SpaceBench

Can you trust AI for space work? SpaceBench asks AI models space-industry questions twice: once on their own (closed book) and once connected to a live satellite and regulatory database (Orbit Sentinel, via MCP). Every grounded answer is frozen from a re-runnable SQL query, and results are reported per practitioner job as trust, verify, or don't use.

Live leaderboard, verdicts and methodology: spacebench.net

This release: corpus 299-21b69d52efdc, revision 2026-10-28, grader 2026-10-10. 8 models, 27 runs, 295 scored questions.

Headline: grounded accuracy, closed book vs. connected to live data

Model	Closed book	With Orbit Sentinel	Wrong when answering, closed	Wrong when answering, with data
claude-opus-5.5	37.3%	94.3%	22.3%	2.0%
claude-sonnet-5.5	34.9%	92.1%	36.2%	2.7%
claude-fable-5.1	39.7%	91.3%	26.5%	2.1%
grok-4.7	18.7%	83.3%	6.0%	3.0%
claude-haiku-5.5	15.5%	82.2%	44.3%	7.2%
gpt-oss-20b	5.6%	68.2%	91.2%	28.3%
gemma4-12b	4.9%	56.3%	90.3%	26.3%

Accuracy is correct answers over the 252 grounded questions. "Wrong when answering" is wrong answers over answers given (abstaining is allowed and not penalised). deepseek-r1-14b was run closed book only. Run conditions for every model are on the methodology page.

Jobs

Questions are grouped by the work a space professional does, because a model can be reliable for one job and unsafe for another:

Job	What it covers
catalog	NORAD and COSPAR IDs, owners, launch and decay dates, orbit class
constellation	Satellites launched, on orbit or lost per constellation
debris	Breakup events, debris and rocket-body populations, collision-risk thresholds
regulatory	FCC, ITU and FAA filings, milestones, surety bonds, licensing rules
market	SEC filings, FCC filing activity by company, bonds and launch licenses
ground	Earth stations, gateways, call signs and where they are
baseline	Space history and orbital mechanics

Verdict rule per AI x job x track: trust if wrong <= 5% and correct >= 80%; don't use if wrong > 20%; otherwise verify. Confidence intervals are in verdicts.

Configs

Config	Rows	Contents
questions	295	252 grounded questions in full (question, answer, tolerance, as-of date, ground-truth SQL). 43 reasoning questions as id, job and category only; their text is held private to resist contamination.
results	7,965	One row per model x track x repeat x question: outcome (correct, incorrect, abstained, no_answer), tokens, latency, cost, tool calls. Model answer text is not included.
verdicts	120	Trust / verify / don't use per model x job x track, with counts and intervals.
leaderboard	15	Per model x track summary: score, accuracy, hallucination rate, cost.
errata	7	Answer-key corrections made after review, with reasons.

raw/ holds the original JSON bundle exactly as published on spacebench.net (corpus.json, leaderboard.json, runs/).

How grounded answers are made

Each grounded question has a ground_truth_sql query over Orbit Sentinel's database, which aggregates public sources (Space-Track SATCAT, FCC IBFS/ICFS, ITU, FAA, SEC EDGAR, national earth station registers). The answer is frozen on its as_of date and checked against a public source before promotion. Numeric answers are graded with a tolerance (tolerance_pct); text answers by word-boundary match. The kind field marks public-verifiable questions (anyone can check the answer from public sources) and os-derived ones (inputs are public, but the derivation needs the database).

Limitations

• Closed-book and tool-grounded runs use different harnesses: Claude models ran through Claude Code on a subscription, Grok through the xAI API, local models through Ollama. Each is labelled per row (provider, harness, tools).
• Answers that depend on live data drift after their as_of date.
• The grounded set is published, so future models may have seen it. The private reasoning set is the contamination check.

Citation

@misc{spacebench2026,
  title  = {SpaceBench: Can you trust AI for space work?},
  author = {Caracappa, Anthony Vincent},
  year   = {2026},
  note   = {Corpus 299-21b69d52efdc, revision 2026-10-28},
  url    = {https://spacebench.net/}
}

Built by Viventine Space Systems. License: CC BY 4.0.