MSI Stealth 18 HX AI review

Frame rates

1920×1080, High preset, upscaling off.

Measured on this machine. Source: Notebookcheck, 2025-11-07.

In use

The short version

The MSI Stealth 18 HX AI, in the A2XW configuration, is an attempt at something genuinely difficult: an 18in laptop with the top mobile graphics part that does not weigh what 18in laptops normally weigh. It carries an RTX 5090 with 24GB of video memory, a Core Ultra 9 275HX, 64GB of system memory and a 2TB solid-state drive, behind an 18in 3840x2400 panel running at 120Hz. It weighs 2.957kg. There is no confirmed UK price for this exact configuration and we will not guess at one.

Notebookcheck measured eleven titles on 7 November 2025, all at 1920x1080 on the High preset with upscaling switched off. The results run from 381fps in Strange Brigade down to 88fps in Assassin's Creed Shadows. Four of the five games in our scoring suite are covered — Baldur's Gate 3 at 207fps, Dota 2 at 184fps, Cyberpunk 2077 at 168fps and X-Plane 11 at 117fps — for a four-game mean of 169fps.

There is no ranked score on this page, and the reason is specific. Our leaderboard averages every laptop across the same five games so that the comparison is genuinely like for like. For the Stealth 18 HX AI we have no GTA V figure, so the set is incomplete. Averaging four games and presenting the result next to machines averaged over five would produce a number that looks precise and is not. We would rather leave a gap in the table than fill it with something that quietly gives one machine a different exam from all the others. That is a hole in the data, not a criticism of the laptop.

The comparison data places this machine carefully. Against the MSI Cyborg 15, an RTX 5050 laptop at GBP 984, the Stealth averages 99fps more across four shared games — a decisive, unambiguous separation. Against the Aorus Master 18, an RTX 5090 running at a 175W power limit, it averages 3fps more. Against the Medion Erazer Beast 16 X1 Ultimate, another RTX 5090 machine costing GBP 4,884, it averages 1fps less. And against the Schenker XMG Neo 16 A25, a third RTX 5090 laptop, it averages 35fps less. Four results, three of them against machines with the same graphics chip, and they do not all say the same thing.

Buy it if you want an 18in flagship display and the top graphics part in a chassis that weighs under three kilograms, and if you have workloads — large memory footprints, high-resolution content work, local machine-learning experiments — that make 64GB of system memory and 24GB of video memory the reason the machine exists rather than a line on a spec sheet. Buy something else if the plan is to drive that 3840x2400 panel natively in the heaviest current games, because the published measurements do not answer that question and there is an important specification missing from the data.

Performance

The number we do not have

We have no graphics power limit for this machine, and on this particular laptop that absence matters more than it would on almost any other. It deserves stating at the top of the performance section rather than buried in a caveat.

Nvidia licenses the same graphics silicon to manufacturers across a wide range of sustained power budgets. The chassis, the cooling design and the firmware decide where in that range a given laptop settles, and the differences at the extremes are large. Two machines can both print RTX 5090 on the lid and behave like meaningfully different products once the cooling system has stopped being able to bank heat and has to settle into a steady state.

The reason this matters especially here is the machine's whole premise. A laptop named Stealth, built at 18in and weighing 2.957kg, is by definition a thinner and lighter design than the 18in norm — and thickness, mass and internal volume are precisely what determine how much sustained graphics power a chassis can dissipate. The natural question about a light 18in flagship is exactly what its power budget is, and we do not have the figure. One of the rivals below, the Aorus Master 18, has a stated 175W limit for comparison; for this machine the equivalent number is simply absent from our data.

We will not fill that gap with an assumption. What we can do is read the measurements, which are a more reliable guide than a specification anyway — with the important caveat that a benchmark sweep at 1920x1080 is a mild workload by the standards of what an RTX 5090 is designed for, and mild workloads are exactly the conditions under which a power limit is least likely to reveal itself.

What a power limit decides, and when

A graphics power limit governs the clock speed a card can hold once it is genuinely saturated: high native resolution, heavy shader work, ray tracing, large texture working sets, long sessions at steady state. Under those conditions the power budget is one of the most consequential specifications on the machine.

