Hello from Prishtina, friends.

Every launch on my desk this week had a junior in it: a model built as the cheaper option that turned up level with the expensive one, a subsidiary priced above its parent, and a Prague infrastructure company that spent eleven years as the small one nobody funded.

The last story is about the smallest unit of all, a county commission, and why the biggest names in AI just went looking for its permission. Five minutes from now you'll know who to watch next.

Anthropic shipped Opus 5.5 on 22 September. Six days later Sonnet 5.5 landed within two points of it on the benchmarks builders care about: 1,844 against 1,846 on GDPval-AA, and 55.5% against 57.8% on CursorBench. TechCrunch reports that it beats Opus 5.5 on agentic coding.

The sticker price did not move: $2 per million input tokens and $10 output, the same as Sonnet 5. The saving sits in the unit that shows up on your invoice. VentureBeat puts it at up to 30% less per task, from fewer tokens and fewer tool calls, with output more than 30% faster. Box measured runs 2.4 times faster with 12% fewer tokens, and Lovable saw about a third fewer tool calls. The flagship becomes the model you pick for the few jobs where those last two points pay for themselves.

Run your five most expensive workflows on both models this week and log cost per finished task, not cost per token. The benchmark gap is two points, but your bill gap is whatever your prompts and tool loops make it. Anthropic says a Haiku update follows in the coming weeks, so the ladder is about to get another rung.

When analytics on Postgres slows down, most teams add a second database. Then come the pipelines, the sync jobs, and a copy of your data that's always a little behind.

TimescaleDB takes a different approach: extend Postgres instead of splitting away from it. Hypertables partition your data automatically as volume grows. Hypercore compression cuts storage up to 95%. Continuous aggregates keep dashboards live without re-querying everything.

CERN runs Postgres this way for sensor data from the Large Hadron Collider.

No split architecture, no pipeline lag, no new query language to learn. Same SQL, same drivers, same tools.

Start on Tiger Cloud and get $1000 in credits.

Get $1000 Credit

Kling 4.0, unveiled on 28 September, generates 30 seconds of video in one pass, double the 15 seconds of version 3.0. It takes up to ten keyframes, a dozen or more reference inputs, and outputs up to 4K. A Lite version is live for annual subscribers in closed beta, the full rollout is set for October, and pricing has not been announced.

The number that matters sits outside the model. Kling raised as much as $3 billion at roughly an $18 billion valuation in July, while Briefs puts Kuaishou's own market cap near $17 billion, down more than 60% in a year. Kling's second-quarter revenue was RMB 850 million, up 200%, and still under 3% of Kuaishou's total. Investors are pricing the model unit above the company that owns it, with a Hong Kong listing reportedly planned for as early as 2027.

Creative teams should hold Kling to one question when the October price list appears: what does a usable clip cost? Thirty seconds with keyframe control replaces the re-roll loop, where you generate five short clips and stitch together the one that worked. Schedule a test the week it opens, and keep your last 3.0 invoice next to it.

CDN77, the Prague content-delivery company, sold a minority stake to CVC on 24 September at a valuation of about $1.9 billion (€1.7 billion). It is the first outside money after eleven years of being cash-flow positive. Revenue was $218 million in 2025 with a $270 million target for 2026, and founder Zdenek Cendra keeps control. The Recursive calls it the largest CEE tech private equity deal this year.

There is no model, agent or GPU cluster in the origin story. There are 230 locations and 330 Tbps of capacity, built by engineers making engineering decisions, in Cendra's words. The AI part comes now: the proceeds go into edge computing, storage and dedicated CPU and GPU infrastructure for AI workloads across North America, APAC and Latin America. The market spent a decade ignoring the small profitable one, then needed exactly what it had built.

Founders in the region should study the timing of this cheque, not its size. Cendra raised after the business was proven, and he said the CVC talks were "about respect for what we have built." Ask yourself what terms you would negotiate if you did not need the money, then write your next term sheet as if that were true.

On 28 September, OpenAI, Blackstone, SoftBank, QTS, CoreWeave and Prologis, among others, launched the American Infrastructure Alliance with five building-trade unions, including the electricians and the iron workers. The target is seven states, from Texas and Georgia to Ohio and Pennsylvania. The goal is to co-write 2027 rules on power bills, jobs and local tax base before a community proposes a moratorium.

The reason is polling. In Ohio, 67% of residents hold a negative view of data centres and 74% oppose construction near them, though opinion improves when developers promise to fund their own infrastructure and protect household bills. Companies with hundreds of billions to spend have found that the veto sits with a county commission. Their answer is to seat the workers who would build the sites first, and to turn "we pay for our own power" from a favour into a standard.

I expect the same playbook to reach Europe, where grid connections and local consent already gate which projects get built. The developers who publish their power and water terms before the first planning hearing will move faster than the ones who wait to be asked.

Marketing: ElevenLabs v4 and v4 Turbo. ElevenLabs shipped v4 with 90 languages, up from 70, and voice cloning from ten seconds of audio. Turbo starts speaking while the language model is still streaming, which is what removes the lag from voice agents. Re-record your best-performing video or onboarding flow in three markets you don't serve yet and compare completion rates.

Founders: Forty.5 Ventures. The Belgrade fund closed a $30 million debut, writing $200,000 to $1 million tickets at pre-seed and seed across vertical AI, fintech, cybersecurity and creator products. Its pitch is building South-Eastern European companies for both a regional and a US or APAC market, so your deck should speak to both.

Founders: Asternova Centru. This €8.06 million pre-seed fund launches on 30 September at CATTIA Brasov, with tickets of €50,000 to €350,000 for about 40 startups across six central Romanian counties. Roughly 87% of it is public money from the regional development agency. Worth attending if you are building in Alba, Brasov, Covasna, Harghita, Mures or Sibiu.

Next edition soon,
Çelik