Best Financial News API
for AI Agents面向 AI Agent 的
最佳金融新闻 API

Key findings at a glance

There is no honest single winner across global discovery, strict ticker filtering, multilingual retrieval, sentiment, and regional specialist feeds. This guide keeps those product shapes separate.

What we tested: 7 financial news APIs, 406 live observations

The broad benchmark completed 406 publishable observations from 410 live API invocations across 203 tool-case cells. Every cell ran twice. Cases covered company, macro, crypto, forex, invalid input, and relevant specialist feeds across nine market regions and seven requested languages.

The ticker benchmark added 120 completed observations across nine listing markets and an invalid ticker control. Two suppliers met the multi-market ranking gate. A market counted as supported only when both rounds returned news relevant to the target company. Each supplier received its own native ticker format rather than one shared symbol dialect.

All overview visuals use one fixed seven-supplier public comparison shortlist: Brave, Caidazi, EODHD, Finlight, Gildata, Linkup, and NewsAPI. The shortlist is the union of suppliers that qualified in at least one published buyer scenario in the baseline edition, including the strict ticker workflow. It is now frozen for comparable retests: a shortlisted supplier that later performs poorly becomes Not qualified but remains one of the seven rows. This also keeps a specialist such as Finlight visible in the global-company view even when that broad input does not clear the scenario gate.

“Valid API response,” “scenario returned the expected empty/non-empty shape,” and “the returned news was relevant” are different metrics. Treating them as one success rate would reward irrelevant fallback results and punish honest empty responses.

To prevent a global-search chart from hiding regional or specialist strengths, we compare five buyer scenarios. Every overview row uses the same seven shortlisted suppliers. Qualified means the supplier's best applicable tool cleared that scenario's published gate; Not qualified means it was tested but fell below the gate; N/A means no tested tool from that supplier exposed the applicable contract. A percentage is the share of applicable live rounds that returned at least one relevant result, with the observation count in parentheses.

Each scenario's detailed leaderboard remains narrower: it includes only qualified suppliers and only one tool per supplier. When several tools from the same supplier qualify, the representative is selected by relevant-result repeatability, then top-five relevance, applicable-case count, expected-response rate, valid-response rate, field completeness, and latency.

This is a qualification map, not a universal quality score. Below-gate entries for shortlisted suppliers remain visible as Not qualified, while detailed leaderboards include qualified entries only. The detailed metrics below distinguish relevance, freshness, fields, source diversity, duplicate behavior, sentiment coverage, latency, and negative controls.

The matrix uses each product's broad discovery or feed input. The strict ticker section is a separate suite using dedicated native ticker parameters and provider-specific symbol formats. Its results are not interchangeable with broad keyword results.

How we scored the financial news APIs: the published evaluation standard

The scoring unit is one observation: one product, one applicable test case, and one live round. A cell is one product-case pair, and an API invocation is one physical request. We report all three to keep the scoring denominator separate from the traffic volume.

At the individual case level and in the supplier comparison matrix, N/A means the inspected endpoint contract cannot express that request. Not qualified means applicable cases were completed but the best supplier tool did not clear the scenario gate. We do not convert contract-level N/A into a failure, and we do not grant untested credit based on another endpoint from the same supplier. Provider errors, wrong-language output, stale or irrelevant fallback data, and unexpected empty results remain in the underlying evaluation. Every comparison must complete two live rounds for every applicable cell before it can be considered for ranking.

A scenario ranking requires at least 50% valid API responses, relevant results in at least 50% of applicable rounds, and an endpoint contract covering at least 50% of that scenario's fixed cases. Only the best qualifying tool from each supplier is shown. The multi-market ticker ranking separately requires at least 50% valid calls and repeatable relevant results in at least five of the nine tested markets. These are eligibility gates, not extra points added after the test.

Ticker syntax is normalized per provider before the final run. Language claims are split into three separate facts: what the vendor documents, what the current integration exposes, and what the fixed live cases actually returned. Charts are generated from the same sanitized scorecard as the tables.

Is this evaluation scientific and accurate?

It is a controlled, reproducible product-selection benchmark, but it is not a statistically conclusive academic study or a long-term reliability certification. The fixed cases, explicit denominators, two-round completion gate, provider-native ticker formats, negative controls, and preserved run evidence make comparisons auditable and useful for choosing an API.

