Skip to main content

Savings

Anchor exists to cut inference spend — not by making models dumber, but by stopping you from paying frontier prices for keystrokes.

Frontier models are excellent judges and poor economics for bulk edits, boilerplate, renames, and “try again” thrash. The Playbook says ~80% of a typical build never needed the big model. Anchor makes that 80% run on mid-tier APIs or hardware you already own, with the same external discipline that keeps quality checkable.

If this project is already helping your bill, a quiet donation helps keep the work going — no pressure.

The model in one picture

About 80% of backlog work is execution (scoped edits, swarm steps); ~20% is judgment (plan, review, hard bugs). Anchor maps that onto a cheaper token mix and keeps quality checkable (tooling verifies; two fails → escalate once).

Baseline: ~100% of agent tokens at frontier-class pricing ($30 / 1M blended, illustrative).
Anchor mix: pie above (adjust to your stack). Calendar: 22 workdays / month · 12 months / year. Rates are order-of-magnitude, not a vendor quote.

Cost driverWithout AnchorWith Anchor
Routine editsFrontier / max planMid or local task specs
Retries on sloppy prompts3× spendTune cheap first (prompt_tuner)
Wrong tier on whole backlogExpensive agent grabs small workPreferred models + /work fit skip
Re-planning in chatRe-read repo on frontier.plans/; cheap executors pull
“Done” without checksSilent reworkTooling verifies

Saving something real? Please consider a small donation to support Anchor.


First year after adoption (Jan 1 = day zero)

Adoption day = 1 January. Usage volume is steady, but discipline is not instant: scaffold, endpoints, first plans, and fleet wiring take a quarter. Monthly savings are scaled by a ramp (fraction of work under the Anchor mix):

MonthRampWhat’s typical
Jan25%Scaffold, /config, first .plans/ / /draft
Feb50%Preferred models, mid/local endpoints, early /work
Mar75%Fleet timers / multi-tier habit
Apr–Dec100%Steady-state mix

That month’s spend on the Anchor path is
frontier_rate × (1 − ramp) + mix_rate × ramp
so early months still look partly “all frontier.” Cumulative savings = sum of ramp × full monthly gap.

Steady-state monthly rates (after ramp):

ScaleTokens / workdayFrontier / moAnchor mix / moFull gap / mo
Solo~8M$5,280$1,800$3,480
Team of 5~30M$19,800$6,730$13,100
Org of 20~100M$66,000$22,400$43,600

What changes at steady state: solo → mid + laptop/Mac Mini for execution; team → shared plans + fleet workers; org → H100/large mid for the 80%, frontier for architecture and final review.

Solo builder — cumulative spend vs savings

Color key

All-frontier spendNever adopt; full frontier bill every month
Anchor-path spendRamped mix (partly frontier in Q1)
Cumulative savingsGap between the two spend lines

By Dec (with ramp): ~$63k spent without Anchor vs ~$27k on the adoption path → ~$37k kept (vs ~$42k if you were at 100% mix from day one).

All scales — cumulative savings over 12 months

Same Q1 ramp, three org sizes:

Color key

Solo~$37k by Dec
Team of 5~$138k by Dec
Org of 20~$458k by Dec
ScaleSteady saved / dayEnd Q1 (Mar)End H1 (Jun)End year (Dec)
Solo~$158 at full ramp~$5.2k~$15.7k~$36.5k
Team of 5~$594 at full ramp~$19.7k~$59.0k~$138k
Org of 20~$1,980 at full ramp~$65.4k~$196k~$458k

Day ≈ full monthly gap ÷ 22 (after Apr). Scale linearly with your token volume and $/1M rates.

Extra lever: move bulk execution off mid API

Always-on mid execution (~40M tokens / workday on that slice) → owned hardware, same Q1 ramp, frontier judgment unchanged. Full gap ~$6,700 / mo after April:

Color key

Execution-slice savingsMid API → local only (not full org bill)

Hardware is capex or rent (Mac Mini → H100). Payback is often weeks to a few months once past ramp — run fit_device and your power bill.

If those curves look like money you get to keep, supporting this project is genuinely appreciated.


Solar: powering local compute

Once Anchor has moved a large slice of work off API meters, the residual cost of that slice is mostly hardware amort + electricity. Solar does not replace the orchestrator pattern — it attacks the power line of owned iron so always-on executors (Mac Minis, desktop towers, small GPU boxes, even denser fleets) run closer to pure amort.

When solar is in scope

More attractiveLess attractive
Always-on pullers (/fleet-watch, multi-box)Sporadic laptop inference a few hours/week
High local utilization (daytime batch + night batch with battery or night grid)Low duty cycle — panels sit idle with the GPUs
Expensive or rising grid $/kWh, peak demand chargesVery cheap industrial power already
Roof / ground / carport space, incentives, net meteringNo siting rights, heavy HOA/landlord constraints
Multi-year ops horizon (5–15 years)Need cash payback in <12 months

Rough power context (order-of-magnitude, duty-cycle dependent):

BoxContinuous-ish draw (illustrative)~kWh / year if ~50% duty
Mac Mini / small NUC~20–60 W~90–260 kWh
Gaming / 4090-class tower~150–450 W average under mixed load~650–2,000 kWh
Dense multi-GPU nodehundreds of W to multi-kWsite-specific — measure

Local token cost in the unit model (~$0.40 / 1M tokens) already bakes a little power + amort. Solar can shrink the power share of that number; it does not erase GPU purchase price.

