OpenAI 的 840 個 Ashby 職位中,587 個標示 isRemote: true,只有 22 個屬 Remote
587 of OpenAI's 840 Ashby Jobs Report isRemote: true. Twenty-Two Are Remote
作者於 19 November 2026 擷取 16 個 Ashby 職位板的 1,875 條資料,發現 OpenAI 的 840 個職位中,587 個標示 isRemote: true,workplaceType 為 Remote 的只有 22 個。
I run a visa agency in Bali. The part of my week that is not visas goes into scrapers on Apify, and a few of them read hiring data. Ashby is the quiet one: it is the applicant-tracking system behind a lot of well-funded tech companies, every customer gets a public board at jobs.ashbyhq.com/<slug>, and Ashby serves that board through a public JSON API that hands you the whole board with descriptions in one request. No browser, no proxy, no login.
That makes it the cheapest job data on the internet to collect and one of the easiest to misread.
Yesterday's measurement session, 19 November 2026: I asked 20 slugs, 16 boards answered, and the run returned 1,875 live postings with 13.9 MB of description text in 10.8 seconds of wall clock. OpenAI alone was 840 rows. Everything below is from that pull plus the raw Ashby responses behind it.
The remote flag is true on 1,483 of 1,875 jobs
Ashby gives you two things that look like the same thing: a boolean isRemote and an enum workplaceType (Remote / Hybrid / OnSite). Across the 1,875 postings in the raw API:
isRemote: true 1,483
workplaceType Hybrid 928 (all 928 also carry isRemote: true)
workplaceType Remote 560
workplaceType OnSite 46
workplaceType missing 341
Every hybrid job in my sample reports isRemote: true. Per board it gets worse the bigger the company:
| board | jobs | isRemote: true |
workplaceType: Remote |
|---|---|---|---|
| openai | 840 | 587 | 22 |
| notion | 169 | 97 | 1 |
| replit | 37 | 37 | 0 |
| ramp | 185 | 172 | 27 |
| clickhouse | 212 | 182 | 179 |
Replit is the clean demonstration: 37 of 37 postings say isRemote: true and 37 of 37 say Hybrid. If you build a remote-jobs board on Ashby's boolean, you ship 1,483 rows where 560 are the product.
So the Actor does not pass the boolean through. isRemote is derived from workplaceType when the company fills it in, which is how a hybrid job stops being remote. The 341 postings with no workplace field at all are the interesting residue — there the only remote signal is the location text, and Ashby keeps a primary location plus a secondaryLocations array that the board UI shows as a city list.
A real run, workplace: "remote" over openai + notion + replit, returned 37 job rows: 28 that the board itself flagged Remote, and 9 OpenAI roles with no workplace field whose location list reads ["San Francisco", "US - Remote"] — Threat Detection and Response Engineering Manager, ChatGPT Performance Engineer, three Account Directors. Those nine are genuinely remote-eligible and no boolean in the API says so. In the same pull, 67 rows that are not remote list a Remote location among their cities, 57 of them hybrid. Location strings alone are as bad as the flag; you need both.
salaryMin and salaryMax can come from two different countries
Ashby is unusually good about pay: 1,232 of 1,875 postings (66%) carry structured compensation, ranging from 96% of Ramp's board to 0% of Notion's. Numbers, not a sentence to regex — minValue, maxValue, currencyCode, interval. 1,219 are yearly, 7 monthly, 6 hourly, across 13 currencies.
Then there is the tier problem. 274 of the 1,232 priced jobs are "Multiple Ranges": the company published a separate range per location or level — 2 tiers on 156 of them, 3 on 71, up to 8 — and Ashby's summary flattens them into one span. On 178 of the 274, the summary is literally min(lowest tier) .. max(highest tier).
That span is wide: the median (max-min)/min is 52% for multiple-range jobs against 33% for single-range ones. And on 114 of 274 the tiers do not share a currency, in which case Ashby's summary silently reports only the tiers in one of them:
openai Software Engineer, Research - Human Data
row says 295,000 – 385,000 USD
tiers: US 295K–385K USD | London 178K–245K GBP ← not in the row
ramp Fraud Operations Specialist
row says 80,000 – 155,000 USD
tiers: US 80K–155K USD | Canada 105K–185K CAD ← not in the row
locations: New York, Remote (US), San Francisco, Remote (Canada)
That Ramp row is the trap in one picture: the posting is open to Canada, the row's location list says so, and the row's salary is the US-only tier.
My favourite row in the whole pull is Cohere's Software Engineer, Security Agents:
{"title": "Software Engineer, Security Agents", "board": "cohere",
"salaryMin": 195, "salaryMax": 415000, "salaryCurrency": "USD", "salaryPeriod": "year",
"salaryText": "$195 – $415,000 • Offers Equity • Multiple Ranges",
"location": "United States", "allLocations": ["United States", "Canada"]}
The job has three tiers: California/New York/Washington 215K–325K, "USA - All Other States" 195 – 295,000, Canada 280K–415K. Someone dropped a K in one tier, and because the summary is a span, that typo is now the minimum salary of the whole posting while the maximum comes from a different country. Ashby is reporting the company's own data faithfully, which is exactly why you should not feed salaryMin into an average without a floor check: exactly one yearly range in my 1,232 priced rows has a minimum under $1,000, and it belongs to a $415K job.