The main limitation is sample size and time dependence. Two rounds can confirm repeatability during the test window, but cannot estimate a provider's monthly SLA, full language inventory, or long-run recall with narrow confidence intervals. News availability also changes with the news cycle, licensing, subscription tier, and indexing delay. We therefore publish observed results with their test date and denominators, avoid a single universal score, and recommend quarterly reruns plus reruns after material integration changes.

Best financial news API by agent workflow

China/HK company-news trade-offs

The table below shows one qualifying representative tool per supplier. All four tools cover both fixed China/HK company cases and returned relevant results in at least half of their live rounds.

These numbers expose different product shapes. EODHD and Gildata Stock News were highly relevant in these fixed company cases, while Caidazi exposed richer observable source and core-field data. A 0.0 source count means the normalized response did not expose a usable source value; it does not mean that the supplier ingested zero publishers.

Chinese specialist-news trade-offs

This cohort requires coverage of at least two of the three fixed fund, industry, and organization cases, then keeps only the strongest qualifying tool from each supplier.

Both representatives returned relevant results in all six observations. Their trade-off is different: Caidazi was faster and exposed more complete normalized fields and sources, while Gildata returned a higher share of records inside the freshness window.

Macro and cross-asset trade-offs

The following supplier representatives met all ranking gates in the macro, crypto, and forex scenario. Applicability still differs: NewsAPI had 10 observations, while the other rows had 6 or 12.

Freshness and field values are measured only from normalized returned rows. An invalid-control failure means the endpoint returned fallback content for the deliberately nonexistent query; it does not invalidate its positive-case results, but agents should add their own relevance guard.

Finlight: best ticker news API for structured agent workflows

Finlight completed 20/20 calls in the strict ticker suite, passed both invalid-ticker controls, and returned relevant news in both rounds for US, HK, JP, DE, FR, IN, and ES. CN and BR returned empty results in both rounds. Median latency was 1,574 ms.

Finlight also returned structured article fields useful to agents: title, URL, publication time, source, summary, language, countries, sentiment label, categories, and expanded company entities. Its confidence field is classification confidence, not directional sentiment, so we did not score it as a sentiment value.

Which languages do financial news APIs support?

Language coverage has three layers:

1. what the vendor documents;
2. what the current integration lets an agent request;
3. what returned relevant requested-language news in live tests.

The chart preserves the same seven-supplier public comparison shortlist. Suppliers for which we did not independently verify a vendor or integration language-control profile remain visible as N/A; they are not treated as supporting zero languages. The compact table lists the four verified profiles and reports observed case results without inferring an undocumented language inventory.

Finlight's own documentation says that the API indexes news in many languages, filters one language per request, and defaults to English. It also documents strict ticker filtering and exchange-aware company entities. The missing language control is therefore an integration gap, not evidence that the underlying Finlight API is English-only. See the Finlight REST endpoint and multilingual query guide.

NewsAPI documents its exact language enum on the Everything endpoint.

Stock ticker lookup by market: which API resolves symbols reliably?

Ticker support is not the same as keyword recall. A product can find an article containing “Toyota” yet fail a strict 7203.T lookup, and different providers use different suffixes for the same listing. Agents that accept user-entered symbols should normalize symbols per provider before calling the news endpoint.

Country and requested-language results

This matrix preserves the same seven-supplier public comparison shortlist and uses each supplier's global-company representative, whether qualified or below the gate. Every applicable tool-case cell was called twice; N/A means the case is outside that endpoint's tested contract, not that the test was skipped.

Each percentage is the number of passing live rounds divided by two. A round passes only when the call has the expected response shape, returns at least one relevant result, all language-observable top-five results match the requested language, and any available country metadata matches the requested market. Therefore, 50% means one of two live rounds passed the complete rule; it is not an article-level language accuracy or a long-term provider support rate.

Country tags in financial news can mean source country, company domicile, listing market, or article topic. Buyers should verify which semantic a provider uses before treating country filters as exchange filters.

How to choose a financial news API for your AI agent

Choose Finlight when your agent starts with a ticker and needs structured entities, sentiment labels, categories, provenance, and clean negative-control behavior. If multilingual Finlight retrieval matters, expose its native language parameter first and rerun the same fixed cases.

