diff --git a/src/intervals/exact_literals.jl b/src/intervals/exact_literals.jl
index 013f0eea..d240db3f 100644
--- a/src/intervals/exact_literals.jl
+++ b/src/intervals/exact_literals.jl
@@ -261,6 +261,57 @@ for f ∈ (:+, :-, :*, :/, :\, :^)
     @eval Base.$f(x::ExactReal, y::BareInterval) = $f(promote(x, y)...)
 end
 
+# These specializations improve efficiency, but always agree precisely with
+# the promoting fallbacks above. Several circumstances fall back to promotion: a
+# non-finite point (whose thin interval is not even valid), an unbounded interval
+# times a point (`0 * ∞` is flavor dependent), and division by zero (likewise).
+# The type parameter is bounded so that these methods are more specific than the
+# generic `Base.$f(::ExactReal, ::BareInterval)`, whose own parameters are bounded
+# by `Real` and `NumTypes`.
+
+function Base.:+(x::BareInterval{T}, y::ExactReal{T}) where {T<:NumTypes}
+    v = y.value
+    isfinite(v) || return +(promote(x, y)...)
+    isempty_interval(x) && return x
+    return @round(T, inf(x) + v, sup(x) + v)
+end
+Base.:+(x::ExactReal{T}, y::BareInterval{T}) where {T<:NumTypes} = y + x
+
+function Base.:-(x::BareInterval{T}, y::ExactReal{T}) where {T<:NumTypes}
+    v = y.value
+    isfinite(v) || return -(promote(x, y)...)
+    isempty_interval(x) && return x
+    return @round(T, inf(x) - v, sup(x) - v)
+end
+
+function Base.:-(x::ExactReal{T}, y::BareInterval{T}) where {T<:NumTypes}
+    v = x.value
+    isfinite(v) || return -(promote(x, y)...)
+    isempty_interval(y) && return y
+    return @round(T, v - sup(y), v - inf(y))
+end
+
+function Base.:*(x::BareInterval{T}, y::ExactReal{T}) where {T<:NumTypes}
+    v = y.value
+    isempty_interval(x) && return x # `isbounded` is true for the empty interval
+    (isfinite(v) & isbounded(x)) || return *(promote(x, y)...)
+    signbit(v) && return @round(T, v * sup(x), v * inf(x))
+    return @round(T, v * inf(x), v * sup(x))
+end
+Base.:*(x::ExactReal{T}, y::BareInterval{T}) where {T<:NumTypes} = y * x
+
+function Base.:/(x::BareInterval{T}, y::ExactReal{T}) where {T<:NumTypes}
+    v = y.value
+    isempty_interval(x) && return x
+    (isfinite(v) & !iszero(v)) || return /(promote(x, y)...)
+    signbit(v) && return @round(T, sup(x) / v, inf(x) / v)
+    return @round(T, inf(x) / v, sup(x) / v)
+end
+
+# `x \ y` is `y / x`; the reverse case is a point divided by an interval, which
+# has no shortcut — the zero-crossing tree of `/` is the whole computation.
+Base.:\(x::ExactReal{T}, y::BareInterval{T}) where {T<:NumTypes} = y / x
+
 """
     has_exact_display(x::Real)
 
diff --git a/src/intervals/intervals.jl b/src/intervals/intervals.jl
index 280002ea..053952a1 100644
--- a/src/intervals/intervals.jl
+++ b/src/intervals/intervals.jl
@@ -5,12 +5,15 @@ include("construction.jl")
 include("parsing.jl")
 include("real_interface.jl")
     export numtype
-include("exact_literals.jl")
-    export ExactReal, exact, @exact, has_exact_display
 
 # Rounding
 include("rounding.jl")
 
+# after `rounding.jl`: the interval operations on `ExactReal` round their
+# endpoints with `@round`
+include("exact_literals.jl")
+    export ExactReal, exact, @exact, has_exact_display
+
 # Flavor
 include("flavor.jl")
 
diff --git a/test/interval_tests/exact_literals.jl b/test/interval_tests/exact_literals.jl
index 5c76832b..b740aa68 100644
--- a/test/interval_tests/exact_literals.jl
+++ b/test/interval_tests/exact_literals.jl
@@ -98,3 +98,59 @@
         @test val2 == 3.0
     end
 end
+
+@testset "Exact literals with bare intervals" begin
+    # `+`, `-`, `*`, `/` and `\` on a matching number type bypass the promotion to
+    # a thin interval, so they must be more specific than the generic methods and
+    # must return exactly what those return.
+    specialized(f, S, R) = which(f, Tuple{S,R}).sig isa UnionAll
+    @test all(f -> specialized(f, BareInterval{Float64}, ExactReal{Float64}), (+, -, *, /))
+    @test all(f -> specialized(f, ExactReal{Float64}, BareInterval{Float64}), (+, -, *, \))
+
+    # ... and the mixed number types must not, since the promotion is where the
+    # `ExactReal` gets rounded to the interval's number type.
+    @test !specialized(*, BareInterval{Float64}, ExactReal{Int})
+
+    viapromotion(f, x, y) = f(promote(x, y)...)
+    samebits(x::BareInterval, y::BareInterval) = (inf(x) === inf(y)) & (sup(x) === sup(y))
+
+    for T ∈ (Float64, Float32, Rational{Int})
+        vals = T <: Rational ?
+            (zero(T), one(T), -one(T), T(1//2), T(-5//3)) :
+            (zero(T), -zero(T), one(T), -one(T), T(0.5), T(-2.5), T(0.1), T(-0.1),
+             floatmax(T), floatmin(T))
+        ivs = (bareinterval(T, 1, 2), bareinterval(T, -2, -1), bareinterval(T, -1, 2),
+               bareinterval(T, 0, 1), bareinterval(T, -1, 0), bareinterval(T, 0, 0),
+               bareinterval(T(1//10), T(1//10)), bareinterval(T, -Inf, 2),
+               bareinterval(T, 3, Inf), entireinterval(BareInterval{T}),
+               emptyinterval(BareInterval{T}))
+        for v ∈ vals, x ∈ ivs
+            k = exact(v)
+            @test samebits(x + k, viapromotion(+, x, k))
+            @test samebits(k + x, viapromotion(+, k, x))
+            @test samebits(x - k, viapromotion(-, x, k))
+            @test samebits(k - x, viapromotion(-, k, x))
+            @test samebits(x * k, viapromotion(*, x, k))
+            @test samebits(k * x, viapromotion(*, k, x))
+            @test samebits(x / k, viapromotion(/, x, k))
+            @test samebits(k \ x, viapromotion(\, k, x))
+        end
+    end
+
+    # A non-finite `ExactReal` has no thin interval; both routes warn and return
+    # the empty interval.
+    let x = bareinterval(1.0, 2.0)
+        for v ∈ (Inf, -Inf, NaN)
+            @test isempty_interval(@test_logs (:warn,) x + exact(v))
+            @test isempty_interval(@test_logs (:warn,) x - exact(v))
+            @test isempty_interval(@test_logs (:warn,) exact(v) - x)
+            @test isempty_interval(@test_logs (:warn,) x * exact(v))
+            @test isempty_interval(@test_logs (:warn,) x / exact(v))
+        end
+    end
+
+    # Mixed number types still round through the promotion.
+    @test isequal_interval(bareinterval(1.0, 2.0) * exact(2), bareinterval(2.0, 4.0))
+    @test isequal_interval(bareinterval(1.0, 2.0) + exact(1//3),
+                           bareinterval(1.0, 2.0) + bareinterval(Float64, 1//3))
+end