AI Research — yourtech?us

What an agent fleet asks of a model server

Sun, 25 Oct 2026 00:00:00 +0000

Written by an AI agent; prepared in an interactive GPT-6 Astra session with John; reviewed by John

GothMUD is an environment where AI agents interact with one another. Alongside it, Chassy runs as another agent workload. They generate conversations, tool calls, and background activity—and provide a place to learn about agent harness design and model behavior.

That activity is not a productivity score. It is, however, a substantial workload for a model server. Understanding its shape helps explain why a configuration that feels fast in one conversation can struggle when several agents share it.

Measurement window: October 14, 2026, 00:00 through October 22, 02:40 EDT. This report was prepared October 25 from an analysis of session transcripts and serving logs.

The workload in numbers

During this window, the agents used Qwen3.6-35B-A3B on a local GB10 server. Four agents were retired later, on October 24. To estimate the needs of the remaining group, the analysis separates their earlier traffic from the retired agents’ traffic.

These are daily averages. The retained group consists of seven GothMUD agents plus Chassy; it was not operating alone during the measurement window.

Daily measure	Retained group	Four subsequently retired agents
Model calls	4,597	3,900
Uncached input tokens	6.48 million	4.23 million
Cached input tokens	138.42 million	128.33 million
Output tokens	340,000	347,000

For the retained group, that is about 145 million input tokens per day, roughly 96% served from the prefix cache. Its output averages only about four tokens per second across a day, though actual requests arrive in bursts.

The large input total does not mean the agents receive 145 million tokens of new information every day. Much of it is the same long context presented again on successive turns. Prefix caching lets the server reuse computation for an unchanged beginning of a prompt.

This is why output speed alone tells little about the cost of serving the group. Context length, reuse, simultaneous requests, and how agents trigger one another all matter.

What changed with a different model

The October 24 move to Flash-Next changed more than response speed. The operational analysis recorded more model calls per incoming event and faster context growth. Turning reasoning off shortened individual responses, but the agents also talked more and woke one another more often.

Observation	Earlier 35B period	Flash-Next, low reasoning	Flash-Next, reasoning off
Model calls per incoming event	2.43	3.25	3.34
Output tokens per model call	118	370	158
Context growth per model call	290 tokens	606 tokens	403 tokens

These are different production periods, not a controlled benchmark. Routing, reasoning settings, and agent activity changed. The counts also required correction for duplicated session records.

After four agents were retired, a later observation found approximately 485 fleet requests per hour, an eight-second median response, and a 31-second 90th-percentile response. That is a snapshot of the smaller fleet on a different model, not directly comparable to the earlier daily averages.

The useful lesson is that reducing individual response time need not reduce total demand. In a system where agents interact, their behavior helps determine the workload.

The hosted bill was a surprise

An initial estimate in this analysis put hosted operation at roughly $7–8 per day. Then the OpenRouter export showed $27.83 for one complete hosted day. Seeing that charge was a shock—and the estimate needs an explanation, not a second table with a different monthly total.

The estimate was not a reliable forecast of that day. It used the later retained group’s earlier traffic and local cache behavior. The hosted day covered a larger group, a different volume of work, and a substantially lower cache-hit rate.

The complete day is October 23 UTC, or October 22 at 20:00 through October 23 at 20:00 EDT. It sits within the hosted interval that began the evening of October 22 Eastern time. The main hosted route ended around 21:23 Eastern on October 23, already October 24 UTC.

Measure	Earlier retained-group daily average	Earlier full-fleet daily average	Observed hosted day
Model calls / provider requests	4,597	8,497	7,589
Input tokens	144.90M	277.46M	364.46M
Cached input tokens	138.42M	266.75M	245.22M
Reported cached share of input	95.5%	96.1%	67.3%
Output / completion tokens	0.340M	0.687M	3.749M

The first two columns come from local session records over October 14–22. The third comes from the provider export for one day. The export identifies the model and API key, but does not identify individual agents, so this is not a fully matched fleet-only comparison. Provider requests and recorded assistant calls may also differ because of retries or other traffic.

Even with those limits, the volume mismatch is large. The hosted day recorded fewer requests than the historical full-fleet daily average, but more input and over five times as much output. It also recorded about 119M uncached input tokens, versus the earlier full fleet’s 10.71M per day. We have not established how much of the change came from agent behavior, serving settings, retries, or accounting differences.

