RTX 4090 vs RTX 5090 for LLMs: same-model tok/s, prices and value, November 2026

On the same model and quant, an RTX 5090 decodes 1.3 to 1.6 times faster than an RTX 4090 on most models in published llama.cpp runs (61.4 against 39.6 tok/s on Qwen3 32B, 334.3 against 224.7 on Nemotron 3.5 Lightning) and holds 32 GB instead of 24 GB. At 7 November 2026 prices, a used 4090 at a $2,731 eBay sold median buys about 1.6 times more decode speed per dollar than a new 5090 at $6,549.99 (Newegg); in Germany the gap is similar, EUR 2,447 against EUR 6,249. The 4090 wins on value; the 5090 wins when 24 GB is what stops you.

Sources in the sections below. Ratios and per-dollar figures are my arithmetic.

What changed between the RTX 4090 and RTX 5090?

The 5090 adds 8 GB of memory, 78% more bandwidth, PCIe 5.0 and native FP4 tensor math, and it draws 125 W more. The rows below come from NVIDIA's product pages unless marked.

Read 7 November 2026: NVIDIA RTX 4090 page, RTX 5090 page, NVIDIA compare page (1,792 GB/s), The FPS Review (4090 at 21 Gbps, 1,008 GB/s), Wikipedia RTX 50 series. NVIDIA's product pages do not state which precision the AI TOPS figures use, so do not read the 2.5x as a like-for-like gain at one format.

The format row matters more for LLMs than the TOPS row. NVIDIA's NVFP4 introduction says the format came with Blackwell's fifth-generation tensor cores, and vLLM's quantization hardware table lists FP8 W8A8 as supported on Ada but FP4 only through its Marlin weight-only kernels. So an NVFP4 checkpoint runs on a 4090, but as compressed storage with 16-bit math; on a 5090 the math itself is 4-bit. The quantization docs cover the formats.

How much faster is an RTX 5090 than an RTX 4090 on the same model?

About 1.4 to 1.6 times in single-stream decode on dense models, less on some MoE models, and only 1.0 to 1.3 times in llama.cpp prefill. These pairs come from the same publisher, engine and quant on both cards, which is the only way the ratio means anything.

Tokens per second. Hardware Corner: RTX 4090 page and RTX 5090 page, both "Updated Benchmarks: March 2026", 4k context; the 4090 page labels gpt-oss 20B MXFP4 and the 5090 page Q4_K, so I left that pair out. Atomic Chat: 4090 article (21 September 2026) and 5090 article (18 September 2026), llama-bench, fp16 KV, 4k prompt. CloudRift, 9 October 2025: 8k context, FP8 KV, sustained throughput. Ratios are mine.

The decode ratios sit below the 1.78x bandwidth ratio, as expected when neither card reaches its full bandwidth ceiling; the dense models get closest. The Qwen3 30B-A3B pair at 1.09x is the outlier against 1.28x and 1.49x for the 2026 MoE pairs, so do not draw a MoE rule from one row. The batched vLLM result at 2.02x is the one where the 5090's extra memory shows up as speed: more room for KV cache means more sequences per batch. The full RTX 5090 table is in the RTX 5090 benchmarks.

What does 32 GB fit that 24 GB does not?

Mostly context and higher-quality quants on the same 27B to 35B models, not new model sizes. Atomic Chat's RTX 4090 reports 24,564 MiB (23.99 GiB) and its 5090 32,607 MiB (31.84 GiB). Taking Qwen3.8-27B, whose config.json gives 64 KiB of FP16 KV cache per token, and reserving 1.5 GiB for overhead and 0.14 GiB for its recurrent state:

max context = (VRAM − 1.5 − 0.14 − weights) ÷ 64 KiB

Qwen3.8-27B UD-Q4_K_XL (16.35 GiB)
  RTX 4090: (23.99 − 1.64 − 16.35) =  6.0 GiB → about  98,000 tokens
  RTX 5090: (31.84 − 1.64 − 16.35) = 13.9 GiB → about 227,000 tokens
Qwen3.8-27B Q6_K (20.47 GiB)
  RTX 4090: 1.9 GiB → about  31,000 tokens
  RTX 5090: 9.7 GiB → about 159,000 tokens

My arithmetic; file sizes from Unsloth's GGUF repo, checked 7 November 2026. Hardware Corner's measured runs agree in shape: its 4090 page shows Qwen3 32B stopping at 16k, while the 5090 ran it at 32k (43.8 tok/s). Atomic Chat ran Gemma 4 31B at the smaller IQ4_XS on the 4090 (20,380 MiB at 32k) and at Q4_K_M on the 5090. For agent work with long contexts or several sessions, those 8 GB are the reason to buy the newer card. The model-by-model fit table is in the best LLMs for an RTX 5090, and for 70B-class models neither card is enough, which the memory calculator shows.

What do an RTX 4090 and an RTX 5090 cost in November 2026?