It decides very little about a graphics card that spends much of each frame idle, waiting for the processor to hand over the next batch of work. That second situation describes a substantial share of what was measured here, because every published figure was recorded at 1920x1080. At that resolution, on the High preset, with no ray tracing, an RTX 5090 is well ahead of the workload in most of the eleven titles listed, and the constraint moves elsewhere — single-thread processor throughput, driver overhead, memory latency, engine architecture.

The practical consequence for this review is uncomfortable but honest: the measurement set we have is the one least likely to expose whatever the Stealth's cooling design costs it. Strong 1080p figures are genuinely good news, and they are not evidence that the machine sustains full graphics power under heavy sustained load. Those are separate questions and we only have data on the first.

Working down the eleven-game list

Strange Brigade at 381fps is the headline number and the least useful one in the set. It is a well-built, efficient, lightly loaded shooter that became a benchmark staple precisely because it scales cleanly and runs fast on everything. At 1080p High it is not asking an RTX 5090 for anything difficult. A figure like that is a diagnostic — it confirms the memory subsystem is not choking, the drivers are behaving and power management is not doing anything odd — rather than a performance claim. No panel in this class can display it.

F1 24 at 326fps belongs in the same category. Racing titles built around frame-time consistency tend to be efficient renderers, and this one clearly is. Again the platform, not the graphics card, is setting the ceiling.

Baldur's Gate 3 at 207fps is the first result where the graphics card is doing recognisable work, and it is an excellent figure. Baldur's Gate 3 has a well-documented performance cliff in its dense third act, where the number of simultaneously simulated actors climbs sharply, and a benchmark run captures a specific scene rather than a worst case. At 207fps in the measured scene there is a great deal of headroom for that to be true and still leave a comfortable frame rate everywhere.

Final Fantasy XV at 198fps and Dota 2 at 184fps form the next band and are informative in the same way as each other. Final Fantasy XV is a DirectX 11 title with a rendering path from an earlier era of engine design. Dota 2 is an esports title engineered to run acceptably on modest hardware, with a performance profile dominated by single-thread processor speed. Neither is asking an RTX 5090 for anything, and both figures are statements about the Core Ultra 9 275HX and its memory subsystem far more than about the graphics card.

Cyberpunk 2077 at 168fps is the most useful single figure in the whole set. It remains the most demanding conventional rasterisation workload in wide benchmark circulation, and at 1080p High with upscaling off it is one of the few titles here where a flagship graphics card is meaningfully engaged. 168fps is a strong result. Read the conditions carefully though: High preset, no ray tracing, no upscaling. We have no ray-traced or path-traced figures for this machine, and those are exactly the workloads where an RTX 5090 with 24GB attached is supposed to earn its place. That absence is a real gap in the evidence.

Indiana Jones and the Great Circle at 146fps and Alan Wake 2 at 124fps are the modern-engine results and the ones that best represent what a current, visually ambitious release asks of hardware. Alan Wake 2 in particular is one of the heavier conventional renderers in wide benchmark use, and 124fps at 1080p High is a comfortable, unstressed result.

X-Plane 11 at 117fps is a low figure that is not primarily about the graphics card at all. Flight simulators of that generation lean extraordinarily hard on single-thread processor performance and on the draw-call and geometry pipeline. A machine reaching 117fps there has told you a great deal about its processor and memory latency and almost nothing about whether its graphics card has 16GB or 24GB attached to it.

Black Myth: Wukong at 95fps and Assassin's Creed Shadows at 88fps are the two lowest results and the two that matter most for judging the graphics implementation. Both are recent, graphically ambitious releases with high per-frame rendering costs, and both are genuinely graphics-limited at 1080p even on hardware of this tier. 88fps at the High preset is a good result in absolute terms — smooth and uncompromised — but it is less than a quarter of what the same machine produces in Strange Brigade, and that ratio is the most instructive thing in the data.

Where the spread comes from

The gap between the fastest and slowest results here — 381fps down to 88fps — is a spread of more than four to one, measured on the same machine, on the same day, at the same resolution and the same preset. None of it comes from settings, thermal variation or measurement noise. All of it comes from how eleven game engines distribute their work.

The ordering is close to chronological and that is not accidental. Titles at the top were built for hardware generations that no longer exist in this form, with low per-frame graphics costs, so the platform runs out of processor before it runs out of graphics card. Titles at the bottom are recent releases with substantially higher per-pixel costs — denser geometry, more expensive lighting, heavier post-processing — and on those the graphics card becomes the constraint and the frame rate falls to what it can actually deliver.

