token ceiling, and code executed against assertions the model never sees. The calls go through OpenRouter's OpenAI-compatible endpoint, deliberately not through the DataLLM Lab gateway, so nothing here depends on our infrastructure and you do not have to be our customer to reproduce it.

The comparison that matters is against the model it replaces. We had already run Grok 4.3 on the same harness: 8/9 at $1.75 per 1,000 tasks, 8.4 s, 482 reasoning tokens, priced at list on 2026-07-17 — a rate xAI had not moved by our 2026-07-29 capture, so the two cost figures are comparable despite the different pricing dates. Its one miss was flatten, a nested-list flattening task. Grok 4.5 did not miss it.

So the upgrade is real and it is priced. Grok 4.5 fixed the miss — 9/9 against 8/9 — and cost 67% more per 1,000 tasks to do it. It was also 1.8 s faster per task and spent 40% fewer reasoning tokens, which is the interesting part: the newer model got the harder task right while thinking less, and still cost more, because its per-token rate is higher. Full Grok 4.3 detail is in our Grok 4.3 review.

Claude Haiku 4.5 and GPT-5.4 are in the table as reference points, not as rivals picked to flatter anyone: both scored the same 9/9 on the same nine tasks, both emitted zero reasoning tokens, and both did it for less money. On this workload, Grok 4.5's 9/9 is 3.1x the measured cost of the cheapest 9/9 in the group. That is not a knock on Grok 4.5 — it is what nine short Python functions can and cannot separate.

Two things about the dollar figures. They are measured cost, not a bill we received: token counts the API reported, multiplied by that model's list price on the stated date. And list prices move: 49 of roughly 396 models in our catalog changed price in the twelve days to 2026-07-29, which is why every cost figure here carries the date it was priced on rather than being left to drift into looking current. The wider field these numbers sit in is in our AI coding ranking and the full coding cost benchmark.

What our nine tasks do not tell you

Nine short Python functions say nothing about the work Grok 4.5 is actually pitched for. The 500k context window, agentic tool-calling, multi-step orchestration, image input — our harness exercises none of it. It is single-turn, Python-only, and every task fits comfortably in a few hundred lines. A model built for long agentic runs is being asked to write two_sum.

So read the 9/9 as: Grok 4.5 clears the floor, and clears the one hurdle its predecessor tripped on. Read the $2.93 as: on work this size, you are paying a premium that this workload gives it no chance to earn back. The tau3-Banking and Coding Agent Index results above are where that premium might be justified, and those are Artificial Analysis's measurements, not ours.

Also worth stating plainly: one miss on nine tasks is one data point. We would not tell you Grok 4.3 cannot flatten a nested list. We would tell you it did not, on this attempt, at temperature 0, with a 4,000-token ceiling and no retry bought for a wrong answer. Grok 4.5 got it. That is the whole of the evidence. Full boundaries are on the methodology page.

Cost per completed task

This section is arithmetic on third-party token counts, not on ours. Our measured numbers are above and cover nine short Python functions; the SWE-Bench Pro token counts below come from third-party measurement of a much larger workload, and we did not reproduce them. Read them as directional and attributed.

With that said, the mechanism is worth spelling out, because raw benchmark scores compare quality and not what you pay to get a task finished — which is a function of both per-token price and how many tokens the model burns. The claimed Grok 4.5 advantage is token efficiency: on SWE-Bench Pro tasks it uses roughly 16k output tokens to solve a task, versus about 67k for Opus 4.8 in third-party measurement — roughly 4x fewer tokens for a comparable job.

Put that against its list price and the economics flip. At 6 dollars per 1M output tokens, a 16k-token Grok 4.5 completion costs about $0.096 on the output side — call it a dime a task. A model that needs 67k tokens to finish the same task would have to price its output below about $1.43 per 1M just to match that dime, before you even factor input and reasoning tokens. That is the mechanism by which a #4 intelligence model can be the #1 value model on long agentic runs.

Every figure in that table is third-party. Dashes mark values not published in the datasets we cite; we do not fill them with guesses. The takeaway is not that Grok 4.5 is the best at everything — it plainly is not — but that on the token-heavy, multi-step work these third parties measured, its reported ~4x efficiency edge can make it the cheaper path to a finished result, while raw one-shot quality still favors Opus 4.8 and Fable 5. Note that this does not carry over to short tasks: on our own nine, Grok 4.5 was 3.1x the measured cost of Claude Haiku 4.5 for the same 9/9, because there is no long agentic run for its efficiency to save anything on.

Pricing, context and access

Grok 4.5 lists at 2.00 dollars per 1M input and 6.00 dollars per 1M output, with cache hits around 0.50 dollars per 1M. That is a step up from Grok 4.3 at 1.25 in and 2.50 out — you are paying more per token for a much higher intelligence score. Both rates were still current in our 2026-07-29 price capture, and they are what produced the $2.93 and $1.75 figures above. Our Grok API pricing guide breaks down where each tier makes sense. The 500k context window and function-calling support mean it slots cleanly into existing agent stacks.

Availability at launch spanned the xAI API, Grok Build, Cursor, OpenRouter and others. Per launch-week coverage it was not yet available in the EU on first-party endpoints, though that kind of restriction tends to move quickly. If you are weighing Grok against the Anthropic and OpenAI stacks generally, Claude vs ChatGPT covers the broader trade-offs, and you can inspect the individual model pages for Opus 4.8 and GPT-5.5 to compare specs directly.

Who should actually switch

Coming from Grok 4.3, the upgrade is real and it costs 67% more. That is the one part of this verdict we measured: 9/9 against 8/9 on the same nine tasks, the flatten miss fixed, $2.93 against $1.75 per 1,000 tasks. If those misses were costing you retries, the newer model is worth the difference. If your short tasks were already passing on 4.3, this benchmark gives you no reason to move.

Switch to Grok 4.5 if your workload is agentic and token-heavy: tool-calling pipelines, long coding sessions, automation flows where the model takes many steps. There its reported tool-use leadership and ~4x token efficiency are the argument — but that argument rests on Artificial Analysis and third-party SWE-Bench testing, not on anything we ran. Stay on Opus 4.8 or Fable 5 if you need the highest raw SWE-Bench Pro quality on hard, one-shot problems.

And do not buy it for short work. On nine small Python functions, Claude Haiku 4.5 scored the same 9/9 for $0.94 and GPT-5.4 for $1.69, both priced 2026-07-29. The cheap coding tier is where that class of task belongs. The honest positioning: Grok 4.5 is a genuine generation-over-generation fix at a real price increase, and whether the increase pays for itself depends entirely on whether your workload looks like the long agentic runs it was built for — which our harness cannot tell you.

FAQ