Pick the graphics card you own, the resolution you play at, the smallest speed gain you would pay for and a budget. The tool sets your card to 100%, measures every other card in the benchmark table against it, and answers in one dated sentence: keep the card, or buy a named card at a named price.
It only uses numbers it can show a source for. Where it has no price for a card, it says so and waits for you to type one.
How to use it
- Pick the card you own. The list is grouped by series and covers the same benchmark table as the GPU Comparison Tool.
- Pick the resolution. 1080p, 1440p or 4K. It opens at 1440p.
- Set the smallest uplift worth paying for. It opens at 40%. That is my own starting point, not a measurement. The presets are 25%, 40%, 60% and 100%.
- Enter a budget. The verdict needs one. A card that costs a cent more is over budget.
- Read the verdict. The four numbers under it show where the field narrowed: how many other cards there are, how many clear your bar, how many of those have a price, and how many fit the budget.
- Type a price where you have a real one. Every card in the Upgrade column has a “Your price” field. A typed price replaces the price on file for that card.
- Copy the verdict or the link. The link carries your card, resolution, bar, budget and typed prices, so it reopens the same verdict.
How the verdict is worked out
Everything is whole percents, because the benchmark index only carries one decimal.
- Rebase.
relative score = card's index ÷ your card's index × 100, rounded to a whole percent. Your card is 100%. - Uplift.
uplift = relative score − 100. A card clears your bar when its uplift is at least your minimum. - Dollars per point.
price ÷ uplift. One point is one percentage point of uplift over your card. - Best pick. The lowest dollars per point among cards that clear the bar and cost no more than the budget. A tie goes to the larger uplift, then the lower price.
- Keep. If priced cards clear the bar but none fits the budget, the verdict is keep. If no card in the table clears the bar at all, the verdict is keep and no price is involved.
- No verdict. If cards clear the bar but none has a price, the tool says so instead of guessing.
The date in the sentence is the capture date of the listing prices, not the day you read it. When the pick turns on a price you typed, the sentence says so and carries no capture date. When no card is fast enough, the sentence is dated by the benchmark table.
A worked verdict
Take a GeForce RTX 3070 at 1440p with a 40% bar and a $1,000 budget, using the November 17, 2026 price capture.
The RTX 3070’s 1440p index is 34.8 and the Radeon RX 9070 XT’s is 69.7. 69.7 ÷ 34.8 × 100 is 200%, so the uplift is 100 points. The lowest qualifying RX 9070 XT listing on that capture was $999.99, and 999.99 ÷ 100 is $10.00 per point. The GeForce RTX 5070 also cleared the bar and fit the budget: 166% of the RTX 3070, 66 points, $919.99, which is $13.94 per point. The RX 9070 XT was the best pick.
Drop the budget to $700 and nothing with a price on that capture cleared a 40% bar for the money, so the verdict was keep. A later capture will change these numbers. That is why the sentence is dated.
Where the prices come from
Observed prices come from my GPU Price Index. For each card the index tracks, the tool uses the lowest qualifying listing in Newegg’s affiliate catalogue on the latest capture date. That catalogue names no seller. The listing is one board from one maker, and the price is a dated observation, not a live quote. Amazon prices are never used.
A tracked card can have no qualifying listing on a capture. For that card the tool shows the maker’s launch price, labelled “launch price, not a current price”, and leaves it out of the arithmetic.
Every other card has no price until you type one. The full file is gpu-price-observations.csv.
What this tool does not know
- Anything but native raster. The scores are the Tom’s Hardware GPU benchmarks hierarchy raster table dated June 24, 2026: one 11-game geometric mean at native resolution, with each resolution normalized separately. Ray tracing, DLSS, FSR, XeSS and frame generation are outside it.
- Prices beyond one retailer’s catalogue. A cheaper listing elsewhere, a used card or a rebate is yours to type in.
- Resale value. It does not subtract what your current card would sell for.
- Power and fit. It does not check board power, your power supply, connectors or case clearance. Use the PSU calculator before you buy.
- Your processor. It does not test whether the graphics card is what limits your games.
- Frame rates. A relative score is not FPS and does not promise a result in any one game.
Before you spend anything
Test whether the graphics card is the limit. The CPU and GPU lookup’s diagnostic steps compare a repeatable scene at two render resolutions. If lowering the render resolution raises the frame rate, rendering work was the limit in that test. If it barely moves, the cause needs more diagnosis before you buy a card. Low GPU usage in games covers frame caps, the active GPU and the other usual causes.
To see two cards side by side at all three resolutions, take them to the GPU Comparison Tool. For the whole field in one ranked list, use the GPU hierarchy.
Sources
- Performance: Tom’s Hardware GPU benchmarks hierarchy, June 24, 2026 raster table.
- Prices: my GPU Price Index, which publishes the method, every observation and the launch-price sources.
- Undervolt results in the keep plan: my GPU undervolt database, where every row links its source.
- Settings guides in the keep plan: the guides themselves, each with its own sources.