Reconciling $7.88 with $27.83

The pricing calculation is straightforward when the units and token categories agree:

Cost = uncached input × input rate + cached input × cache-read rate + output × output rate.

Token quantities below are priced per million. The first steps use the same historical price assumption throughout: $0.10 input, $0.05 cached input, and $0.90 output. Each step changes one assumption from the previous row.

Calculation	Daily cost
Earlier retained group, with its recorded local cache reuse	$7.88
Include the four later retired agents, using their historical traffic too	$15.03
Replace input volume with the hosted day’s 364.46M; retain the full fleet’s historical cache share and output	$19.54
Also replace output with the hosted day’s 3.749M	$22.30
Also replace cached input with the hosted day’s recorded 245.22M	$27.56
Price actual token totals separately for each provider using the saved rate snapshots	$28.06
OpenRouter’s recorded charge	$27.83

This is an arithmetic reconciliation, not a controlled experiment or a unique attribution of causes: changing the order of the steps changes the size assigned to each step. It shows why varying prices and cache assumptions around the original small workload could never explain the observed bill.

Provider-specific reconstruction comes within $0.23, or about 0.8%, of the recorded charge. Most of the remaining difference is in the Parasail rows, where the saved rate snapshot reconstructs $2.01 against $1.79 charged. The exact request-level pricing difference remains unresolved. Reasoning tokens are not added on top of exported completion tokens in this calculation.

Otari’s gateway estimate for the same UTC day was $33.72 across 7,560 successful requests, compared with OpenRouter’s $27.83 across 7,589 records. The provider charge is 17.5% lower. That is a separate comparison: conservative gateway pricing overestimated the observed day even though the original workload-based estimate underestimated it. The 29-request discrepancy still needs reconciliation.

At the observed provider charge, 30 identical days would cost about $835. That is a constant-usage extrapolation, not a forecast for the reduced fleet or evidence that every day costs the same.

Making the next estimate more trustworthy

The lesson is not to discard token-based pricing. It is to stop treating one workload’s token volume and cache behavior as a forecast for another configuration.

For a Haiku or Sonnet comparison, the observed hosted day is a better starting point than the retained-group estimate. Repricing it would answer a limited question: what would these input and output quantities cost under the proposed provider’s pricing rules? It would not establish what that model would actually consume while running these agents.

A useful estimate should therefore:

Fix the population and window. Name the agents, include retries and background calls where possible, and use the same UTC interval across sources.
Check the accounting against a known charge. Reconstruct the existing provider’s bill before applying another price schedule. Keep unmatched requests and missing usage visible.
Use the target provider’s billing categories. Tokenization, cache eligibility, cache-write charges and lifetime, reasoning, and tool charges may differ. The 96% local hit rate is not a portable assumption.
Separate a repricing scenario from a behavioral forecast. A different model can make more calls, produce more output, or trigger more agent interactions. Show that uncertainty instead of a single confident monthly number.
Measure a bounded pilot before committing to continuous operation. With an explicit spending limit, measure representative agents on the target model, compare the resulting usage and charge with the estimate, and repeat across more than one day before extrapolating.

The first four steps can improve a paper estimate. The pilot would be a separate, deliberately authorized paid experiment; none has been run for this report.

What local costs mean here

The GB10 is already owned, and local requests have no per-token API charge. That makes it convenient to run this experiment without each conversation increasing a hosted bill.

It does not make operation free. Electricity, hardware purchase and wear, maintenance, and competing uses of the machine are outside this calculation. A complete purchase comparison would need those costs and a suitable hosted alternative with comparable behavior.

This particular fleet is also not representative of every AI workload. A focused assistant task or coding session can have a very different economic balance. The purpose of measuring it is to establish the workload before making that comparison.

Method and limits

The daily volumes come from assistant-message usage records in agent session transcripts, deduplicated by message ID over approximately 8.1 days. The analysis ends before a subsequent restart-loop incident. Model-change observations also use gateway and model-server logs.

The October 24 retirement roster determines the two groups shown above. Rounded daily values should not be mistaken for exact billing quantities. Hosted prices are historical endpoint snapshots collected October 18–25; 30-day projections assume unchanged volume and prices.

The wider token-use overview includes Jane, FreeFiFoFum, and RSP, with different measurement windows and accounting coverage.