New 4090s are gone from US retail, used ones sell for about $2,700, and a new 5090 costs over three times its $1,999 launch price. All prices below were read on 7 November 2026.

Prices include whatever the seller charges; US prices exclude sales tax. pcprice.watch medians are across seven eBay markets; the EUR figures are from its Germany 4090 and Germany 5090 pages. The used 5090 median above the new Newegg price reflects listings sold earlier in the 90-day window. A country-by-country 5090 price table is in the sibling article on RTX 5090 prices.

Which card gives more tokens per second per dollar and per watt?

A used 4090 gives about 1.5 to 1.7 times more single-stream decode per dollar and still about 1.2 times more in batched serving; the 5090 gives 14 to 21% more decode per rated watt, and 58% more in batched serving. Using the used 4090 median ($2,731) and the cheapest new 5090 ($6,549.99):

My arithmetic on the speeds in the comparison table. German prices give the same picture: EUR 6,249 for the cheapest new 5090 is 2.55 times the EUR 2,447 used 4090 median, for 1.3 to 1.6 times the decode speed. "Per rated W" divides by total graphics power (450 W and 575 W), not measured draw; single-stream decode usually draws less than the rating, so treat these as relative figures. Buying a new 5090 instead of a used one changes the 5090 column by under 3%.

The per-watt edge is real but small in money terms. At the EU household average of EUR 0.2832 per kWh (Eurostat, first half of 2026), the 125 W difference in rated power costs EUR 0.035 per hour at full load. The sibling article on RTX 5090 power limits shows that a 5090 capped at 400 W gives up about 1.4% of decode speed, which closes most of that gap anyway.

Should you upgrade from a 4090 or buy a used one?

Keep your 4090 unless 24 GB is blocking a model or context you need; if you are buying today, a used 4090 is the value pick and a 5090 is the capacity pick. Selling a 4090 at the $2,731 median and buying a $6,549.99 5090 costs about $3,800 for roughly 1.5x decode on dense models, 2x batched throughput and 8 GB more memory. That is a fair trade for a team serving a 27B to 35B model to several users, and a poor one for one person chatting with a 14B model. The modded 48 GB RTX 4090 sold by some Chinese shops is a separate gamble, covered in RTX 4090 48 GB.

Used-card risk is the other side. A 4090 bought in November 2026 has been out of US retail for more than a year, and both cards use the 16-pin connector family that Wikipedia's RTX 40 and RTX 50 articles document melting reports for, so inspect the connector and ask how the card was run. If neither card fits the model you care about, a 128 GB machine is the different class of answer, compared in DGX Spark vs RTX 5090, and the GPU buyer's guide walks through the capacity-versus-bandwidth choice. To check whether 32 GB changes anything for your workload before spending $6,550, GPUwerk rents an RTX 5090 by the minute.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• GeForce RTX 5090 graphics card specifications, NVIDIA
• GeForce RTX 4090 graphics card specifications, NVIDIA
• Compare GeForce graphics cards, NVIDIA
• NVIDIA GeForce RTX 4090 Founders Edition Review (specifications), The FPS Review, 2022-10-11
• Introducing NVFP4 for Efficient and Accurate Low-Precision Inference, NVIDIA Technical Blog, 2025-06-24
• Quantization: supported hardware, vLLM on GitHub
• Performance of llama.cpp on Nvidia CUDA (discussion #15013), ggml-org/llama.cpp on GitHub, 2025-08-01
• RTX 5090 LLM benchmarks, Hardware Corner, 2026-03
• RTX 4090 LLM benchmarks, Hardware Corner, 2026-03
• Best Local LLMs for RTX 5090: Benchmarks, Atomic Chat, 2026-09-18
• Best Local LLMs for RTX 4090: Benchmarks, Atomic Chat, 2026-09-21
• Benchmarking RTX GPUs for LLM inference (RTX 4090 vs RTX 5090 vs RTX PRO 6000), CloudRift, 2025-10-09
• llm-performance-tests, DeutscheKI on GitHub
• NVIDIA RTX 5090 in stock tracker, NowInStock, 2026-11-07
• NVIDIA RTX 4090 in stock tracker, NowInStock
• RTX 4090 price, used and specs, pcprice.watch, 2026-11-07
• RTX 5090 price, used and specs, pcprice.watch, 2026-11-07
• Used RTX 4090 in Germany, pcprice.watch, 2026-11-07
• Used RTX 5090 in Germany, pcprice.watch, 2026-11-07
• INNO3D GeForce RTX 5090 X3, Geizhals, 2026-11-07
• NVIDIA GeForce RTX 4090 Founders Edition, Geizhals, 2026-11-07
• Used RTX 4090 prices, Dealhunter, 2026-11-07
• GeForce RTX 50 series, Wikipedia
• GeForce RTX 40 series, Wikipedia
• Electricity price statistics, Eurostat, 2026-10