diff --git a/baml_language/crates/baml_tests/Cargo.toml b/baml_language/crates/baml_tests/Cargo.toml
index c733eac42df..6e6484d0f2d 100644
--- a/baml_language/crates/baml_tests/Cargo.toml
+++ b/baml_language/crates/baml_tests/Cargo.toml
@@ -69,6 +69,10 @@ harness = false
 name = "runtime_benchmark"
 harness = false
 
+[[bench]]
+name = "profiling_overhead"
+harness = false
+
 [[bench]]
 name = "cache_profile"
 harness = false
diff --git a/baml_language/crates/baml_tests/benches/compiler_benchmark.rs b/baml_language/crates/baml_tests/benches/compiler_benchmark.rs
index b825267a83f..6916b82d8d1 100644
--- a/baml_language/crates/baml_tests/benches/compiler_benchmark.rs
+++ b/baml_language/crates/baml_tests/benches/compiler_benchmark.rs
@@ -28,6 +28,15 @@ fn main() {
         eprintln!("Skipping compiler_benchmark in debug/test profile.");
         return;
     }
+    // Hermetic wall-time: pin BAML profiling OFF unless the caller explicitly
+    // chose a value, so compile timings don't include the default-ON tracing
+    // pipeline. See runtime_benchmark.rs for the full rationale; tracing cost
+    // is measured deliberately in the `profiling_overhead` bench target.
+    if std::env::var_os("BAML_PROFILE").is_none() {
+        // SAFETY: single-threaded at the very top of main, before any engine
+        // or divan code reads the environment.
+        unsafe { std::env::set_var("BAML_PROFILE", "0") };
+    }
     divan::main();
 }
 
