
Why teams trust the engine
Real astronomy, not approximations.
Open Ephemeris computes planetary positions straight from the NASA JPL DE440 kernel, within about 0.1 arcsecond of JPL Horizons near the present — the same data behind spacecraft navigation. Here's the proof, not the pitch.
AI chart accuracy test
Does AI get your birth chart right?
We gave Claude the same 210 birth dates twice. First on its own. Then with Open Ephemeris connected. Each square below is one chart.
Claude on its own
1 of 210
charts had every sign right. It would not answer 40, and 169 had at least one planet in the wrong sign.
- 1 all signs right
- 169 at least one wrong sign
- 40 no answer
Claude with Open Ephemeris
209 of 210
charts had every sign right. 1 chart had a wrong sign.
- 209 all signs right
- 1 at least one wrong sign
Of the charts Claude did answer on its own, 99% had at least one wrong sign. It warned the reader that its figures were estimates in every reply. The warning did not make the chart right.
Where it goes wrong
Share of answers with the wrong sign. Claude usually works out the Ascendant by hand and takes the planets from memory. It is mostly right about the Sun and the slow outer planets, and most often wrong about Mars, Venus, Jupiter and Mercury.
What the misses have in common
With the tool, 3 of 290 replies had a wrong sign, all for the same reason. The tool returned the historical time zone, and Claude replaced it with its own assumption: standard time for a Los Angeles birth in December 1948 (California stayed on daylight time through the end of 1948), and UTC+8 for Beijing in 1943. The time-zone database agrees with the tool in both cases.
How close the numbers are
With the tool, every planet came back within 0.007° at the median. On its own, Claude's Mars was off by a median of 46°.
Ask twice, get the same chart?
We asked for 40 charts three times each. On its own, Claude changed its Ascendant sign on 14 of them between runs. With the tool it changed on 2, both from the time-zone overrides above.
The data, scoring code and every raw reply are public: github.com/openephemeris/ai-chart-accuracy
Useful insight begins with correct information.
How we tested(show)
- 210 invented birth dates: 30 per decade from the 1940s to the 2000s, across 38 cities in both hemispheres, including polar latitudes and births close to a clock change.
- The same plain request every time: date, time and city, tropical zodiac, Placidus houses, and a fixed answer format so scoring needs no judgement. A chart counts as right only if every planet and the Ascendant are in the correct sign.
- The answer key is NASA JPL DE440. We checked it against JPL Horizons on 30 charts: planets agreed to within 0.0001°, and the Ascendant matched the textbook formula to within 0.01°.
- Claude Sonnet 5.5 at chat-style settings, run on 2026-11-03. No system prompt. The tool run used the Open Ephemeris connector.
- We make the tool, so this is our test, not an independent one. The dataset, answer key, scoring code and every raw reply are public, so anyone can check or repeat it.
- Not yet tested: ChatGPT, other models, other ways of asking.
Built on Open Ephemeris
ELA Map
ELA Map — our own locational astrology and Human Design platform — is built directly on the Open Ephemeris engine. Every line, crossing, relocated chart, and bodygraph it renders comes from the same public endpoints, which is why they stay fast, precise, and honest: we depend on them ourselves.
In their words
“I built the astrology tool I always wanted, Chinese, Vedic and Western, with Claude over a couple of nights. Working like a charm.”
Measured against JPL Horizons
Run on 2026-10-28: 10 bodies at 12:00 UT on 1 January every 50 years, 1550 to 2650. 171 of 230 points compared; Horizons had no body-centre data for the other 59 (outer planets on some dates). Error is the angular gap in arcseconds, longitude and latitude together. We ran this ourselves, so it is not an independent test. Frame, nutation and time-scale conventions could add a few tenths.
| Years | Points | Median | 95th percentile | Worst |
|---|---|---|---|---|
| 1950 to 2050 | 30 | 0.09″ | 0.34″ | 1.62 (Moon, 2050)″ |
| 1900 to 2100 | 50 | 0.12″ | 1.05″ | 10.4 (Moon, 2100)″ |
| 1800 to 1949 | 28 | 0.35″ | 1.00″ | 1.29″ |
| 1550 to 1799 | 30 | 0.77″ | 2.07″ | 2.18″ |
| 2101 to 2199 | 10 | 0.91″ | 25.0″ | 25.0″ |
| 2200 to 2299 | 16 | 3.97″ | 95.0″ | 95.0″ |
| 2300 to 2649 | 43 | 25.1″ | 593″ | 715″ |
Worst single point: Moon, 2650, 905″. Near the present the gap is under an arcsecond for most points. It grows away from today and sharply after about 2100. For the Sun and Moon the far-future gap matches a clock offset (Delta T) between the two systems, and Delta T cannot be predicted that far, so this run does not say which side is closer. Not tested: asteroids, Chiron, nodes, house cusps, sidereal modes.
Rerun it yourself: download accuracy-vs-horizons.mjs (Node 18+, no dependencies) and run OE_API_KEY=<your key> node accuracy-vs-horizons.mjs. The full grid is 230 calls at 1 credit each; add --step 100 for a 120-call run inside the free 150 credits.
Why the numbers are trustworthy
Straight from JPL
Positions come from the NASA JPL DE440 kernel via the CALCEPH and SuperNOVAS libraries — the peer-reviewed standard for solar-system geometry.
Self-validating
Every hour a built-in check compares one position (the Sun, Moon, Venus, Mars or Jupiter) with JPL Horizons and raises an alert if it drifts by more than 0.1 arcseconds.
Secure & typed
RFC 7807 errors, authenticated and rate-limited requests, vulnerability scans on every build, and typed MCP tools so an AI calls real math.
Try it against your own chart.
Explorer tier is free — 150 credits to start (one-time), no card required. The fastest proof is your own.