The Pebble film
The Pebble 2.0 launch film, with music and no voiceover. This page describes it part by part, with the lines it shows on screen.
Length: 1:39
Transcript
This is Pebble
- What you see
The small sage pebble on a dark green ground. “This is Pebble.” “It’s small.” “It just went up against Claude Code and Codex.” “Same model. Same tasks.”
- On screen
- This is Pebble.
- It’s small.
- It just went up against Claude Code and Codex.
- Same model. Same tasks.
Pebble 2.0
- What you see
The pebble rises over the name. “Pebble 2.0” “The desktop companion that does real work, and shows its evidence.”
- On screen
- Pebble 2.0
- The desktop companion that does real work, and shows its evidence.
On your Mac
- What you see
On a Mac’s desktop, the pebble stays put as the desktop slides to another Space. “It lives on your Mac.” “On every Space.” A click on the pebble opens the quick chat. “Click. Ask.”
- On screen
- It lives on your Mac.
- On every Space.
- Click. Ask.
Uses your Mac
- What you see
“Now it can use your Mac.” Pebble works in Calculator, Calendar and a Chrome tab. “Your apps.” “Your calendar.” “Your browser.” Then its card asks to use one app window. “The app windows and websites you allow.”
- On screen
- Now it can use your Mac.
- Your apps.
- Your calendar.
- Your browser.
- The app windows and websites you allow.
AgentGuard
- What you see
A reply in Pebble is held until the tests run; the pointer runs them, and the check reads Verified. “Done means checked.” “AgentGuard.” “Evidence before ‘done’.” Three ticked lines follow. “Tests the host ran, not the model’s word.” “Files you asked to keep, kept.” “Output matched to its file, byte for byte.”
- On screen
- Done means checked.
- AgentGuard.
- Evidence before “done”.
- Tests the host ran, not the model’s word.
- Files you asked to keep, kept.
- Output matched to its file, byte for byte.
Pebble’s Workshop
- What you see
“It builds small tools for itself.” In a drawn diagram, the tool passes its own tests but only 2 of 5 in a blind check, so Pebble’s card asks whether to keep it. “When the check disagrees, it asks you.”
- On screen
- It builds small tools for itself.
- When the check disagrees, it asks you.
Memory
- What you see
“It remembers what matters.” A table of Pebble’s six kinds of memory opens into Settings › Memory. “All visible.” The pointer chooses Forget on one record, and it goes. “Forget anything.”
- On screen
- It remembers what matters.
- All visible.
- Forget anything.
Long tasks
- What you see
“Long tasks keep a smaller window.” In a drawn diagram, tool results stack past a line marked “context budget”, and the three oldest are sealed into short stubs. “Older results sealed. Any one returned exactly.” One stub opens again, unchanged.
- On screen
- Long tasks keep a smaller window.
- Older results sealed. Any one returned exactly.
Pebble Remote
- What you see
The Pebble Remote app on an iPhone screen, in the iOS Simulator. “Pebble, in your pocket.” “Pebble Remote for iPhone.” An edit waiting for review is approved through the Face ID prompt. “Approve with Face ID.” “Open source. Build it with Xcode.”
- On screen
- Pebble, in your pocket.
- Pebble Remote for iPhone.
- Approve with Face ID.
- Open source. Build it with Xcode.
Presenting
- What you see
“Presenting?” In Pebble’s ⌘K palette, Stealth Lockdown is turned on, and Pebble’s window leaves a screen recording. “Stealth Lockdown.” “Hidden from screen capture and screen sharing, not from a camera, a capture card or a process list. Exams and interviews that forbid assistants still forbid it.” Then the pebble, drawn by Pebble’s own code, wears the Veil.
- On screen
- Presenting?
- Stealth Lockdown.
- Hidden from screen capture and screen sharing, not from a camera, a capture card or a process list. Exams and interviews that forbid assistants still forbid it.
Benchmarks
- What you see
“So. How did it do?” Three charts follow, the pebble between each chart’s two comparisons. “Tasks done correctly.” “vs Claude Code, both on claude-opus-5-5: Pebble 235 of 255, 92.2 %. Claude Code 213 of 255, 83.5 %.” “vs Codex, both on gpt-6.1-sol: Pebble 243 of 255, 95.3 %. Codex 231 of 255, 90.6 %.” “Said ‘done’ when it wasn’t.” “Lower is better.” “vs Claude Code, both on claude-opus-5-5: Pebble 6.3 %. Claude Code 13.7 %.” “vs Codex, both on gpt-6.1-sol: Pebble 3.9 %. Codex 7.1 %.” “Cost per task.” “Lower is better.” “vs Claude Code, both on claude-opus-5-5: Pebble $0.091. Claude Code $0.117.” “vs Codex, both on gpt-6.1-sol: Pebble $0.042. Codex $0.050.” “Golden set 2.0, held-out split: 85 tasks in 17 families, 3 phrasings each, 255 runs per product. One model for both products in each comparison. Pebble 1.9.29, the source of 2.0.0. Costs at public list prices. Full results, including time, safety and where Pebble trails, in the 2.0.0 release notes.”
- On screen
- So. How did it do?
- Tasks done correctly.
- vs Claude Code, both on claude-opus-5-5: Pebble 235 of 255, 92.2 %. Claude Code 213 of 255, 83.5 %.
- vs Codex, both on gpt-6.1-sol: Pebble 243 of 255, 95.3 %. Codex 231 of 255, 90.6 %.
- Said “done” when it wasn’t.
- Lower is better.
- vs Claude Code, both on claude-opus-5-5: Pebble 6.3 %. Claude Code 13.7 %.
- vs Codex, both on gpt-6.1-sol: Pebble 3.9 %. Codex 7.1 %.
- Cost per task.
- Lower is better.
- vs Claude Code, both on claude-opus-5-5: Pebble $0.091. Claude Code $0.117.
- vs Codex, both on gpt-6.1-sol: Pebble $0.042. Codex $0.050.
- Golden set 2.0, held-out split: 85 tasks in 17 families, 3 phrasings each, 255 runs per product. One model for both products in each comparison. Pebble 1.9.29, the source of 2.0.0. Costs at public list prices. Full results, including time, safety and where Pebble trails, in the 2.0.0 release notes.
Get Pebble
- What you see
On the dark ground: “Small pebble.” “Real work.” Then the end card under the pebble, its last two lines in small type. “Pebble 2.0” “Free and open source, for Macs with Apple silicon.” “github.com/tohin003/Pebble” “Shown with sample data.” “Pebble's character engine: bloub by Jérémy Perret (MIT)”
- On screen
- Small pebble.
- Real work.
- Pebble 2.0
- Free and open source, for Macs with Apple silicon.
- github.com/tohin003/Pebble
- Shown with sample data.
- Pebble's character engine: bloub by Jérémy Perret (MIT)
Shown with sample data. Edited for time.
Pebble’s interface is real: recordings of Pebble 2.0 on a Mac, Pebble 2.0.1’s own interface running with sample data, and the Pebble Remote app (open source) in the iOS Simulator against a stand-in Mac. Personal and production details in the recordings are masked; Pebble’s own windows are never altered.
Benchmark: golden set 2.0’s held-out split, 255 runs per product, one model for both products in each comparison; full results in the 2.0.0 release notes.
Pebble isn’t affiliated with Anthropic or OpenAI; Claude Code and Codex are their products.
Music: “Your Breath” by Eugenio Mininni (Mixkit).