Open weights, permissive licenses, and API prices far below the US frontier. The real question with GLM-5.2 and Kimi K2.7 Code is whether you can run them yourself, and what happens to your code when you cannot.
Sakana Fugu sells itself as AI sovereignty, launched days after rival frontier models were cut off by US export control. Read the Terms and the story narrows: routing you cannot see, the models it names all in one jurisdiction, and orchestration cost you cannot forecast. An audit of the gap between the pitch and the paperwork.
Claude Fable 5 was live for about 72 hours before a US export-control directive barred foreign nationals worldwide and Anthropic disabled it. A Tier C briefing on what the shutdown proves about buying frontier AI when access can be revoked by nationality, as of June 2026.
TensorZero, the open-source LLMOps stack, shut down in June 2026 with most of its $7.3M seed unspent. The gateway is replaceable; the learning loop it was built around is not.
On June 18, 2026, Google cut Gemini CLI access for free, Pro, and Ultra users and moved them to Antigravity CLI. The Apache 2.0 code stayed open; the free entitlement to Google’s backend did not. Here is what changed, who wins, who loses, and the dependency most migration guides skipped.
Instantly sells Credits and email sending as separate subscriptions. The $9 Nano Credits plan funds AI lead search, enrichment, and agent drafts, but it cannot send a single email on its own. FSR hit that wall: the AI Sales Agent stopped at the sending-account check, not the credit balance.
Mitte runs many image, video, and audio models on one credit balance, and the images are strong. The pricing is the part to check before you pay. In our paid testing, two of six configurations billed more than the panel showed on every run, one shown at 45 credits that settled at 280. Budget from the receipt, not the panel.
Claude Fable 5 is free on paid Claude plans through June 22, then it bills like the API. Anthropic publishes the token price but no conversion from your five-hour plan limits to Fable usage, so you cannot forecast the cost. Measure your own token burn while it is free.
We bought Zebracat credits before subscribing, and the refund was refused. This Zebracat review audits the buying layer where the risk actually sits: refund rules, credit metering, model costs, review incentives, and enterprise paperwork, all with receipts and dates. It does not score video quality, because we generated none.
Odysseus, PewDiePie’s free self-hosted AI workspace, is private only along the paths you keep local. FSR captured the traffic on one machine: the local model sent nothing out, a cloud model left for OpenRouter, and the default setup returned a 404 on first run.