diff --git a/README.md b/README.md
index dd4dab68..f6e0d9c3 100644
--- a/README.md
+++ b/README.md
@@ -180,35 +180,25 @@ Output on a modern terminal with Unicode support:
 
 ![image](https://github.com/fmtlib/fmt/assets/576385/2a93c904-d6fa-4aa6-b453-2618e1c327d7)
 
-# Benchmarks
+# Performance
 
-## Speed tests
+{fmt} can be tens of percent to 20–30 times faster than iostreams and
+`sprintf`, especially for numeric formatting. It minimizes dynamic memory
+allocations and can optionally [compile format strings](
+https://fmt.dev/latest/api/#compile-api) into efficient formatting code.
 
-| Library           | Method        | Run Time, s |
-|-------------------|---------------|-------------|
-| libc              | printf        |   0.66      |
-| libc++            | std::ostream  |   1.63      |
-| {fmt} 12.1        | fmt::print    |   0.44      |
-| Boost Format 1.88 | boost::format |   3.89      |
-| Folly Format      | folly::format |   1.28      |
+See [format-benchmark](https://github.com/fmtlib/format-benchmark) and
+[dtoa-benchmark](https://github.com/fmtlib/dtoa-benchmark) for benchmarks
+and methodology.
 
-{fmt} is the fastest of the benchmarked methods, \~50% faster than
-`printf`.
+**Time per double (smaller is better):**
 
-The above results were generated by building `tinyformat_test.cpp` on
-macOS 15.6.1 with `clang++ -O3 -DNDEBUG -DSPEED_TEST -DHAVE_FORMAT`, and
-taking the best of three runs. In the test, the format string
-`"%0.10f:%04d:%+g:%s:%p:%c:%%\n"` or equivalent is filled 2,000,000
-times with output sent to `/dev/null`; for further details refer to the
-[source](https://github.com/fmtlib/format-benchmark/blob/master/src/tinyformat-test.cc).
+[![Time per double for floating-point formatting methods](
+https://github.com/user-attachments/assets/3678bc4a-9405-489e-8ce1-ca702829cdaa)](
+https://github.com/fmtlib/dtoa-benchmark)
 
-{fmt} is up to 20-30x faster than `std::ostringstream` and `sprintf` on
-IEEE754 `float` and `double` formatting
-([dtoa-benchmark](https://github.com/fmtlib/dtoa-benchmark)) and faster
-than [double-conversion](https://github.com/google/double-conversion)
-and [ryu](https://github.com/ulfjack/ryu):
-
-[![image](https://user-images.githubusercontent.com/576385/95684665-11719600-0ba8-11eb-8e5b-972ff4e49428.png)](https://fmt.dev/unknown_mac64_clang12.0.html)
+`ostringstream` and `sprintf` are omitted because they are an order of
+magnitude slower than the other methods.
 
 ## Compile time and code bloat