The practical lesson is that this laptop's relationship with a given game is largely a function of when the game was made. That is a more useful predictive tool than any average, and it applies to future releases as directly as to past ones. Assassin's Creed Shadows at 88fps is not a warning about this machine's capability — it is the highest-quality information in the set about what the current front edge of game development costs to render, on hardware at the top of the mobile stack, at a resolution well below this laptop's own panel.

The four-game mean, and the missing fifth title

Our scoring suite covers five titles: Dota 2, GTA V, Cyberpunk 2077, X-Plane 11 and Baldur's Gate 3. For the Stealth 18 HX AI we have four of them, averaging 169fps. GTA V is absent.

It would be easy to shrug at that and publish the four-game mean as though it were comparable to a five-game mean. It would also be wrong. GTA V is a decade-old engine that predates modern multi-core scheduling and is heavily limited by single-thread processor performance at 1080p on fast hardware — which is to say it behaves quite differently from the graphics-bound titles in the set, and its inclusion would move a mean by an amount that depends entirely on the platform underneath it. Adding a fifth game to a four-game average is not a small perturbation; the composition of a benchmark set is a substantial input to the result, not a neutral background detail.

So there is no score here, and there will not be one until a GTA V figure exists from a source we are willing to stand behind. That cuts both ways, and it is worth saying so explicitly: nobody should read the absence of a rank as evidence that the machine underperforms. On the four suite games we do have, it is comfortably a fast laptop.

What these numbers cannot tell you

The unmeasured territory here is large and it is precisely the territory that would justify the money. There are no ray-traced figures, no path-traced figures and no results with DLSS or frame generation enabled — and on an RTX 5090 those are not side dishes, they are a substantial part of the argument for buying the top part rather than the one below it. There are no measurements at the panel's native 3840x2400, which is the resolution anyone buying this display will actually want to use.

We also have no measured battery life, and indeed no battery capacity figure at all in our data for this configuration, so we will not comment on unplugged endurance beyond the structural observation that no laptop of this class runs its graphics card near full power on battery. We have no noise measurements, no surface temperature data, and nothing on how the machine behaves after an hour of sustained load. On a thin 18in chassis carrying a flagship graphics part, that last gap is the one we would most want filled.

The screen

The panel is 18in at 3840x2400 running at 120Hz. Our data does not state the panel technology for this configuration, so we will not tell you whether it is OLED, mini LED or conventional IPS — that is a meaningful difference in contrast, brightness behaviour and longevity, and we are not going to guess at it. We also have no brightness measurement, colour gamut figure or response time data.

What we can discuss with confidence is the relationship between the resolution and the measured frame rates, and it is the most important thing about this machine. 3840x2400 is 9,216,000 pixels. 1920x1080, the resolution every published figure on this page was recorded at, is 2,073,600. The native panel is four and a half times the pixel count of the benchmark resolution.

That is not a gap you can extrapolate across responsibly, and we are not going to try. What can be said without inventing anything is how the two ends of the measured list would behave differently. Strange Brigade at 381fps and F1 24 at 326fps are not graphics-limited at 1080p, so their frame rates would fall less steeply with resolution than a purely graphics-bound title would. Black Myth: Wukong at 95fps and Assassin's Creed Shadows at 88fps are already graphics-limited at 1080p, and raising the pixel count more than fourfold is the change that hits graphics-limited titles hardest. The spread between the top and bottom of the list, already more than four to one, would widen considerably at native resolution. Exactly where each title lands is not something this data can tell you.

The refresh rate is 120Hz, which is the sensible ceiling for a panel of this pixel count — asking for more would specify a display no mobile graphics card could feed. At the measured 1080p, eight of the eleven titles exceed 120fps and would saturate the panel: Strange Brigade, F1 24, Baldur's Gate 3, Final Fantasy XV, Dota 2, Cyberpunk 2077, Indiana Jones and the Great Circle and Alan Wake 2. Three fall short — X-Plane 11 at 117fps, Black Myth: Wukong at 95fps and Assassin's Creed Shadows at 88fps — though all three are perfectly comfortable frame rates.