diff --git a/baml_language/crates/baml_tests/benches/profiling_overhead.rs b/baml_language/crates/baml_tests/benches/profiling_overhead.rs
new file mode 100644
index 00000000000..04d2998e3cc
--- /dev/null
+++ b/baml_language/crates/baml_tests/benches/profiling_overhead.rs
@@ -0,0 +1,139 @@
+//! Profiling (BEX tracing) overhead benchmarks.
+//!
+//! Run with: cargo bench --bench profiling_overhead
+//!
+//! The ordinary wall-time benches (`runtime_benchmark`, `compiler_benchmark`)
+//! pin `BAML_PROFILE=0` so their numbers are hermetic. This target is the
+//! deliberate counterpart: the same VM harness with profiling pinned ON, over a
+//! small fixed workload subset chosen to characterize tracing cost —
+//! per-call ring-pair overhead (`pure_call_1m`), allocation-heavy loops
+//! (`array_build_sum_100k`), the known ring-overflow reproducer
+//! (`fib32_recursive`, which aborted at the default 1 GiB cap on bare metal),
+//! and a string baseline (`concat_loop_10k`).
+//!
+//! Overhead ratio = this target's median ÷ the same workload's median in
+//! `runtime_benchmark`. The subset sources come from the same speedtest corpus
+//! export as the main benches (single source of truth; see `build.rs`).
+//!
+//! Environment pins (each only when the caller hasn't set a value):
+//! - `BAML_PROFILE=1` — profiling explicitly ON.
+//! - `BAML_RING_MAX_OVERFLOW_BYTES=4294967296` — 4 GiB headroom so hot
+//!   workloads measure tracing cost instead of aborting on the 1 GiB default.
+//! - `BAML_PROFILE_DIR=<fresh temp dir>` — events sink to a throwaway
+//!   location instead of the working directory.
+
+use std::{path::Path, sync::Arc};
+
+use baml_compiler2_emit::{CompileOptions, generate_project_bytecode};
+use baml_project::ProjectDatabase;
+use bex_engine::{BexEngine, FunctionCallContextBuilder};
+use divan::{Bencher, black_box};
+use sys_native::{CallId, SysOpsExt};
+
+fn main() {
+    if cfg!(debug_assertions) {
+        eprintln!("Skipping profiling_overhead in debug/test profile.");
+        return;
+    }
+    if std::env::var_os("DIVAN_MAX_TIME").is_none() {
+        // SAFETY: single-threaded here at the very top of main, before divan
+        // reads its args/env. No other thread can observe the environment.
+        unsafe { std::env::set_var("DIVAN_MAX_TIME", "2") };
+    }
+    if std::env::var_os("BAML_PROFILE").is_none() {
+        // SAFETY: as above.
+        unsafe { std::env::set_var("BAML_PROFILE", "1") };
+    }
+    if std::env::var_os("BAML_RING_MAX_OVERFLOW_BYTES").is_none() {
+        // SAFETY: as above.
+        unsafe { std::env::set_var("BAML_RING_MAX_OVERFLOW_BYTES", "4294967296") };
+    }
+    // Drop-backed so the sink and its artifacts are removed when the run ends;
+    // only used when the caller didn't choose a sink of their own.
+    let _sink_dir = if std::env::var_os("BAML_PROFILE_DIR").is_none() {
+        let dir = tempfile::Builder::new()
+            .prefix("baml-profiling-overhead-bench-")
+            .tempdir()
+            .expect("failed to create profile sink dir");
+        // SAFETY: as above.
+        unsafe { std::env::set_var("BAML_PROFILE_DIR", dir.path()) };
+        Some(dir)
+    } else {
+        None
+    };
+    divan::main();
+}
+
+// ============================================================================
+// Helpers — duplicated from runtime_benchmark.rs (bench targets are separate
+// binaries and cannot share modules without moving these into the crate's lib)
+// ============================================================================
+
+/// Compile BAML source into a ready-to-run engine.
+fn compile_source(source: &str) -> (ProjectDatabase, BexEngine) {
+    let mut db = ProjectDatabase::new();
+    db.set_project_root(Path::new("."));
+    db.add_file("bench.baml", source);
+    let bytecode = generate_project_bytecode(
+        &db,
+        &CompileOptions {
+            emit_test_cases: false,
+        },
+    )
+    .expect("benchmark compilation failed");
+    let engine = BexEngine::new(bytecode, Arc::new(sys_native::SysOps::native()), vec![])
+        .expect("benchmark engine creation failed");
+    (db, engine)
+}
+
+/// Compile once, then measure only the cost of calling `main()` — identical
+/// harness to `runtime_benchmark::bench_vm_main` so medians divide cleanly.
+/// Workload sources are empty when the speedtest corpus was unavailable at
+/// build time; the bench then skips itself.
+fn bench_vm_main(bencher: Bencher, source: &str) {
+    if source.is_empty() {
+        eprintln!("speedtest corpus unavailable at build time; skipping");
+        return;
+    }
+    let (_db, engine) = compile_source(source);
+    let engine = Arc::new(engine);
+    let rt = tokio::runtime::Runtime::new().expect("failed to build tokio runtime");
+    bencher.bench(|| {
+        black_box(
+            rt.block_on(engine.call_function(
+                "main",
+                vec![],
+                FunctionCallContextBuilder::new(CallId::next()).build(),
+                true,
+            ))
+            .expect("benchmark execution failed"),
+        )
+    });
+}
+
+// Subset sources exported by build.rs from the speedtest corpus.
+include!(concat!(env!("OUT_DIR"), "/speedtest_profiling_sources.rs"));
+
+/// compute::pure call 1m — per-call CallFunction/EndFunction ring-pair cost.
+#[divan::bench]
+fn prof_on_compute_pure_call_1m(bencher: Bencher) {
+    bench_vm_main(bencher, PROF_SRC_COMPUTE_PURE_CALL_1M);
+}
+
+/// compute::array build sum 100k — allocation-heavy loop under tracing.
+#[divan::bench]
+fn prof_on_compute_array_build_sum_100k(bencher: Bencher) {
+    bench_vm_main(bencher, PROF_SRC_COMPUTE_ARRAY_BUILD_SUM_100K);
+}
+
+/// compute::fib32 recursive — the historical ring-overflow reproducer.
+#[divan::bench]
+fn prof_on_compute_fib32_recursive(bencher: Bencher) {
+    bench_vm_main(bencher, PROF_SRC_COMPUTE_FIB32_RECURSIVE);
+}
+
+/// string::concat loop 10k — string-path baseline under tracing.
+#[divan::bench]
+fn prof_on_string_concat_loop_10k(bencher: Bencher) {
+    bench_vm_main(bencher, PROF_SRC_STRING_CONCAT_LOOP_10K);
+}
diff --git a/baml_language/crates/baml_tests/benches/runtime_benchmark.rs b/baml_language/crates/baml_tests/benches/runtime_benchmark.rs
index 896c3c94d92..0c49d516346 100644
--- a/baml_language/crates/baml_tests/benches/runtime_benchmark.rs
+++ b/baml_language/crates/baml_tests/benches/runtime_benchmark.rs
@@ -39,6 +39,17 @@ fn main() {
         // reads its args/env. No other thread can observe the environment.
         unsafe { std::env::set_var("DIVAN_MAX_TIME", "2") };
     }
+    // Hermetic wall-time: BAML profiling ships default-ON, which both skews
+    // per-call timings and can abort hot workloads outright when the event
+    // ring's overflow cap is hit (compute::fib32_recursive reproduced this at
+    // the 1 GiB cap on bare metal). Pin it OFF unless the caller explicitly
+    // chose a value; tracing cost is measured deliberately in the
+    // `profiling_overhead` bench target instead of contaminating every number
+    // here.
+    if std::env::var_os("BAML_PROFILE").is_none() {
+        // SAFETY: as above — single-threaded, before any engine reads env.
+        unsafe { std::env::set_var("BAML_PROFILE", "0") };
+    }
     divan::main();
 }
 
