Company Funding Benchmark

Two boards, split on how old the round is

Finding a round announced this week and holding a correct historical record are different capabilities, and a vendor can be strong at one and weak at the other. Averaging them into a single number hides that, so they are measured separately. The split is on the announcement date: rounds inside the trailing 30 days are freshness, everything older is enrichment.

Three ways to answer the same question

Both boards are subdivided by how a provider produces its answer, because a single ranked list quietly compares things that are not substitutes for each other. Every provider is asked the same question about the same companies and judged by the same policy, so the rows remain directly comparable. The grouping is a reading aid, not a separate scoring rule, and nothing is excluded or weighted differently.

• Long Running Agent APIs dispatch an agent that researches each company at request time, reading sources and deciding what the latest round was. They are the slowest and most expensive per call, and they are priced on how much work they do.
• Web search APIs query a search index and extract the answer from what comes back. Fast and cheap per call, with accuracy that tracks how well a round was covered on the open web.
• GTM data providers return a stored record from a maintained company database. Latency and per-record cost are predictable, and coverage is set by the vendor's own ingestion rather than by the query.

The distinction is about mechanism, so Exa and Parallel each appear in two groups. Exa's Agent API and Parallel's Task API dispatch an agent that researches the company; Exa's two /search arms and Parallel's Responses API resolve the question against an index. The instruction and output schema sent to all of them are identical, so the endpoint is the only variable. The grouping also explains results a single list would make look contradictory, since an agent that researches on demand can lead on historical rounds and still be beaten on rounds announced last week.

What each column means

Every result is aligned to the same company domains. A latest stage counts as correct under the documented Ground Truth matching policy, which includes selected stage equivalents and exact date or amount evidence for non-Series disagreements. A missing provider stage lowers the correct-stage result, except that a blank vendor stage is accepted when Ground Truth is Undisclosed.

• Measured on is that provider's own denominator: each provider/company/snapshot observation it was scored on, and how many snapshots those came from. A company repeated in a later snapshot counts again because it is a new query moment. Two rows with different values here were not measured on the same cohort.
• Latest stage correct is the headline. Of the scoreable companies, the share where the provider named the right stage.
• Correct when returned divides correct answers that actually returned a stage by every case where the provider returned one.
• Funding fields returned counts how often latest stage, date, round amount, total raised, and round count were populated. It does not check whether the four non-stage fields are correct.
• Official docs links each provider's own API, endpoint, or access documentation. These describe vendor configuration and do not influence the score.
• Est. cost is modeled from observed billing units at published rates, scaled to the companies that board scored.

• Whether the other funding fields are right. Only the latest stage is judged for correctness with announced date as fallback. Round amount, total raised, and round count are counted for presence only, so a provider can score well on funding fields returned while carrying wrong amounts.
• Full funding history. One round per company is evaluated: the latest one. A provider with an excellent complete round history and a weak latest round scores badly here.
• Investor, valuation, or board data. Not requested and not scored, even where a provider returns it.
• Everything else the vendor sells. Most of these are broad platforms. Contact data, intent, technographics, and workflow tooling are different products. A result here says nothing about them.

• Rows do not always contain the same measurement observations. Providers join at different times, so each is scored on its own denominator. Comparing two rows with very different Measured on counts compares performance on two different cohorts.
• Cost is modeled, not invoiced. It is derived from observed billing units at published rates and scaled to each board. Plan minimums, negotiated rates, and credit rules will move real spend. Explorium's figure is an explicit request proxy because its per-endpoint credit burn is not public.
• Latency is one runner on one network. Median request time is indicative, and batch endpoints are timed per batch rather than per company.
• Reviewed exports are not timed runs. Crunchbase and Harmonic are mapped from supplied exports, so they carry no latency or cost and cannot be read as evidence of API behaviour.
• Freshness is a rolling number. Each snapshot is a different set of companies. A change between snapshots can be a real shift in a provider, or a harder cohort.