What Federal Procurement Data Costs in Tokens

If you fed one fiscal year of awards data from USASpending's REST API into Claude Haiku, it would cost about $9,500 in token costs. Pulled from Tango, the same awards cost about $3,600.

The other ~$5,900 is formatting: duplicated descriptions, empty fields, and codes the model never asked for.

To be clear, it would be silly to handle data this way at scale. But token efficiency is a very real issue when it comes to working with LLMs (see Anthropic's write-up on context engineering for a good explanation).

A model's context window is a finite resource, and every token you put into it has to justify its cost. Tokens don't just cost more in terms of money. LLMs perform worse with extra tokens. On raw federal data, most of your tokens aren't eating up context window on filler. We wanted to measure how much.

The benchmark

We took a federal award and pulled it two ways: once from the public USASpending API and once from the Tango API. Then, we minified both responses so formatting couldn't skew the count, and tokenized them using tiktoken. We did this across seven awards, spanning agencies from DLA to the VA, and sizes from $185 to $381M.

# Raw USASpending award record
curl -s 'https://api.usaspending.gov/api/v2/awards/CONT_AWD_HC102823F0318_9700_HC102822A0002_9700/' \
  | jq -c . | ttok

# The same award from Tango
curl -s 'https://tango.makegov.com/api/contracts/CONT_AWD_HC102823F0318_9700_HC102822A0002_9700/' \
  -H "X-API-KEY: ${TANGO_API_KEY}" \
  | jq -c . | ttok

Although ttok counts with a GPT tokenizer (not Claude's tokenizer), the ratio is the robust part, and it holds across models.

On average, the USASpending API cost about 2.6 more tokens than Tango for the exact same award. This means you're using more tokens for the same amount of information, leading to higher costs and worse performance.

The interesting part is why, and it's part of the reason this data is hard to work with in the first place. A raw contract award record spends most of its payload on structure, not substance:

• Code + description. Roughly fifty fields each ship the raw code and its English description as separate keys ("extent_competed":"A" sitting right next to "extent_competed_description":"FULL AND OPEN COMPETITION").
• Duplicate agency trees. The funding agency and the awarding agency each carry the full department → sub-agency → office hierarchy, usually nearly identical.
• Null scaffolding. A lot of data comes back null, which adds to the token counts.
• Pre-exploded hierarchies. Three to four levels of NAICS and PSC code-plus-description you didn't request.

All awards come with that verbosity. That's why a $185 delivery order for "DOUGHNUTS, FRESH, VARIETY PACK" at DLA costs an agent essentially the same number of tokens as a $381M IT contract at the VA. The schema overhead is the same size for everything.

Put a dollar on it

Federal contracting runs on the order of 5.5 million prime contract awards every fiscal year. At Claude Haiku 4.5's input rate of $1 per million tokens:

Award reads	Raw API	Tango	You overpay
1,000	$1.73	$0.66	$1.07
1,000,000	$1,732	$658	$1,073
1 fiscal year (~5.5)	$9,526	$3,619	$5,907

You can do even better with response shaping. Simply pass a shape parameter and the API returns just those:

curl -G 'https://tango.makegov.com/api/contracts/' \
  -H "X-API-KEY: ${TANGO_API_KEY}" \
  --data-urlencode "shape=key,piid,description,recipient(display_name,uei),total_contract_value" \
  --data-urlencode "limit=25"

Full details are in the response-shaping guide.

The bottom line

The takeaway isn't that Tango's API is inherently better than USASpending in all cases. Instead, the takeaway is that Tango is built for builders who want data that matches their workflows and can help reduce costs and improve quality over time. If you're using workflows that incorporate LLMs, Tango is going to give you better otucomes than raw federal procurement data. Because small decisions like better schema design give you more token leverage: more of the window left for reasoning, and better answers on what's there, and at a lower price.