Where my AI tokens go

Sun, 25 Oct 2026 00:00:00 +0000

Written by an AI agent; prepared in an interactive GPT-6 Astra session with John; reviewed by John

AI runs here for several reasons: an assistant handles conversations and scheduled work; social agents interact; FreeFiFoFum analyzes financial data; and experiments test models and the infrastructure around them.

Token counts help establish the scale of those workloads. They do not measure productivity. Some activity produces useful task results; some is recreation or a way to learn about harness design and model behavior. This page keeps consumption and spending visible without treating a larger number as a better outcome.

Evidence reviewed: October 25, 2026. This is a partial view assembled from records with different windows, not a single monthly bill. The intention is to maintain it as better accounting becomes available, without promising a reporting schedule.

What is covered

Workload	Available evidence	Window or unit
GothMUD and Chassy	Local usage, observed hosted charges, and a reconciled cost estimate	Local: October 14–22 EDT; hosted comparison: October 23 UTC
Jane	Selected gateway request totals and corrected cost estimates	September 9–October 24, 23:45 EDT
FreeFiFoFum backtesting	Saved analysis of calls and tokens	One historical experiment with 27 rebalance dates
FreeFiFoFum recurring committee	Older weekly-run projection; current nightly total not established here	Coverage incomplete
RSP agent sessions	Costs reported in 46 saved session results	September 28–October 23
RSP consultants	Four reported Opus costs; four GPT costs missing	October 2, 4, 11, and 18

These rows should not be summed into a consumption total. They mix daily averages, cumulative periods, and an individual backtest. Some source reports also describe the same requests from different viewpoints.

Existing hardware, hosted usage, and subscriptions

“Zero cost” in a request log can mean several things:

Execution or billing category	What the accounting means
Local, on existing hardware	No incremental API charge. Electricity, equipment, maintenance, and other uses of the machine are not priced here.
Hosted pay-as-you-go	A recorded or reconstructed usage charge, which still needs reconciliation with the provider’s bill.
Hosted subscription or prepaid plan	A request may show zero while the subscription has a separate charge.
Free allowance	No charge within that allowance; it is not necessarily a sustainable price for more usage.
Missing cost record	Unknown, not zero.

There is no assumption that owning a model server is the cheaper way to accomplish a task. A price comparison needs the workload, quality requirements, usage frequency, and the costs included on each side.

The agent fleet: repeated context dominates

For seven retained GothMUD agents plus Chassy, the October analysis finds daily averages of 4,597 model calls, 144.90 million input tokens, and 340,000 output tokens. About 138.42 million input tokens were cached.

Those agents were part of a larger fleet during the measurement window. The figures select the agents that remained after a later reduction; they are not a measurement of the smaller fleet operating alone.

The hosted day changed the estimate

The initial $7–8/day estimate used that smaller group’s earlier traffic and local cache reuse. It was not a reliable forecast of the larger workload that actually ran hosted.

For October 23 UTC—October 22 at 20:00 through October 23 at 20:00 EDT—the OpenRouter export records 7,589 requests, 364.46M input tokens, and 3.749M completion tokens. About 67.3% of input was reported cached. That is materially different from the earlier full-fleet averages of 277.46M input, 0.687M output, and 96.1% cached input per day, even before considering the smaller retained group.

Cost measure	Amount	What it represents
Initial retained-group scenario	$7.88/day	Earlier traffic for eight agents, assuming local-like cache reuse
Reconstruction of the observed hosted day	$28.06	Actual provider token totals priced with saved provider-specific rates
OpenRouter charge for that day	$27.83	Provider-reported charge for the exported model/API-key scope
Otari estimate for that day	$33.72	Gateway deployment pricing across 7,560 successful requests

The large underestimate came from the workload assumptions: population, token volume, and cache behavior. Repricing the observed usage gets within $0.23, about 0.8%, of the provider charge. Most of that residual comes from the Parasail records; their exact pricing difference remains unresolved.

The provider charge is also 17.5% below Otari’s estimate. These are two different estimation errors: the initial scenario understated demand, while the gateway’s pricing overstated the cost of the observed day. The provider and gateway counts differ by 29 records, so their comparison is not yet a matched-request reconciliation. They describe overlapping activity and must not be added together.

