A news API for language models
Grounding a model in current events needs three things most news APIs make hard: article bodies, honest metadata, and access without a human in the loop.
Why headlines are not grounding
A model handed twenty headlines will produce a fluent summary of twenty headlines. It will not have the numbers, the quotes or the attributions, so it will fill those in from its own weights — which is exactly the failure retrieval was supposed to prevent.
Grounding needs the body. Every competitor in this category puts the body behind a paid
tier; here full_text=true is open to everyone.
curl "https://freenewsapi.ai/v1/search?q=interest+rate+decision&full_text=true&size=5"What matters when a model is the consumer
| Requirement | Why | Others | Here |
|---|---|---|---|
| Full article text | Headlines cannot ground a claim | paid tier | free |
| No API key | An agent cannot complete a signup form | required | none |
| Provenance labels | A model should know which fields are guesses | not published | country_source, date_source |
| Known token cost | Context budgets are finite | not published | published |
| Cheap counting | Measure coverage before spending tokens reading | varies | /v1/stats, ~200 tokens |
| Stable ids | Deduplicate across runs | varies | MD5 of URL |
| Forward-only cursor | Poll without missing backdated articles | varies | sort=crawled |
The pipeline that works
- Measure — is there coverage? ~200 tokens.
- Retrieve 60 candidates without bodies. ~9,000 tokens if a model read them, but your code reads them.
- Rank in code: one article per publisher, drop the stubs.
- Fetch bodies for the surviving eight. ~7,000 tokens.
- Answer with a citation on every claim.
Roughly 9,000 tokens end to end, against 45,000 for the naive "fetch fifty with full text" approach, and the answer is better because the eight were chosen. Full code in the RAG recipe.
Ready-made integration
OpenAI
Function-calling schema and dispatcher.
Anthropic
Tool-use definition and conversation loop.
Prompts
When to search, how to phrase it, how to cite.