diff --git a/baml_language/crates/baml_tests/build.rs b/baml_language/crates/baml_tests/build.rs
index d3f675db72b..1a772887b7d 100644
--- a/baml_language/crates/baml_tests/build.rs
+++ b/baml_language/crates/baml_tests/build.rs
@@ -91,12 +91,22 @@ fn generate_speedtest_benches(manifest_dir: &str) {
         Ok(w) => w,
         Err(e) => {
             println!("cargo:warning=speedtest benches disabled: {e}");
-            // Emit an empty (but valid) file so the include! in the bench compiles.
+            // Emit empty (but valid) files so the include!s in both bench
+            // targets compile.
             fs::write(
                 &dest_path,
                 "// speedtest workloads unavailable at build time; no benches generated.\n",
             )
             .unwrap();
+            fs::write(
+                Path::new(&out_dir).join("speedtest_profiling_sources.rs"),
+                "// speedtest workloads unavailable at build time.\n\
+                 pub const PROF_SRC_COMPUTE_PURE_CALL_1M: &str = \"\";\n\
+                 pub const PROF_SRC_COMPUTE_ARRAY_BUILD_SUM_100K: &str = \"\";\n\
+                 pub const PROF_SRC_COMPUTE_FIB32_RECURSIVE: &str = \"\";\n\
+                 pub const PROF_SRC_STRING_CONCAT_LOOP_10K: &str = \"\";\n",
+            )
+            .unwrap();
             return;
         }
     };
@@ -148,6 +158,41 @@ fn generate_speedtest_benches(manifest_dir: &str) {
 ";
 
     write_formatted_code(&dest_path, benches, header);
+
+    // Also emit the small fixed subset consumed by the `profiling_overhead`
+    // bench target: same single source of truth, but only the workloads chosen
+    // to characterize tracing cost (per-call ring pairs, allocation-heavy
+    // loops, the known ring-overflow reproducer, and a string baseline).
+    let subset = [
+        "compute::pure call 1m",
+        "compute::array build sum 100k",
+        "compute::fib32 recursive",
+        "string::concat loop 10k",
+    ];
+    let prof_consts: TokenStream = subset
+        .iter()
+        .map(|name| {
+            let slug = slugify(name);
+            let ident = format_ident!("PROF_SRC_{}", slug.to_uppercase());
+            let source = all_workloads
+                .iter()
+                .find(|w| w.name == *name)
+                .map(|w| w.baml.as_str())
+                .unwrap_or("");
+            quote! {
+                pub const #ident: &str = #source;
+            }
+        })
+        .collect();
+    let prof_path = Path::new(&out_dir).join("speedtest_profiling_sources.rs");
+    write_formatted_code(
+        &prof_path,
+        prof_consts,
+        "// Auto-generated profiling-overhead workload sources by build.rs.\n\
+         // Source of truth: tools/speedtest/workloads/*.md (expanded via export_baml.py).\n\
+         // An empty const means the corpus was unavailable at build time; the\n\
+         // corresponding bench skips itself.\n",
+    );
 }
 
 /// Run `export_baml.py` and parse its JSON output into the workload list.