|
$ ai-news --date 2026-10-08 | mail
AI News — October 8, 2026
// the most viral AI topics of the last 24 hours, curated for your stack
$ cat top-story.md
Anthropic released Claude Haiku 5.5 on October 7: $0.10/$0.50 per million input tokens and $0.50/$2.50 output, about 75% cheaper than Haiku 4.5. It scores 72.4% on OSWorld and 39.2% on Terminal-Bench, supports effort levels up to Max, and is available on AWS, Google Cloud and Azure.
Why it matters: It is built to work as a cheap subagent next to larger models, which cuts the cost of agent fleets. Simon Willison notes a new tokenizer that uses ~25% more tokens, which eats part of the price advantage.
// source: Anthropic · Simon Willison
## 📦 Models & Releases
Haiku 5.5 matches GPT-6 Luna's $0.10/$0.50 price up to 100k tokens and benchmarks better, but beyond 100k tokens Luna is the better value. Why it matters: A rule of thumb for the cheap tier: context length and tokenizer overhead decide, not list price. // simonwillison.net
## 🤖 Agents & MCP
v2.1.293 makes Haiku 5.5 the default Haiku model (1M context) and fixes an HTTP MCP memory leak; v2.1.292 adds claude plugin install --marketplace, an effort parameter for the Agent tool and sandbox/UNC-path permission fixes. Why it matters: You can now tune sub-agent effort per call, and the MCP leak fix matters for long sessions. // GitHub
Per an October 5 report by The Information, Microsoft lowered per-employee monthly AI caps from ~$100k to ~$10k, and Meta's internal Claude Code users fell from ~60k to ~30k as it pushes MetaCode and Muse Code. Not officially confirmed. Why it matters: Agentic coding costs are becoming a budget line item, and in-house tools are gaining ground. // Let's Data Science · AI Weekly
## 🏠 Self-Hosting & RAG
717 commits: DeepSeek-V4.1-Flash optimizations (FlashMLA with NVFP4 compressed KV cache), a weight-cache daemon for fast engine restarts, Model Runner V2 speculative-decoding improvements and a new --max-num-active-seqs scheduling flag. Why it matters: Faster restarts and the concurrency cap are directly useful for a long-running Gemma serving setup. // GitHub
Faster sparse-vector upserts, better proxy-segment persistence, and fixes for HNSW search issues and a storage-corruption case. Why it matters: A patch worth applying if you use sparse/hybrid retrieval in your RAG pipeline. // GitHub
The slim container image is ~89% smaller (~175 MB), terminal skills appear alongside workspace skills, and every reachable skill is listed to the model by name and description. Why it matters: Smaller images speed up deploys, and skill discovery moves OpenWebUI toward agent-style workflows. // GitHub
$ echo "automatically researched on 2026-10-08"
> ai-news · news.reichenberg.ruhr
[ unsubscribe ]
[ view in browser ]
|