OLLAMA CLOUD, UNLEASHED

Our tests

Every experiment behind the field manual, with its question, method, raw numbers, date and verdict. These are our own results, not vendor claims.

How to read these

  • Each card is one experiment. The numbers are as we recorded them.
  • Two accounts, one old plan and one new plan. Your mix may differ.
  • Session points are the share of an old-plan five-hour session that one call used.

The experiments

  1. Our test

    Old-plan sessions run in fixed five-hour blocks

    Question
    Do old-plan five-hour session windows roll forward with each request, or do they start on fixed boundaries?
    Method
    We compared each observed session reset with a fixed five-hour grid that ends at minute :55.
    Raw numbers
    • Six resets observed, zero drift.
    • Blocks end at minute :55. Five and 24 share no common factor, so the boundary walks through every clock hour over five days.
    Date
    August 2026
    Verdict
    Fixed five-hour blocks ending at :55. Not a rolling window.
    Caveats
    • Two accounts, one old plan and one new plan. Your mix may differ.
    • Old-plan accounts only.
  2. Our test

    The old-plan weekly window is seven days

    Question
    Is the weekly window a fixed seven days from one reset to the next?
    Method
    We compared consecutive weekly resets.
    Raw numbers
    • One reset landed exactly seven days after the previous one.
    Date
    August 2026
    Verdict
    A fixed seven-day cycle.
    Caveats
    • Two accounts, one old plan and one new plan. Your mix may differ.
    • One observed pair is a small sample.
  3. Our test

    Calibrating the old-plan estimate

    Question
    How much of a session or a week does one request use?
    Method
    We set the request count against a dashboard reading, and used two independent session measurements.
    Raw numbers
    • Weekly: 64.9% used at 3,672 requests (8 Oct 2026, mostly glm-5.3-flash traffic).
    • Session: about 0.10% of a five-hour session per request. Two independent measurements: 0.103 and 0.093.
    Date
    8 Oct 2026 for the weekly reading. The session figures are from our earlier measurements.
    Verdict
    Usable as an estimate, labelled EST. Ollama stopped reporting the quota percentage on about 7 Oct 2026.
    Caveats
    • Two accounts, one old plan and one new plan. Your mix may differ.
    • The weekly figure reflects one traffic mix. Calibrations are not API measurements.
  4. Our test

    Per-call cost, round 1

    Question
    Which of two models uses less of a session for the same call?
    Method
    Twenty identical calls per model. Each prompt was about 400 to 500 tokens.
    Raw numbers
    • glm-5.3-flash: 0.030 session points per call.
    • deepseek-v4-flash:0731: 0.015 session points per call.
    Date
    27 Aug 2026
    Verdict
    In this round the DeepSeek model used about half the points that GLM used.
    Caveats
    • Two accounts, one old plan and one new plan. Your mix may differ.
    • One prompt shape. Old-plan session points only, not dollars.
  5. Our test

    Per-call cost, round 2

    Question
    Does the ranking hold one model generation later?
    Method
    Fifteen identical calls per model, with the same prompt shape as round 1.
    Raw numbers
    • glm-5.3-flash: 0.093 session points per call.
    • deepseek-v4.1-flash: 0.140 session points per call.
    Date
    30 Sep 2026
    Verdict
    The ranking reversed. DeepSeek v4.1 Flash used about 1.5 times GLM's points. The cheapest model changed within one generation.
    Caveats
    • Two accounts, one old plan and one new plan. Your mix may differ.
    • Small sample. Re-test before you commit to a model.
  6. Our test

    A hard stop at $60.00 on a new Pro plan

    Question
    What happens when the included credits on a new plan run out?
    Method
    We watched cumulative usage on a new-plan Pro account until its calls stopped.
    Raw numbers
    • Calls stopped at exactly $60.00 of cumulative usage, the Pro included credits.
    • After that, calls returned HTTP 429 with the monthly usage limit message.
    Date
    Logged in our test notes. Verified 12 Oct 2026.
    Verdict
    Included credits are a hard stop. Calls fail until the monthly reset, or until credits are added.
    Caveats
    • Two accounts, one old plan and one new plan. Your mix may differ.
    • One Pro account. We did not test Max or Team.
  7. Our test

    An agent-heavy account over 30 days

    Question
    What does an agent-heavy account look like on a new plan?
    Method
    We read the 30-day usage totals for one agent-heavy, new-plan account.
    Raw numbers
    • About 1.47B input tokens, of which about 93% were cached input.
    • About 12.8M output tokens.
    • About 142,000 input and 1,200 output tokens per request, on average.
    Date
    The 30 days to 12 Oct 2026
    Verdict
    Stable prompt prefixes decide most of the input cost. Cached input is priced far below fresh input.
    Caveats
    • Two accounts, one old plan and one new plan. Your mix may differ.
    • Your agent's prompts and model choice will change the mix.
  8. Our test

    The usage API allows about five reads a minute

    Question
    How fast can a monitoring script read the usage API?
    Method
    We made repeated usage reads and watched for the rate-limit response.
    Raw numbers
    • About five usage reads per minute. Past that, the API returns 429 "too many requests".
    • No rate-limit headers are returned on chat or usage responses.
    Date
    Verified 12 Oct 2026
    Verdict
    Space reads at least 12 seconds apart. Eight seconds is too fast for about five reads a minute.
    Caveats
    • Two accounts, one old plan and one new plan. Your mix may differ.
    • Limits can change. Check them before you rely on them.

Check the real numbers

Our tests cover two accounts. Paste your own key on the home page to see your plan’s windows.

Check your own usage