Choose EODHD when broad exchange-coded ticker coverage matters and your agent can handle provider errors explicitly.

Choose Linkup when multilingual recall and citations matter more than having a strict news-feed contract. Choose Brave for English discovery within its tested scope. Choose NewsAPI when you need a familiar keyword API and explicit language selection, while accepting that declared language options do not guarantee relevant local-market results.

For China/HK company workflows, compare Gildata Stock News, EODHD, Linkup, and Caidazi Hybrid Search V2 against the response shape your agent needs. Gildata Stock News and EODHD were the most consistently relevant in the two fixed company cases; Caidazi exposed more complete normalized fields and observable source diversity; Linkup traded structure for broader retrieval.

For Chinese fund, industry, and organization workflows, Caidazi Hybrid Search V2 was faster and more structurally complete in the fixed suite. Gildata Public Opinion matched all six relevant-result rounds with stronger observed freshness, but higher latency.

For macro, crypto, and forex discovery, Linkup had the strongest relevant-result repeatability, followed by NewsAPI and Gildata Public Opinion in their applicable cases. Linkup and Brave both failed the invalid-query control, so an agent using them should validate entity/topic relevance before accepting fallback results.

Limitations

This is a dated live observation, not a permanent statement about a provider's entire inventory. News availability changes with the news cycle, subscription tier, source licensing, indexing delay, and integration schema. Empty results do not prove a market is contractually unsupported; they mean that the fixed query returned no relevant article in the tested window.

The benchmark scores the endpoints and parameters actually called. It does not transfer a capability from another endpoint owned by the same provider.

How providers can join or improve the benchmark

Providers can submit a comparable news discovery or ticker-news endpoint, document its market and language controls, and supply a test credential or QVeris integration. Inclusion and ranking cannot be purchased. Corrections to contract facts can be made from documentation; performance results change only after the fixed live suite is rerun.

Start by checking whether your product is already listed in the QVeris Provider Hub. To request a new listing, correct an endpoint contract, or schedule a comparable retest, email support@qveris.ai with “Financial News Benchmark” in the subject.

For the fastest retest, provide:

• the exact endpoint and stable schema;
• supported ticker formats and exchanges;
• language codes and whether the filter is article-level or source-level;
• country-field semantics;
• source, licensing, and freshness constraints;
• invalid-input and empty-result behavior.

FAQ: financial news APIs for AI agents

What is the best financial news API for an AI agent?

There is no single winner. In this benchmark, Finlight was the best structured ticker news API, Linkup the strongest for multilingual and macro discovery, EODHD the best broad-exchange ticker alternative, and Gildata and Caidazi the best for Chinese-market coverage. Pick by workflow, not by an overall score.

Can I search financial news by stock ticker?

Yes. Several tested integrations exposed dedicated ticker or symbol inputs, but only Finlight and EODHD met the published multi-market ranking gate. Market coverage and ticker syntax differ, so normalize the user's symbol to the provider's dialect (for example 7203.T versus TYO:7203).

Which API had the broadest observed language coverage?

Linkup repeated relevant results in six requested languages. The French case met the strict rule in 1/2 live rounds. It is a Web retrieval product, not a strict structured financial news feed. Among explicit news APIs in the tested integrations, English was the only language repeated in our fixed company cases.

Which financial news APIs return sentiment?

Finlight exposed a sentiment label alongside structured entities, categories and provenance. Sentiment coverage here measures availability on relevant rows, not sentiment accuracy.

Is there a free financial news API for AI agents?

Several tested providers offer free or trial tiers, but the results above were measured on the plans we tested. Free tiers usually reduce source licensing, history depth and rate limits, so rerun the fixed cases on the exact plan you intend to ship.

How fast are these APIs?

Median end-to-end latency in this run ranged from about 1.6 s (Finlight ticker lookups) to roughly 16 s (Gildata Public Opinion). Treat these as observations from the test window, not as an SLA.

How often should these results be rerun?

Quarterly, and after any material endpoint, source, plan, schema, language, or ticker-routing change. Keep the test cases stable so score changes reflect the product rather than a rewritten benchmark.