The honest assessment of this display is that it is specified for content work first. An 18in panel at 3840x2400 gives an enormous amount of usable resolution for video timelines, large 3D scenes, photography and multi-window layouts, and it is genuinely rare in a portable machine. Paired with 64GB of system memory it describes a mobile workstation more clearly than it describes a gaming laptop. As a gaming display it is capable and its 120Hz refresh is well matched to what the hardware produces at sensible resolutions — but a buyer choosing it for gaming should expect to run games below the panel's native resolution, or with upscaling, in anything demanding. That is not a criticism of the machine so much as an accurate description of what any mobile graphics card can do against nine million pixels.

Build, size and portability

2.957kg for an 18in laptop with a flagship graphics part is the most interesting specification on this machine, and the one that justifies the name. Eighteen-inch gaming laptops are normally substantially heavier than this, because sustaining a high graphics power budget alongside a high-end processor requires a large vapour chamber or heat-pipe array, extensive fin area, fans big enough to move air through it, and a chassis rigid enough to hold the whole assembly without flexing. Coming in under three kilograms at this screen size represents a real engineering result.

It also raises the obvious question, which we cannot answer: what was traded to achieve it. The candidates are cooling capacity, sustained graphics power, acoustic behaviour under load, or some combination. We have no power limit figure, no noise measurement and no sustained-load data, so we are not going to assert that anything was compromised. We will say that this is the specific area where a prospective buyer should look hardest for evidence beyond a 1080p benchmark sweep, because a 1080p sweep is the workload least likely to reveal a cooling constraint.

What 2.957kg means in practice is that this is an 18in machine you might genuinely move, which is unusual. It will go from desk to bag and out at the other end without the resentment that a heavier 18in machine generates. What it will not do is stop being 18in. The footprint remains awkward on an aeroplane tray table, in a standard backpack sleeve, and on a small desk — and those constraints are about area, not mass. A lighter 18in laptop is still a large object. The real travelling weight is also higher than 2.957kg because a power supply for a flagship graphics part has to come with it, and we do not have the adapter's weight in our data so we will not put a number on the total.

On the rest of the configuration, the story is consistent and it is a workstation story. 64GB of system memory is far beyond what any current game requires; it is the specification of a machine expected to hold large video timelines, substantial 3D scenes, several virtual machines or big local datasets. The 2TB solid-state drive is generous enough to carry a serious game library alongside working files without constant management. And 24GB of video memory is a genuine long-term asset — far more than any current title needs at the resolutions this machine will realistically render at, which makes it insurance against future texture budgets and immediately useful for local machine-learning work, large scenes in 3D applications and high-resolution video.

Put the 18in 3840x2400 panel, the 64GB of memory, the 24GB of video memory and the sub-three-kilogram weight together, and the machine's identity is clear. It is a portable workstation with flagship gaming hardware inside it, aimed at someone who needs both and does not want to own two machines. That is a coherent product and a genuinely underserved niche.

How it compares

Four machines from the same graphics generation, chosen to give a spread rather than just the nearest neighbours. Notably, three of the four carry the same RTX 5090 graphics part, which makes this comparison set unusually informative about how much the rest of a laptop matters. All comparisons use only games both machines have figures for, from the same 1080p High measurement conditions.

MSI Cyborg 15 — RTX 5050, GBP 984

Across four shared games the Stealth 18 HX AI averages 99fps more than the Cyborg 15. That is the widest margin in this comparison set by a very large distance, and it is the number that makes the rest of the section interpretable. A 99fps average gap across four titles is not measurement noise or run-to-run variance — it is the difference between an entry-level graphics part and a flagship, and the benchmark reports it emphatically.

Hold onto that when reading the near-parity results below. The temptation, on seeing a 3fps or 1fps difference, is to conclude that the test cannot resolve differences between machines. It plainly can. It resolved 99fps here and 35fps in the other direction against the XMG Neo. When it reports parity between two machines, that is a measurement about those machines in those games, not a failure of the instrument.

The Cyborg 15 costs GBP 984. We have no confirmed price for the Stealth 18 HX AI, but an 18in machine with an RTX 5090, 64GB of memory and a 3840x2400 panel is not competing in that bracket and the comparison is not a value judgement. It is a calibration of what the benchmark can see.

