Methodology

1. Model + agent catalog

The catalog is curated by gpt.buzz editors. Each entry — name, vendor, release date, context window, parameter count, modality, license, weights URL, base-model dependencies — is verified against the vendor's official announcement, system card, API documentation, or research paper before publication.

When vendors release a successor (e.g., GPT-5 → GPT-5.5), the prior entry is kept for historical continuity and marked as superseded in its description.

2. News + capital aggregation

News and capital activity are auto-aggregated every 15 minutes from 12 active RSS, GitHub-release, and capital feeds. Items are deduplicated by source URL hash.

Two filters keep noise out: a 90-day age cutoff and a relevance gate that requires either a tracked-model/agent mention or a domain keyword (LLM, foundation model, context window, RLHF, etc.). Capital items run through a stricter gate that requires a deal-flow keyword (raised $, Series A/B/C, valuation, acquired, IPO) plus AI context.

Article summaries are generated by a small GPT-class model via an OpenAI-compatible endpoint. The prompt forces a strict two-sentence format: what happened, why it matters. Raw RSS body is the fallback when AI is unavailable.

3. Provider status

Provider availability is polled every 5 minutes from the official Statuspage.io feeds (where exposed) for OpenAI, Anthropic, DeepSeek. Each fetch updates current operational state and ingests new/changed incidents.

Status reflects the vendor's own dashboard — we do not invent or extrapolate. Providers without Statuspage v2 (Google, xAI, Mistral, Meta, Alibaba) display as "status not tracked" with a link to their official page.

30-day uptime is computed from the duration of non-resolved major/critical incidents in the rolling window.

4. Pricing history (in progress)

A pricing_history table records per-model token costs over time. Phase 5 wires in scrapers for vendor pricing pages and OpenRouter to populate it.

Once populated, pricing changes will fire Pro alerts and feed the cost calculator. Until then the schema is empty and per-model "Pricing history" sections show a placeholder.

5. Benchmarks (in progress)

Benchmark data ingest from the Open LLM Leaderboard (Hugging Face), LMSys Chatbot Arena, OpenRouter, and vendor-published evaluations is planned for Phase 5. Until then, model pages show a placeholder.

6. Editorial standards

gpt.buzz is independent — no vendor pays for placement, no sponsored content, no affiliate links inside model descriptions.

When a piece of data conflicts between sources, the vendor's own announcement wins. When a vendor stops publishing details (e.g., closed-weight parameter counts), we say "unknown" rather than guess.

Spot a wrong entry? Open an issue, message us, or hit the "report a fix" link on the model page — we patch and roll forward.

Active sources