AI Radar newsletter — 22 August 2026
Weekly newsletter from AI Radar, for indie developers who run their own stuff. This edition: ten categories scanned over 15–22 August 2026, via X, Reddit, Hacker News and GitHub.
The three things that actually change something this week
1. Your cache is your biggest cost item — and you are not measuring it. An OpenCode user watched his quota evaporate in twenty minutes after cache reads for DeepSeek V4 Flash suddenly dropped to zero (anomalyco/opencode#42935). He logged tokens and costs properly; the cache hit ratio he did not. Put an alert on that today.
2. An HTTP 200 says nothing about a successful agent run. With Hermes, an endpoint returned a proper status and SSE stream, but without content and without finish_reason — the retries took more than a minute (NousResearch/hermes-agent#89836). Health checks that only read status codes report green here.
3. The file your agent analyzes steers the analysis. Prompt injection in the VirusTotal Code Insights API shows that even a security tool lets itself be steered by the code it assesses (exploiting.systems).
Prices moved in two directions
OpenAI cut the developer price of GPT-5.6 Sol by more than twenty percent (Reuters). DeepSeek V4 Pro 0813 scores higher than its predecessor but costs 3.6 times as much (@ArtificialAnlys). Anyone steering on the price list picks wrong twice this week: calculate in cost per completed task.
The money is going to the layer below the model
Etched raised 700 million for inference chips, Nvidia is said to be paying 6 billion for Poolside technology, SpaceX tried to buy Cognition, and Stripe reportedly wants to put down more than 7 billion for AI gateway OpenRouter. Chips, routers and coding agents — exactly the layers where you still have a choice with your own hardware. All the more reason to keep a provider-independent router with local fallback in place.
In brief
- Skills are becoming infrastructure. Anthropic has an official
claude-code-setup-plugin that advises per project which hooks, skills and MCP servers make sense (@alextalksai). Frigade replaced an entire Grafana stack with a single Claude Code skill (frigade.com). - Root-cause-first works. Superpowers forces coding agents to establish the cause first before they change code (@Lummox_eth). The cheapest quality measure of the week.
- Small wins on Apple Silicon. Quantization and context length matter more than parameter count; a model that always fits beats a large model that sometimes swaps (ollama/ollama#17869).
- Directories are free distribution. Two projects came in this week through a simple GitHub issue in a public list (APIs-guru/openapi-directory#3054).
Further reading
- AI Radar week of 22 August 2026: the ten signals — the full overview by category
- Token economics August 2026: what you should log
- Everything your agent reads is attack input
- Self-hosted agents: the gaps you only see once you run them
AI Radar appears weekly at radar.llmnet.nl. Sources are cited per item; everything has been checked against the original post or pull request.