Aorus Master 18 — RTX 5090 at 175W, no confirmed UK price

Across four shared games the Stealth averages 3fps more than the Aorus Master 18, which carries the same RTX 5090 at a stated 175W power limit. Three frames is not a meaningful separation; these two machines produced the same result.

The explanation is the one that governs this entire data set. Of the four shared suite games, Dota 2 and X-Plane 11 are limited by processor throughput and memory behaviour rather than graphics throughput at 1080p, and Baldur's Gate 3 at 207fps is not stressing a flagship card either. In a processor-limited title, adding graphics capability produces no additional frames at all — the card finishes each frame earlier and waits longer for the next batch of work. Only Cyberpunk 2077 gives the two graphics implementations meaningful room to differ, and one title out of four is not enough to move an average appreciably.

This result is genuinely useful for one specific reason, though. The Aorus Master 18's power limit is stated at 175W and the Stealth's is not stated at all. Parity on this test does not prove the Stealth runs at a comparable power budget — a 1080p sweep would not distinguish them even if it did not — but it does establish that whatever the Stealth's budget is, it is not so constrained that it falls behind a 175W machine even in the one graphics-bound suite title. That is a modestly reassuring data point rather than a conclusive one, and we will not stretch it further than that.

Medion Erazer Beast 16 X1 Ultimate — RTX 5090, GBP 4,884

Across four shared games the Stealth averages 1fps behind the Medion Erazer Beast 16 X1 Ultimate, another RTX 5090 machine, this one with a confirmed UK price of GBP 4,884.

A single frame across four games is as close to identical as two measurements get, and the reasoning is the same as above: the same graphics chip, in games that mostly do not stress it, at a resolution well below either machine's capability. What makes this comparison more useful than the previous one is the price. GBP 4,884 is the only confirmed figure among the three RTX 5090 rivals, and it establishes the order of magnitude that a flagship machine of this class occupies. It also frames the question a buyer should ask about the Stealth when its own price appears.

The Beast is a 16in machine and the Stealth is 18in with a considerably higher-resolution panel and 64GB of memory. Those differences are the whole basis for choosing between them, because the frame rate data offers no separation whatsoever. If the display size and resolution and the memory capacity are what you need, the Stealth is the machine. If they are not, this comparison is telling you that a 16in alternative delivers the same measured gaming performance, and the money should follow whichever set of non-gaming attributes you actually use.

Schenker XMG Neo 16 A25 — RTX 5090, no confirmed UK price

Across four shared games the Stealth averages 35fps behind the Schenker XMG Neo 16 A25 — a fourth machine with the same RTX 5090 graphics part. This is the result that makes the whole section coherent.

Set the four figures side by side: plus 99, plus 3, minus 1, minus 35. Three of those comparisons are against machines with the same graphics chip, and they land at plus 3, minus 1 and minus 35. That last figure is the important one, because it demonstrates that even among identically specified graphics parts, this benchmark can resolve a substantial difference. Two 5090 machines can genuinely be 35 frames apart on the same four games.

What produces that gap, on titles that are mostly processor-limited, is almost certainly not the graphics tier — it cannot be, since the tier is identical — but the rest of the platform: processor performance and how aggressively it is permitted to boost, memory bandwidth and latency, how the firmware distributes power between processor and graphics card when neither is fully loaded, and how quickly the whole system can feed work forward. The XMG Neo produces high figures against a range of machines on this test, which points to a platform tuned particularly well for exactly these conditions.

For the Stealth, the honest reading is that it sits in the middle of the RTX 5090 field on this test rather than at the top of it. That is a perfectly respectable position and it does not undermine the machine — but it does mean that buying it on the strength of "it has a 5090" would be a mistake. Three other laptops here have the same chip and land at plus 3, minus 1 and minus 35 relative to it. The chip is a necessary condition for top-tier performance, not a sufficient one.

Reading the four results together

The comparison set says something specific and slightly unusual. This machine is decisively faster than the tier below it and functionally indistinguishable from two other flagship machines, one of which costs GBP 4,884. It is also clearly behind the fastest 5090 platform in the group. In other words, it belongs firmly in the top group and is not the best of that group on this particular test.

That position has to be justified by something other than raw frame rate, and in this machine's case there is an obvious candidate: it is the only laptop in the comparison set that combines an 18in 3840x2400 display, 64GB of system memory and a weight under three kilograms. If those attributes are what you need, the frame rate parity with more conventional flagships is exactly the result you would want — it means the unusual form factor costs you nothing measurable in games. If those attributes are not what you need, the comparison table is quietly pointing you at alternatives.

