Overview
Today felt like a tug of war between raw capability and real-world friction. We got bigger, smarter models and flashier creative tools, but also the bill shock, token burn, and the creeping sense that “just try it” can get expensive fast. Outside AI, Starlink’s march across airline fleets and a wild insurtech valuation jump kept the tempo up.
The big picture
The pace is not easing. Model releases are landing with new “effort” modes, parallel agent workflows, and glossy demos that make whole roles look compressible. At the same time, the practical bits are biting harder: usage limits, runaway spend, and the awkward gap between what these tools can do and what people can afford to do with them day to day.
Claude Opus 4.8 lands with a clearer, more natural feel
Anthropic’s Alex Albert says Opus 4.8 is the direct response to 4.7 feedback: better at nuance, more natural to talk to, and a stronger collaborator for coding and knowledge work. The subtext is confidence: sharper judgement, more honest boundaries, and longer independent runs, without a price bump.
“Low effort” reasoning that still hits hard
Early benchmarking chatter suggests Opus 4.8 is doing more with less, at least on popular tests. @scaling01 points out that 4.8 in low reasoning mode can match older versions running much “hotter”, which is the sort of efficiency jump that gets builders paying attention.
Claude Code’s dynamic workflows, parallel agents on demand
@gregisenberg highlights a new Claude Code feature where you can ask for a workflow, or switch on “ultracode”, and it spins up swarms of parallel agents that cross-check each other. The pitch is simple: the unit of work grows from a file to something closer to a whole project, though you can almost hear the meter running.
Token burn meets hard limits, even on the £££ tiers
The fun part about parallel agents is the speed. The un-fun part is how quickly you can hit caps. @theo says one prompt pushed his session limits to 100% after resubbing at the $100 tier, which matches the broader mood that premium features can be intense, but brief.
When AI spend goes off the rails
The most jaw-dropping story of the day: Polymarket amplifying a report of an enterprise client allegedly spending $500,000,000 in a month on Claude after failing to set employee limits. Whether every detail survives scrutiny or not, the lesson is plain enough: without guardrails, usage-based AI can turn into a finance incident.
AI dubbing that keeps the performance, not just the words
ElevenLabs is pushing dubbing beyond flat transcript readouts. Dubbing v2 conditions on the source audio so emotion, pacing, and delivery carry across languages, with timing-aware translation to keep it watchable. If it works as advertised at scale, it changes localisation from a big project into a routine button press.
Google’s “Nano Banana” image models go general release
Google has put Nano Banana 2 and Nano Banana Pro into general availability via AI Studio and the Gemini Enterprise Agent Platform API. The headline is developer access, higher-res output, and tools aimed at consistent editing, which is the part businesses actually care about when they need repeatable assets, not random surprises.
App Store link in, sizzle video out
@minchoi shares a demo where a single App Store URL becomes a finished app promo video using Pika’s MCP Founder Kit skills inside Claude. No storyboard, no timeline fiddling, just “here’s the link”. It is hard not to see where this goes for small teams who need passable marketing output yesterday.
Starlink Wi‑Fi becomes an airline arms race
@SawyerMerritt says United expects Starlink on 80% of its fleet by the end of the year, with Southwest targeting 300+ aircraft by end of 2026. In-flight internet used to be a mild annoyance you tolerated. Now it is edging into “reason to choose an airline”, especially for streaming and calls.
Corgi’s valuation jump, fast enough to make investors blink
TechCrunch reports insurtech startup Corgi raising $106M at a $2.6B valuation, triple its valuation from three weeks earlier. Rapid mark-ups always invite scepticism, but the company’s AI-native insurance pitch and reported revenue momentum are clearly pulling capital in at speed.
Episode #415: 29 May 2026