Practical rule: for medians and histograms, drop rows whose salaryText contains Multiple Ranges or treat them as a bucket, and keep salaryText next to the numbers so a human can see what happened.
One role, eight rows
OpenAI's 840 rows are 793 distinct titles. 25 titles repeat across 72 rows, and the repeats are not a bug in the scraper — they are separate postings with separate ids. "Applied AI Engineer, Codex" exists eight times:
Singapore published 2026-10-08 no salary
Sydney 2026-09-30 no salary
Paris 2026-09-23 no salary
Munich 2026-09-23 no salary
Madrid 2026-09-23 no salary
London 2026-09-11 no salary
São Paulo 2026-09-02 no salary
San Francisco 2025-11-21 211,000 – 278,000 USD
Eight ids, one byte-identical description, one priced row, and the only row carrying pay is the oldest by eleven months. Across the whole pull, 34 of 93 duplicated titles have identical descriptions, and in 10 of OpenAI's 25 duplicate groups exactly one row has a salary. If you count open roles per company, dedupe on (title, description) and not on id; if you build salary stats, the one city that publishes pay is the one that legally must.
Posting dates are soft too: median age 55 days, 290 of 1,875 older than six months, 100 older than a year, and Linear has a live Fullstack Engineer listing first published 2,031 days ago. postedAt is when the requisition was opened, not when anybody last looked at it.
The slug is not the company name
Four of my twenty guesses 404'd: gusto, anduril, scaleai, huggingface. Two of them are not missing, they are on a different ATS under the same slug — boards-api.greenhouse.io/v1/boards/gusto/jobs has 82 jobs, scaleai has 168. A 404 from Ashby means "wrong system", not "no jobs", and the fastest way to find out which is to look at what the company's own careers page links to.
Two slugs behaved worse than a 404: deel and mercury answered 200 with zero jobs. A missing board gets you a free error row —
{"type": "error", "url": "ashby:gusto", "kind": "board",
"error": "board ashby:gusto not found — check the company slug …"}
— but an empty board produces no row at all, so in my three-company control run (gusto, deel, substack) I got 21 Substack jobs, one error row and nothing whatsoever about Deel. Read the run summary row, not the row count.
Two more field shapes worth knowing before you write a schema. allLocations goes up to 20 entries — ClickHouse has a Lifecycle Marketing Manager listed in 20 US cities, and 40 rows list ten or more — so a location column is a join, not a string. And descriptions are not uniformly sized: median 6,585 characters, but 3 rows of 1,875 carried inline base64 images worth 1,017,816 characters, 7.3% of all the text in the run. One Ashby posting is 357,583 characters of which 339,272 are a PNG. That is the row that breaks your CSV export and your embedding job.
Contacts, finally, are a myth on this source: 35 of 1,875 postings contained an e-mail, and 32 of those are one security address Linear puts in a boilerplate footer. Ashby boards are structured and generous about pay; they are not a lead list.
Calling it
{
"companies": ["openai", "https://jobs.ashbyhq.com/ramp", "clickhouse"],
"titleKeywords": ["engineer"],
"workplace": "remote",
"postedWithin": "week",
"includeDescription": false
}
import requests
rows = requests.post(
"https://api.apify.com/v2/acts/lergassy~ashby-jobs-scraper/run-sync-get-dataset-items",
params={"token": "apify_api_..."},
json={"companies": ["ramp", "cohere"], "workplace": "remote",
"includeDescription": False, "maxJobsPerCompany": 50},
timeout=600,
).json()
for r in (x for x in rows if x.get("type") == "job"):
multi = "Multiple Ranges" in (r.get("salaryText") or "")
print(r["company"], r["title"], r["workplaceType"],
None if multi else r["salaryMin"], r["salaryCurrency"],
"|", ", ".join(r.get("allLocations") or [r.get("location") or ""]))
The Actor is Ashby Jobs Scraper — $1.20 per 1,000 job rows, no start fee, error rows free. The 14-board, 1,875-row pull above cost $2.03 and took 10.8 seconds; a single board with descriptions off is 2–3 seconds.
When it is the wrong tool. There is no global directory of Ashby boards, so you cannot search "all remote Python jobs on Ashby" — you bring the companies, which means your real work is maintaining a slug list. Per-location pay tiers live on the job page, not in the board API, so if you need "what does this pay in Toronto" you have to open the posting. Applicant counts, recruiter identity and hiring velocity are not published at all. And description is whatever the company wrote — if you want normalised seniority, skills or benefits, that is an extraction job on top of this data, not a field you can scrape.
來源:DEV Community · dev.to