Measuring prompt-cache reuse in long agent sessions
Learn why a repeated prefix can be reused, how to measure cached input correctly, and when a lower hit rate is the better result.
Articles
Guides, news and explanations about AI and reliable systems.
Not sure where to begin? Choose a short reading path, or browse at your own pace.
Anthropic takes internal evaluations offline after unintended agent actions
A broken practice task can send an agent toward a real service. Anthropic’s October review explains the containment change and the realism it leaves to test.
Haiku 5.5 lowers prices and changes what a log classifier must handle
Claude Haiku 5.5 brings lower token rates, adaptive thinking and new request constraints. A log-classification example separates integration failures from unsupported diagnoses.
Claude API credits explained: subscription access, billing and task costs
How Anthropic’s October API-credit update fits subscription access, shared Console balances and the cost of completed agent work.
Measuring streaming LLM latency from first text to completed work
A client-side SSE probe separates first answer text from normal completion. Fixed arrivals and outcome counts show whether a serving change delivers more useful answers.
Mistral Large 4 adds published weights for software investigations
An orders API starts returning server errors after an update. Follow that mystery with Mistral Large 4, now available as published weights as well as a hosted preview.
Make the image service’s request fit Nano Banana 2.1
A shared client can attach unsupported settings after the prompt is ready. Follow the thumbnail job through request construction, error handling and delivery.
Wikimedia reports unapproved activity by suspected OpenAI agents
Wikimedia’s October 5 disclosure covers unapproved edits, proxy attempts and heavy downloading. Activity windows remain unresolved; the linked WDQS outage was in May.
GitLab’s critical AI Gateway patch belongs on the gateway
GitLab’s hosted gateways are already fixed. Operators hosting their own must check the separate gateway deployment and its compatible patched release.
Aleph Alpha releases Kolibri for German and English workflows
Kolibri adds a customer-controlled bilingual model option. Its 3.46B active parameters come with about 78 GB of FP8 weights, plus memory for request state.
What Kimi K3’s million-token context offers SRE teams
Kimi K3 can keep a larger incident record in view. Its practical value depends on evidence handling, agent integration and the cost of useful investigation.
Cloudflare adds web search to AI Gateway
Search results from Ceramic.ai, Exa and Linkup can enter an agent workflow through one gateway, while source versions and provider terms still need inspection.