Thirty days at the observed daily charge would be about $835. That is a constant-usage projection, not a monthly bill or a forecast for the reduced fleet. The full export totals $29.74 across 8,180 requests over October 19–24 UTC; that wider export must not be mistaken for one day’s usage. Neither scope establishes individual-agent attribution from provider records alone.

The fleet-load report walks through the estimate-to-charge reconciliation. It also explains what is still unknown about the changes in traffic and behavior.

What this means for the next price comparison

An estimate for Haiku, Sonnet, or another model should start with a defined workload and reporting window, then use that provider’s token, cache, and other billing rules. Repricing the observed Qwen workload answers what those quantities would cost; it does not predict how another model will behave. Call counts, output length, cache reuse, and interactions between agents can all change.

Before trusting a monthly forecast, first reconstruct a known charge, show the assumptions and uncertainty, then validate with a separately authorized, spending-limited pilot. No such target-model pilot has been run for this report. More than one representative day would be needed to judge whether an observed rate is typical.

Jane: accounting changed the cost story

Jane is a personal assistant with both an ongoing conversation and scheduled work. The available gateway report covers successful requests on two Jane routing policies, not all of Jane’s activity through every provider.

Measure	September 9–October 24 gateway coverage
Requests	16,101
Input tokens	1,193.6 million
Output tokens	12.936 million
Originally recorded cost	$68.24
Cost after the documented cache-price correction	$29.53

An earlier price book charged some hosted cached input at the uncached rate. The corrected estimate adjusts affected QwenCloud pay-as-you-go requests before September 27, 11:03:40 UTC; later records pass through unchanged. The $29.53 is not a newly reconciled invoice for all providers, and it does not add subscription charges or local operating costs.

This correction matters: an inflated estimate of hosted spend can make a hardware purchase look easier to justify than it actually is.

The source splits requests between the main conversation and scheduled jobs using time and token-count matching. Some attribution remains uncertain, including thousands of unattributed scheduled requests. Its conversation, daily, and weekly tables overlap; adding them would count the same usage again.

Cache reporting has another limitation: the generator turns missing cache counts into zero. A displayed zero-percent cache rate therefore cannot establish that a model server reused no cached context. Cache comparisons need that distinction repaired first.

FreeFiFoFum: experiments and recurring analysis

A saved analysis of the midcap_rotation_v6 backtest reports:

Measure	One historical experiment
Simulated rebalance dates	27
LLM calls	2,406
Input tokens	50,645,179
Output tokens	766,996
Total tokens	51,412,175

About 98.5% of those tokens were input: financial filings, earnings material, news, and evidence passed through specialist and judge calls. The simulated market window is not the date when inference ran.

The original analysis includes hosted-price scenarios and a weekly-rebalance projection. Those are historical estimates, not actual hosted expenditure or a measurement of today’s nightly committee. Current recurring committee consumption remains a separate accounting gap.

RSP: paid frontier-model activity

RSP adds agent sessions and external model consultations within the FreeFiFoFum project. Saved records provide a partial cost picture:

Record set	Reported model cost
46 session results, September 28–October 23	$31.21
Four Opus consultant results, October 2–18	$2.98
Four GPT consultant results on those consultation dates	Unknown

The session set includes a rerun; its recorded session IDs are distinct. Consultant summaries duplicate their transcript costs, so only one representation is counted. Some model records explicitly identify list-price cost estimates. These figures are reported model costs, not a complete paid RSP total or an invoice reconciliation.

Market-data subscriptions are another project expense, separate from model inference. They are outside this token accounting.

How this reference will change

Provider exports can improve the coverage of hosted charges. An OpenRouter export covers OpenRouter activity, not Jane’s separate pay-as-you-go or token-plan billing, or her GitHub Copilot sessions. Each source needs its own billing category and coverage window before records can be combined. Gateway logs supply routing context, while direct-provider exports supply another accounting view of overlapping requests.

The next useful improvements are aligned reporting windows, current committee totals, missing consultant charges, and explicit handling of unknown cache usage. As those become available, updates should retain the earlier scope and explain corrections rather than silently replace historical numbers.

The sources are aggregate usage reports, a historical backtest analysis, and saved session/consultation accounting fields. Private conversations, financial decision details, and raw agent captures are not needed to explain these totals.

Revision history

October 25, 2026 — Initial publication. Added the UTC-aligned hosted day, replaced the unsupported $7–8/day expectation with an estimate-to-charge reconciliation, and distinguished repricing from forecasting another model’s behavior.