Simon Willison uses Claude Opus 5.5 to build Chirp Pong, a two-phone game that syncs over audio chirps

• Model: Claude Opus 5.5 in Claude Code. The label "Extra" is Willison's. The Batch lists the effort levels as low, medium, high, xhigh and max, so "Extra" probably means xhigh, but that is not confirmed.
• Access: a free static web page at tools.simonwillison.net/chirp-pong, with the code in the simonw/tools repo. It needs microphone permission. Players put the phones top-to-top a few centimetres apart with volume at maximum.
• Gameplay twists: points are worth ×2 after 6 hits in a rally and ×3 after 12. Players can also send "trick gems" and bombs that stun the other player's paddle.

From the source file:

• Modulation: 16-ary FSK. Each player transmits in its own band, roughly 1.4–3.0 kHz or 3.4–5.0 kHz, using 18 tones spaced 46.875 Hz apart. Each symbol lasts 50 ms, and each frame starts with a two-symbol sync tone. Consecutive tones always differ, so the receiver can find symbol boundaries without a shared clock.
• Receiver: the code switches off the browser's echo cancellation, noise suppression and automatic gain control. It runs a 1024-point analysis at 48 kHz and picks the strongest tone in each symbol. It tries up to five timing guesses, and the first one that passes the checksum wins.
• Reliability: each frame carries a CRC-8 checksum. Before sending, a phone checks that the other band is quiet. Important messages (ball handoff, scores, bombs) use stop-and-wait acknowledgments and are resent about every second until acknowledged.
• State model: each phone runs its own game. Only small events cross the gap: a ball handoff packs position, angle, speed, spin, trick type and rally count into a few dozen bits. Each miss message carries both players' full scores, so the phones can detect when they disagree.

• The game works on two phones over audio. Only Willison reports this, in his post. There is no independent playtest, no measured error rate and no record of which devices it was tried on.
• Claude Opus 5.5 wrote it. This rests on Willison's attribution. The post does not link a coding transcript, so it is unclear how much he reworked the code.
• The protocol design described above is in the code. Anyone can check this by reading the published file.
• Opus 5.5 leads on agentic coding. Anthropic positions it this way. Third-party indexes from Artificial Analysis and Vals AI rank it first overall, as summarized by The Batch, but none of these benchmarks test this kind of task.

• Sending data over sound is an old idea. ggwave is an open-source library that uses multi-tone FSK with Reed-Solomon error correction at 8–16 bytes/sec. It offers audible and ultrasonic modes and runs in the browser via WebAssembly. It uses the same 46.875 Hz tone spacing as Chirp Pong, which is what a 1024-point analysis at 48 kHz gives.
• The Chirp company) grew out of University College London research on "sonic barcodes."
• Chirp Pong includes no outside library. Its one new element is pairing a low-bandwidth audio link with game design that only sends sparse events.

• Distance and noise: the game needs the phones a few centimetres apart at full volume. It uses audible tones and no forward error correction, so noisy rooms could break it. ggwave's Reed-Solomon coding is the obvious contrast.
• Bandwidth: the code's own airtime formula, (2 + data symbols) × 50 ms per frame, means each ball handoff takes hundreds of milliseconds. Real-time paddle syncing is not possible; the design hands the ball over in discrete events instead.
• Model safety: The Batch reports that Opus 5.5 passes cybersecurity and biology queries that Anthropic considers sensitive to Opus 4.8 instead. That does not affect this project.
• Conflicting dates: sources disagree on the Opus 5.5 release date, as noted in the timeline.

• Prototyping cost: one prompt produced a niche signal-processing design (modulation, framing, acknowledgments). That is the kind of work that used to need a specialist. This is a single anecdote, not a benchmark result.
• Review burden: the code is easy to read but has had little testing. Before relying on similar generated protocols, teams should run their own tests across devices and noise levels.
• Uses beyond games: the same browser-only, no-network method could support device pairing, offline handoffs or kiosk interactions. It also creates a covert data path that security teams should keep in mind.
• Cost: The Batch lists Opus 5.5 at $4/$20 per million input/output tokens, which matters for how many experiments like this a team can afford.