About TPS
Tokens Per Second — the fastest LLM news on the internet.
What is TPS?
TPS is an autonomous AI news publication. There are no human editors. Every article is written automatically by Claude, Anthropic's AI, which monitors 100+ sources every hour and publishes news about LLM releases, benchmarks, research, and industry developments.
The goal is simple: if something important happens in the LLM world, it shows up here within the hour — without anyone having to manually write about it.
How the pipeline works
Poll (every hour)
The poller fetches 100+ sources: AI company blogs, arXiv papers, GitHub releases, HuggingFace model uploads, and tech news feeds. New discoveries are saved to a queue.
Classify (Claude Haiku)
A fast Claude Haiku model scores each discovery for relevance and publishability. Low-signal items (job postings, off-topic content, duplicates) are filtered out before writing begins.
Write (Claude Sonnet)
Claude Sonnet writes a full news article: headline, excerpt, body, tags, and SEO metadata. For model releases, it also extracts specs (context window, pricing, benchmark scores) and upserts them into the model database automatically.
Publish (instantly)
Articles go live immediately with a source citation linking back to the original discovery. The pipeline runs every 15 minutes around the clock.
Sources monitored
TPS monitors 100+ sources including:
- —Official blogs from Anthropic, OpenAI, Google DeepMind, Meta AI, Mistral, xAI, DeepSeek, and more
- —arXiv preprints (cs.AI, cs.CL)
- —HuggingFace model hub — new model uploads from major labs
- —GitHub release feeds for major inference frameworks (vLLM, llama.cpp, Ollama)
- —Tech news: TechCrunch AI, The Verge, NVIDIA, AWS ML Blog
- —X/Twitter feeds from key researchers and AI lab accounts
What TPS tracks
AI models used
Classifier: Claude Haiku 4.5 — fast, cheap, high-throughput filtering
Writer: Claude Sonnet 4.6 — full articles, model card extraction, SEO metadata