Claude Fable 5.1 cuts agent costs

Claude Fable 5.1 lands with cheaper agent runs

Claude Fable 5.1 is now available everywhere, positioning a more capable coding and knowledge-work model as a direct option for long-running agent workflows.

• Claude says Fable 5.1 is built for complex, long-running tasks and research work.
• On Terminal-Bench 4.0, Claude reports a 55.8% score for Fable 5.1, versus 42.0% for Fable 5.
• Cache reads cost 75% less than before, which Claude says can reduce typical workload costs by around 25%.
• For highly agentic workloads, Claude says the practical cost reduction can reach 45%.

What it means Test Fable 5.1 on your longest coding and agent loops first, where cheaper cache reads could improve both task quality and unit economics.

Source: claudeai

Gemini gets selective about long videos

Gemini’s latest models can now take a goal-directed approach to video analysis, choosing what to inspect across frames, audio, and transcripts instead of scanning every file at a fixed rate.

• Google DeepMind says the feature delivers better video-analysis accuracy while using up to 88% fewer tokens.
• The model dynamically adjusts frame rate to find the moments needed for a task.
• The biggest efficiency gains are aimed at long-form video, from 10-minute guides to multi-hour content.
• Google AI Studio also published a developer guide for getting started with agentic video understanding.

What it means Revisit video-heavy workflows now, because selective analysis can make searchable archives, call review, and long-form content agents far cheaper to run.

Source: GoogleDeepMind

Perplexity splits private work from cloud work

Perplexity Computer can now start a task in the cloud and shift sensitive steps to a local model running on your Mac.

• Hybrid compute is available today in the Perplexity Mac app.
• Cloud models handle search, planning, and reasoning, while local models handle private files or sensitive information.
• Perplexity says the system combines the work into one output without sending sensitive data to the server.
• A privacy gate detects PII before it leaves the device, and Perplexity is open-sourcing that classifier.
• The feature is available for Pro, Max, and Enterprise subscribers on macOS 15+.

What it means Use hybrid designs as a product pattern for privacy-sensitive agents, keeping planning in the cloud while routing customer documents and personal data to local compute.

Source: perplexity_ai

Replit MCP puts app operations in chat

Replit MCP lets MCP clients steer Replit’s agent and manage Replit Apps from the tools where teams already work.

• Replit says its power users have used MCP to create, find, inspect, update, and publish apps since the beta release.
• The company says you can use Replit MCP from ChatGPT, Claude, Slack, or any MCP client.
• Examples include running a property business through a fleet of apps from chat.
• Users have also audited 50+ apps with scorecards in one prompt and checked every Replit app database in one sweep.

What it means Connect Replit MCP to your team’s preferred AI client and try managing a portfolio of apps, audits, or maintenance checks through one conversational control layer.

Source: Replit

Airtable was just acquired - do you need to migrate?

A look at whether Airtable users need to migrate after its acquisition.