On SWE-bench Pro, running Sonnet 5 as the executor with Fable 5 called as an advisor tool achieves roughly 92% of Fable 5's standalone score at about 63% of the price. Fable 5 is invoked only about once per task to provide high-level steering.
On BrowseComp, pairing a Fable 5 orchestrator with Sonnet 5 worker sub-agents via Claude Managed Agents reaches 96% of solo Fable 5 performance at 46% of the price. Token-heavy research work is offloaded to the cheaper Sonnet 5 workers.
The orchestrator pattern flips the advisor model: Fable 5 plans and delegates tasks to Sonnet 5 workers, so the bulk of tokens are billed at the lower worker rate rather than Fable 5 rates.
Claude Managed Agents supports both the advisor and orchestrator patterns with sub-agents. A notable detail: each sub-agent maintains its own cache, so repeated calls to the same context don't incur full token costs twice.
A team member kicked off a live video editing session using Fable on 59GB of raw conference footage, using an HTML slide deck alongside the video. The thread covers the full pipeline from initial prompt to rendered clips.
The video editing setup is a lightweight harness of scripts and Remotion assets. The team member plans to open source it if there's interest.
A team member received 60GB of raw conference talk footage and is running it through a Claude-assisted video editing pipeline, also feeding in the HTML slide deck. The thread documents the starting prompt and outputs in real time.
A team member is live-testing Claude-assisted video editing on a 60GB conference talk recording. The clip shows Claude switching between multiple layouts and seems to be tracking speaker position to decide on framing.
Video rendering is the bottleneck in the Claude-assisted editing pipeline, so the team member is batching low-resolution renders to iterate faster. Using an HTML deck turns out to be a key enabler: Claude can create dynamic animations directly from it.
During the live video editing session, Claude took individual static slides from the original deck and turned them into animations. The team member found this particularly impressive given the slides started as flat, separate HTML slides.
In the video editing demo, Claude appears to detect where the speaker is standing in frame and uses that to position slide overlays accordingly. The team member found a short closing clip it added unexpectedly charming, though the transcription highlighting remains too busy.
Without being explicitly asked, Claude generated YouTube Shorts-style clips from the conference footage. The transcription highlighting is still too chaotic, but the autonomous creative decision to make short-form clips was unexpected.
The video editing pipeline runs Whisper to transcribe the entire talk before Claude begins editing. This transcription feeds into downstream decisions about cuts, captions, and layout.
A detailed thread traces how Claude Code grew from an internal Anthropic experiment into what it is today, covering origins, team, philosophy, launch, and the agentic future.
A live walkthrough tomorrow at 10am PT covers the evolution from single-player Claude Code to multi-player Claude Tag, including how Claude Tag monitors channels, does proactive work, accepts steering from the whole team, and retains memory across sessions.
A user reported that the repo selection dropdown isn't working correctly. Clicking other repos causes it to snap back to a single default repo instead of switching.
A user wasn't seeing their repos in Fable. The app only surfaces repositories the user has contributed to within the last year, which explains the missing repos for people who mostly open issues rather than commit code.
