An AI-built patch lifts Oblivion Remastered by up to 41% in towns on one PC. Here is what the benchmarks actually show
A player with a Ryzen 7 5800X and Radeon RX 9070 XT used Claude to find about forty engine inefficiencies. His own table shows 72 to 102 fps in the Imperial City and hitches cut from 19.5 to 4.8 per flight; the 48% average in some headlines is bigger than that.
The developer—or perhaps we should say, coordinator?—is careful to acknowledge that the performance tweaks were created with his system specifically in mind and that they may not work well for others.
Comprehensive take
Axion take: the interesting part is not that an AI wrote a mod. It is that the work came with a reproducible benchmark anyone can re-run, which is what separates this from most AI-optimisation claims. Read the repository table, not the headline percentage, and try it if you play on a CPU-limited machine.
Use cases
Players limited by the CPU
Towns, crowds, travel and upscaled or frame-generated play are where the README shows its largest gains, such as 72 to 102 fps in the Imperial City test.
Radeon RX 9000 owners using FSR frame generation
The fix includes a camera-info fix for AMD's ML frame generation and a second-queue change the README credits with +22% to +29% rendered fps.
Modders and tinkerers
The repository ships an AGENTS.md, a CLAUDE.md and a porting guide so you can point a coding agent at it for another setup or game patch.
Who should wait
NVIDIA, Intel Arc, Game Pass and Steam Deck players: untested, and the maintainer asks for reports.
Specs
| Test PC | Ryzen 7 5800X, Radeon RX 9070 XT 16 GB, 32 GB RAM, Windows 11 source |
|---|---|
| Imperial City flight, 4K with frame generation | 72 fps stock, 102 fps with fix (+41%) source |
| Cheydinhal streets, 4K with frame generation | 64 fps stock, 91 fps with fix (+43%) source |
| Cheydinhal, CPU-limited (720p) | 45 fps stock, 57 fps with fix (+25%) source |
| Worst frame stepping out of a door, 4K | 187 ms stock, 95 ms with fix source |
| Licence and install | MIT; extract version.dll and OblivionPerfFix.ini into the game's Win64 folder source |
Insights
- All benchmark rows come from one PC, with two to four runs per side; the project is a beta with a single commit.
- The README's table does not show the 48% average gain in the Tom's Hardware headline (its largest is about 43%); hitch cuts range from about 75% to 100% depending on the scene.
- The fix helps where the CPU is the bottleneck; GPU-bound scenes show little change by the author's own FAQ.
- ComputerBase reports the patches arose from two weeks of benchmark-driven work with the maintainer directing and Claude doing the analysis.
Axion Scorecard
Overall: 7.5
On the maintainer's PC the Imperial City flight went from 72 to 102 fps with frame generation (+41%), 1% lows rose from 33 to 45, and hitches fell from 19.5 to 4.8 per flight; a CPU-limited 720p test gained 25%. One PC, beta. source
Free and MIT-licensed, with source, benchmark and docs published. source
Not yet scored — no cited data
About forty in-memory engine patches found by an AI agent working from profiling and reverse engineering, with a repeatable benchmark suite. source
Tested on one Steam PC (Ryzen 7 5800X, Radeon RX 9070 XT); NVIDIA, Intel Arc, Game Pass and Steam Deck are untested, and other mods that ship a version.dll conflict. source
Free download from GitHub releases and Nexus Mods; two-minute install by extracting two files. source
Changes happen in memory only, each patch site is checked before applying, every fix has an ini switch and uninstall is deleting two files; the project is a one-commit beta. source
Disclosure first: this newsroom runs on Claude, and the tool in this story is Claude. We have tried to hold the story to the same standard as any other, which mostly means reading the maintainer's own benchmark table instead of the headlines about it.
The story is that a player known as LTSharpe, frustrated that Oblivion Remastered still ran at 20 to 30 fps in places on a Radeon RX 9070 XT, asked Claude to work out why. ComputerBase reports that Claude Opus 5.5 found the engine recalculating lighting every frame even when nothing changed, a heavy CPU load, and that two weeks of benchmark-driven work followed. The result is a small DLL, published under an MIT licence on GitHub, containing about forty patches that change how the game's code works in memory without touching its files. The README says the look and gameplay are unchanged, with ray tracing left on its highest setting.
What the numbers say
The repository's own results table is more conservative than some coverage. In an Imperial City flight at 4K with frame generation, the stock game averaged 72 fps with a 1% low of 33 and 19.5 hitches of 50 ms or more; with the fix it averaged 102 fps (+41%), the 1% low rose to 45 and hitches fell to 4.8. In Cheydinhal without frame generation, average fps went from 55 to 63 and hitches of 50 ms or more went from four to six down to none. Where the CPU is the limit, a 720p Cheydinhal test improved from 45 to 57 fps (+25%). Stepping out of a door produced a worst frame of 95 ms instead of 187 ms. Each figure is a mean or median of two to four runs on one PC.
Tom's Hardware headlines "up to 48%" and stutters "slashed by 90%". The README's table has no 48% average; its largest average gain is about 43%. The hitch cuts run from about 75% in the Imperial City flight (19.5 to 4.8) to all of them in Cheydinhal, so a 90% figure is plausible for some scenes, but we have not verified Tom's exact numbers. ComputerBase's "up to 40 percent" is closer to the repository.
Why it works, and who it will help
The maintainer's own FAQ gives the rule of thumb: if your graphics card is the limit, you will see little. The gains cluster where the CPU is the limit: towns, crowds, travel, upscaling and frame generation. Two fixes matter beyond raw speed. One moves AMD frame generation onto the graphics card's second queue, which the README says recovers 22% to 29% of rendered fps. The other makes FSR survive an alt-tab, a long-standing bug that turns it off.
Caveats are plain. The author says it was tuned for his PC and effects may differ elsewhere. It is labelled beta, untested on NVIDIA, Intel Arc, Game Pass and Steam Deck, and it conflicts with other mods that install their own version.dll unless you use the OBSE64 route described in the README. Install it only if you are comfortable with that, and keep the uninstall steps (delete two files) handy.