What does a task cost on plori? Two real tasks and parallel workers on v349

• Run a repository's test suite. Clone github.com/pallets/itsdangerous at tag 2.2.0 (commit 096c8d42545d3b68ea21a4f890fb2b2d8979c0bd). Install it and its test needs, run pytest, and write the counts to results.txt. Pass: the saved file reads passed=297 failed=0 errors=0 skipped=0.
• Research from several sources. Daylight saving time rules in the European Union, the United States and Mexico, and the laws passed or proposed since 2018. At least five sources, one Markdown file per region plus a summary, every source cited with its full URL. Pass: the saved notes exist, contain at least 5 distinct URLs, and name all three regions. This check tests the structure of the notes. It does not check that the cited facts are correct.

Every run passed on the first message, in one run, with no approval or question card.

In the research task, the model also decided how often the agent searched. Per run, gpt-6-luna made 23 to 52 metered web searches, claude-haiku-5.5 made 3 to 7, and the gpt-6.1-sol run made 8.

• In this v349 measurement on 14 October 2026, run time was charged per started minute ($0.01), so every short run paid at least $0.01. Run time costs $0.01 per minute, billed by the second.
• The research tools cost is web search only. Web fetches and the platform's own checks cost $0.
• One research run made one advisor call on claude-opus-5.5, which cost $0.1700. It is part of the model cost above.
• Prompt cache hit rate (cached prompt tokens over prompt tokens, executor rounds): tests 89.7%, research 81.8%. The single gpt-6.1-sol research run had 19.0%.
• VM claims: all 10 test-suite runs took the warm-pool VM. Of the research runs, 9 took the warm-pool VM and 1 booted a new VM.

• The first 100 web_search calls each month are free, then $0.01 per call. A search that returns no results is not charged.
• All 10 research runs used one account, so they shared its 100 free calls. They made 158 metered searches. The free calls ran out during the 8th run, and 58 searches were billed: $0.5800.
• No single run made 100 searches (the most was 52), so each run on a fresh account would have paid $0 for search.
• At list price, the 158 metered searches cost $1.5800.
• The run events show 185 web search calls, 27 more than the metered count. Those calls have no usage record, which is consistent with the rule that a search with no results is not charged. We did not check this call by call.

• In this v349 measurement on 14 October 2026, run time was most of the cost: each worker run paid at least one started minute ($0.01). Copies, the task group, the changes and the accepts added no charge.
• The warm pool holds one ready VM, but in some repeats more than one worker was recorded as a pool hit. Under load some of those claims took up to 11.6 s. We did not trace why. Claims that booted a new VM took 2.6 to 10.8 s.
• The fleet started with one VM host. During N = 8 repeat 1, the autoscaler logged a scale-up at 14:30:31 UTC and created a second VM host at 14:30:34. The autoscaler logs no "ready" line. The first VM was placed on the new host at about 14:33:10, so it was ready within 2 min 36 s of creation. Repeat 1's eight workers had all been placed on the first host before the new host existed. The new host served repeats 2 and 3.

A third task, "build and activate a scheduled workflow that posts a Hacker News front-page digest", could not be completed on v349. We stopped it after two trial runs. First, the workflow's chat notice is a fixed status line, so it cannot carry a digest. Second, with the digest kept in the run output, the agent wrote a step without its required authentication field. The workflow engine rejected the step before it ran, but the error the agent received did not list the validation issue, and the agent's own validation check returned valid. The agent ran 11 tests, of which 10 were rejected, and activation was refused. That trial cost $0.225. These are known gaps that we are fixing.

• Production release v349, driven from one host on the public internet through the public REST API. The fleet was shared with real users, our hourly test probe and, during the task runs, another benchmark.
• Tasks: one new registered free account per task, marked as test traffic. Each sample is a new agent with its own Workspace on that account, so the 10 samples of a task share the account's monthly allowances. Samples ran one at a time.
• Workers: one account per N, three free and one on the power plan.
• Each test account received a $20 welcome credit. What a new account receives is on the pricing page.
• Task times come from the server's run record. Costs come from the account balance and the usage records. Worker dispatch is a client timestamp, and every later phase is a server timestamp. In the N ≤ 4 repeats the first worker's run was admitted 0.23 to 0.77 s after dispatch, so any clock difference is under a second.
• Percentiles use linear interpolation between closest ranks. p50 needs n ≥ 5 and p90 n ≥ 10. Smaller groups list their values.
• No sample or repeat was excluded.

• n = 10 per task with a four-model draw. The per-model rows have n = 1 to 6, and the pooled percentiles describe this draw, not a fixed model.
• Two tasks. This page does not measure longer tasks, tasks that need several messages, or other models.
• The worker edit is tiny on purpose, so the phases measure platform overhead (VM claim, save, compare, accept) more than model work. The accepts were driven through the public API about once every 0.5 s. An agent that accepts the changes itself adds its own model time.
• 3 repeats per N on a shared production fleet that started with one VM host. The N = 8 repeats ran on different fleets (one host, two hosts, mostly the new host).
• All numbers are from 2026-10-14 only.
• No cross-vendor comparison: this page measures plori against itself, under a stated method, and claims nothing about other platforms.