Capex vs power savings (illustrative)

Not a quote. Regional panels, labor, incentives, and irradiance dominate. Treat this as a worksheet skeleton.

ItemIllustrative range
Small residential / shop array (3–6 kW DC)~$6k–$18k after common incentives (highly regional)
Usable annual production (good sun, that size)~4,000–9,000 kWh / year
Grid retail rate (example)~$0.12–$0.35 / kWh
Avoided grid spend at $0.20/kWh × 6,000 kWh~$1,200 / year
Simple payback on $12k net system~10 years (before rate inflation / demand charges)

For compute-heavy sites the avoided kWh is only the share the fleet actually draws — a single Mac Mini cannot “use” a whole 6 kW array; surplus usually needs net metering, other loads, or storage. Pairing solar with always-on fleet workers plus house/office baseload is usually more honest than “panels only for the GPU.”

Stacking with Anchor (same spirit as the Jan 1 charts):

  1. Year 0–1: software mix (API → mid/local) — largest $/mo gap; short payback.
  2. Year 1–N: hardware amort of local boxes.
  3. Year 3–15: solar (and optional battery) — slower payback, locks in low marginal kWh for the local slice and hedges rate spikes.

Color key

Grid-only power~$1,200 / year avoided-load example (no solar)
Solar path~$12k net system year 0, then ~$100 / year O&M stand-in

Crossing point ≈ simple payback (~year 10 in this sketch). Real systems often cross earlier with incentives, higher retail rates, or more kWh displaced — or later with poor sun / high install cost. Battery adds capex and resilience; model it separately.

Benefits vs costs (weigh before you buy)

BenefitsCosts / risks
Lower marginal cost per local token after paybackHigh up-front cash; multi-year horizon
Hedge against retail rate inflation and peak pricingInstall, permitting, interconnection delay
Aligns always-on fleet workers with daytime generation (or storage)Intermittency without battery; night jobs still on grid
Can serve whole site (not just GPUs) — better capacity factorRoof structure, shade, HOA, landlord, insurance
Narrative fit: own judgment and own electronsWrong-size array for tiny duty cycle → poor ROI
Optional ESG / resilience story for the shopInverters, O&M, eventual panel degradation

Practical order of operations

  1. Measure actual wall power for the boxes you will run 24/7 (kill-a-watt / PDU / UPS logs).
  2. Finish Anchor’s API → local move for fit work (hardware, fleet workers) so utilization is real.
  3. Size solar for site loads (fleet + office), not a fantasy GPU-only load factor.
  4. Run installer quotes + incentive calculators in your jurisdiction; re-do the cumulative chart with their kWh and net cost.
  5. Only then add battery if you need night-only local inference or backup — do not assume battery is free “for Anchor.”

Solar is a second-order savings lever: it matters most after local compute is already the right place for the 80%. It is not a substitute for Preferred models, /work fit, or prompt tuning.

Owning the stack end-to-end? A thank-you donation helps us keep the docs and tooling free.


Worked arithmetic (solo day, steady-state)

Full-ramp month (Apr–Dec). Early months use savings = ramp × $3,480 (Jan 25% … Mar 75%).

Tokens/day = 8,000,000
All frontier = 8 × $30 = $240 / day

Anchor mix (100% ramp):
20% frontier = 1.6 × $30 = $48
50% mid = 4.0 × $8 = $32
30% local = 2.4 × $0.40 ≈ $1
Total ≈ $81 / day

Savings/day ≈ $159
× 22 workdays ≈ $3,480 / month (steady)
Year with Q1 ramp ≈ $36.5k cumulative (not 12 × full month)

Swap in provider usage exports and real $/1M; the structure is the product, not the sample dollars.

If the arithmetic works in your favor, please help support Anchor if you can.

What Anchor does not claim

  • That local 8B models replace frontier judgment on greenfield architecture
  • That you will hit these exact dollars without measuring your own mix
  • That seat licenses (IDE, chat) disappear — this page is about inference routing and thrash
  • That solar is free or always ROI-positive — panels are multi-year infrastructure, not a software toggle

Measure with benchmark, provider dashboards, and a real power meter; treat this as a business case sketch, then validate.

Get the savings in practice

  1. Playbook — five moves (orchestrator pattern first)
  2. Doctrine — process that makes cheap models checkable
  3. /work + .plans/ — path-authoritative backlog, Preferred models
  4. /fleet-watch — durable multi-tier pullers
  5. Hardware — own the 80% when volume justifies it
  6. Solar (above) — optional; lower kWh for always-on local fleet after utilization is real

Anchor is free open source. If the savings on this page are more than theoretical for you, please consider a donation — it funds continued work on the doctrine, fleet tooling, and docs. Source on GitHub; project by Carefree Investments LLC.