Who it's for

The Stealth 18 HX AI makes most sense for someone with a genuinely mixed workload and a strong preference for one machine rather than two. The combination of an 18in 3840x2400 panel, 64GB of system memory, 24GB of video memory and a 2TB drive is a mobile workstation specification: enough screen area and resolution to edit video or work on complex 3D scenes without an external display, enough memory to hold large datasets, and enough video memory for local machine-learning experiments or heavy GPU-accelerated content work. That it also runs Cyberpunk 2077 at 168fps and Baldur's Gate 3 at 207fps at 1080p makes it a serious gaming machine as well.

It suits someone who has previously rejected 18in laptops on weight grounds. 2.957kg is not light in absolute terms, but it is markedly lighter than 18in flagship machines usually are, and for a buyer who wants the screen area but has been unwilling to carry what normally comes with it, this is one of the few machines that addresses the objection directly.

And it suits a buyer who reads the comparison data correctly and finds it reassuring. Parity with a GBP 4,884 flagship and with a 175W RTX 5090 machine on the shared games means the thin chassis is not costing measurable gaming performance under these conditions. That is the result the machine needed to produce.

Who should buy something else

If you intend to game at the panel's native 3840x2400, look for evidence this page cannot give you. That resolution is four and a half times the pixel count every published figure here was recorded at, and we have no measurements at it and no power limit figure to reason from. Assassin's Creed Shadows at 88fps and Black Myth: Wukong at 95fps at 1080p are already the graphics-limited end of the list, and the pixel count is the change that hits graphics-limited titles hardest. Plan on running below native resolution or with upscaling in demanding titles, or find measurements that answer the question directly.

If ray tracing is central to what you play, there are no ray-traced figures for this machine in our data. That is an evidence gap rather than a verdict, but it is a gap in exactly the workload where a flagship graphics part is supposed to distinguish itself, and where a chassis with unstated cooling headroom carries the most uncertainty.

If you want the highest measured frame rate available with this graphics part, the XMG Neo 16 A25 averages 35fps ahead across four shared games. It is a 16in machine with a smaller screen and no confirmed UK price in our data, but on this test it is simply faster and that is the honest recommendation for anyone whose criterion is the number.

If you do not need the 18in screen area or the 3840x2400 resolution, buy a 16in machine. The Medion Erazer Beast 16 X1 Ultimate delivers the same measured performance across four shared games at GBP 4,884, in a smaller and more manageable format. Paying for an 18in panel you do not use is the most expensive mistake available in this category.

If portability genuinely matters day to day, 2.957kg is a good result for 18in and still a poor result in absolute terms, and the footprint is unchanged by the weight saving. Anyone commuting daily should be looking at 16in or smaller and accepting the reduced screen area.

And if value is your deciding criterion, this page cannot help, because there is no confirmed UK price for this configuration. What the data does give you is the reference point to judge one against when it appears: the same measured gaming performance as a GBP 4,884 machine across four shared games. Whatever the Stealth costs relative to that, the frame rates are not the reason to pay more.

The bottom line

The MSI Stealth 18 HX AI is an unusual machine that does an unusual thing well: it puts an 18in high-resolution display, a flagship graphics part and 64GB of memory into a chassis under three kilograms, and the measured frame rates show no obvious penalty for it at 1080p. A 169fps average across the four scoring-suite games we have is a strong result, and the eleven-title spread from 381fps down to 88fps describes a laptop that handles everything currently available at that resolution without compromise.

The reservations are all about evidence rather than about the machine. There is no GTA V figure, so there is no ranked score. There is no stated graphics power limit, which on a deliberately thin 18in design is the specification a buyer would most want. And there are no measurements at the panel's own native resolution, which leaves the single most important performance question about the machine unanswered.

Judged as a mobile workstation that games extremely well, it is a coherent and genuinely distinctive product with few direct competitors. Judged purely as a gaming laptop, the comparison data places it mid-pack among RTX 5090 machines — level with two of them, 35fps behind the fastest — which means the case for it rests on the display, the memory and the weight rather than on the frame rates. That is a legitimate case. It is just not the case the graphics part on the box appears to be making.