Google says its Antigravity /teamwork AI system rewrote libgav1, its AV1 video decoder, in Rust and made it 20–30% faster than the original C++ and hand-optimized assembly code on sample video clips. The result, disclosed in a November 13 engineering update, followed a port that initially ran around four times slower than the C++ version.
The company reports around 170,000 lines of Rust and 100% bit-exact parity across a broad conformance test suite, meaning the rewritten decoder matched the originals output in those tests.
The performance reversal came after /teamwork received profiling tools and a goal of maximizing speed on sample clips, Google says. It reports the advantage on both x86_64 and ARM.
Related Company Google An American multinational technology company
Google preserves behavior before tuning speed
The migration separates two engineering problems: keeping an established program's behavior and making the replacement fast enough to use. Google's account illustrates why a successful translation alone does not settle the performance question.
Matching output and improving speed are separate checks
Google reports a Rust rewrite of libgav1. Explore the output comparison and the profiling loop behind its reported performance result.
Google reports migrating libgav1 from C++ to RustLIBGAV1 / GOOGLE'S REPORTED MIGRATIONC++Original decoder
/teamwork rewritesRustAround 170,000 lines
Google reports migrating libgav1 from C++ to RustLIBGAV1 / REPORTED MIGRATIONC++Original decoder
Rust~170k linesRewritten with /teamwork
Explore what each check measuresExplore what each check measuresMatching outputMeasuring speed
Google reports complete bit-exact parity across a broad conformance test suite.
Conformance tests compare original and Rust decoder output; Google reports full parity within the suiteCOMPARE THE DECODED OUTPUTOriginal decoderRust port
Conformancetest suite
100%bit-exact parityMatching output in the reported suite; no claim of all-input correctness.
Conformance tests compare original and Rust decoder output; Google reports full parity within the suiteCOMPARE DECODED OUTPUTOriginaldecoderRust port
Conformance test suite
100%bit-exact parity in the suiteUntested inputs remain outside this result.
The Rust port began slower; profiling and optimization on sample clips led to a reported advantage over C++ and assembly.
Profiling and sample-clip optimization led to Google's reported 20 to 30 percent speed advantagePROFILE, OPTIMIZE, RUN AGAINInitial Rust port: around 4x slower than C++Profiling toolsOptimize Rust
Run sample clips
20–30%fasterthan C++ +optimized assembly
- Reported on x86_64 and ARMSample-clip result; the full benchmark configuration is not disclosed.
Profiling and sample-clip optimization led to Google's reported 20 to 30 percent speed advantagePROFILE, OPTIMIZE, RUN AGAINInitial Rust port: around 4x slowerthan the original C++ProfilingtoolsOptimizeRust
Run sample clips20–30% fasterthan C++ + optimized assemblyReported on x86_64 and ARM
- Sample clips; full benchmarkconfiguration not disclosed.
Matching a test suite answers a different question from measuring speed on sample video clips.
Google's November 13, 2026 account. Conformance parity is limited to its reported suite; speed is a sample-clip comparison. Hardware models, build settings and clip identities are not disclosed. These results do not establish universal correctness, security or production deployment.
Sources and evidence
Sources
- Google Research engineering report
- Reported statement Google says /teamwork rewrote its libgav1 AV1 video decoder from C++ into around 170,000 lines of Rust. [1]
- Reported statement Google reports 100% bit-exact parity for the Rust libgav1 port across a broad conformance test suite. [1]
- Reported statement Google says the Rust libgav1 port initially ran around four times slower than its original C++ code. [1]
- Reported statement Google reports a Rust libgav1 port 20–30% faster than the original C++ and hand-optimized assembly on sample video clips on x86_64 and ARM after profiling; the report does not specify processor models, compiler settings, clip identities, thread counts or timing distributions. [1]
Supporting facts
Method
Schematic migration and two selectable evaluations. Connectors show conceptual comparisons and a profiling feedback loop, not measured durations or a published test-harness topology. Values are reported by Google; no ratios or speed estimates were calculated.
Scope
Google's November 13, 2026 account. Conformance parity is limited to its reported suite; speed is a sample-clip comparison. Hardware models, build settings and clip identities are not disclosed. These results do not establish universal correctness, security or production deployment.
Infographic by CryptoSlate
For code migrations, Google describes a forthcoming general rewrite pattern that schedules each module only after its dependencies compile, then tests leaf modules against the original, unmodified test suite. Keeping those checks intact gives the agents a fixed target while they change the implementation. Profiling then provides a separate feedback loop for finding where the new version spends its time.
Google places this work within collaborations with human experts. Researchers frame questions and define constraints, while agents explore possibilities, test candidate solutions and check results.
The decoder work also predates this month's update. In its September 30 Argon announcement, Google described agents optimizing an existing Rust libgav1 port to run 2.7 times faster than that earlier Rust version, bringing it closer to optimized C++. November's comparison is against C++ and assembly instead. The different baselines prevent treating those figures as one measured progression.
CryptoSlate's recent AWS Emulator coverage concerned agent-written software and human-designed release checks. Google's decoder example adds the challenge of preserving an existing implementation's outputs while improving its speed.
The libgav1 project repository linked by Google still documents a C++17 and CMake build, including architecture-specific optimization switches. The November report does not specify processor models, compiler settings, clip identities, thread counts or timing distributions. It also does not identify a reproduction package for the Rust benchmark or announce that the rewritten decoder has entered production.
That leaves a concrete next test for the disclosure: a specified Rust revision, comparison build and video corpus would let other developers examine whether the reported advantage holds under their own conditions.

