Most reframing starts with:
“Make this 9:16.”
It shouldn't be the instruction.
First understand what’s happening in the video.
Then decide what deserves to stay in frame.
The shape comes second.
Have you ever thought about how much content you're leaving inside your long-form videos?
One conversation can easily become weeks of posts, but not every moment deserves to become a clip.
The real work is knowing which ones deserve to stand alone and deciding which ones are worth publishing.
There is a real difference with Clipping tools & Video Agents.
A clipping tool finds moments. You still run every step around it: settings till the review what goes out.
An agent takes one instruction. Last week I gave it this one, in Claude with reap connected:
"Find every strong take where he talks about why he didn't sell Buildspace, and turn them into vertical clips with captions."
What came back, without me touching a setting:
→ it read the 1hr interview and picked portrait, captions and face tracking on its own
→ 30 clips, cut, captioned and ranked
→ a read on the batch: lead with #1, hold #4 for a follow-up two days later
→ one more prompt, and the same interview came back captioned in Spanish, Bengali and Arabic, 30 clips each
The shift is not that AI can find clips.
It is that you stop doing every step yourself. You give the direction, it handles more of the work.
This is the direction we are building toward at reap.
You have your long-form video (podcast, interview, webinar etc) in English language, but for your audience reach you want to Translate it to other language here's how you can do it in few easy steps:
-> Drop in a video link that you have
-> Select the Target language from the Translate to dropdown
-> Hit get clips reap will automatically translate your clips
With one video you can easily localized it to 100+ other languages.
Watch it below.
Claude Opus 5 is out. Nearly all of Fable 5's intelligence at half the price.
But the real tell is the new low/medium/high effort toggle. The frontier labs have stopped selling "the smartest model" and started selling "the right amount of smart for this task."
Intelligence isn't the product anymore. Metering it is.
AI search visits grew 42.8% last year. Google grew 2.4%.
And AI-search visitors convert 4.4x better than regular organic, because they arrive already sold by the answer.
You're not losing traffic. You're losing the traffic that was going to buy.
Kimi K3, a model nobody mentioned last week, just hit #1 on Arena's frontend board, past Claude and OpenAI.
17-place jump. One release.
Nine days: Grok 4.5, GPT-5.6, Muse Spark, now this.
If your strategy depends on which model is best, you don't have a strategy.
@deredleritt3r The benchmaxxed question is the one builders should care about most. A model that wins frontend pairwise but degrades everywhere else isn't a frontier model, it's a very good frontend tool. Both are useful. They're just different purchases.
@arena The 17-place jump from K2.6 to #1 is the real story, not the rank. Frontier capability is now something you can leapfrog in a single release. Nobody's lead lasts long enough to build a moat on it.
@mweinbach This is the demo that lands better than any benchmark. Nobody feels a 3-point eval bump. Everyone feels "it rebuilt macOS in my browser." Capability you can see travels further than capability you can measure.
@spacexai "#1 on research" is the line worth pulling out here. Everyone benchmarks code gen, but research ability is what decides whether an agent can actually run a long task unsupervised. That's the number to watch.
@JasonBud Open-sourcing the harness is the underrated half of this. The models converge anyway. What compounds is the tooling everyone builds on top, and the goodwill you get from not gating it.
@leerob Every lab and tool is racing to give away more compute right now. Worth noticing what that implies: access is being commoditized on purpose. When usage stops being the constraint, the constraint becomes knowing what's worth building.
@AlexFinn Agreed. And the uncomfortable flip side: when everyone can build anything for pennies, building stops being the differentiator. Taste in what you build, and distribution once it exists, are the only scarce things left.
@thsottiaux Getting ahead of this publicly, this fast, is underrated. Every model will eventually break something. The differentiator isn't never failing, it's how quickly you explain what happened. Trust compounds faster than capability.