Local models vs cloud frontier: a true-cost & efficacy analysis
Built on your measured Claude Code usage (last 30 days, 2026-07-29 → 08-28) and primary-source pricing
verified 2026-08-28. Amortized hardware, electricity, labor, residual value, sensitivity analysis and validation included.
Plan detected: Claude Max 20x, $200/mo (billed via Apple, extra-usage enabled).
$14,914
API-equivalent value consumed /mo (list prices)
$200
what you actually pay (Max 20x)
75×
subscription leverage
14.69B
tokens processed /mo · 96% from prompt cache
71,350
requests /mo · 31 active days
Demand side
Your workload, measured
Parsed from local Claude Code transcripts (2,317 JSONL files, deduplicated by message + request ID) — the same data behind /usage:
Last 30 days
Input (uncached)
Cache write
Cache read
Output
API-equiv $
Claude Fable 5
0.1M
89.8M
4,868M
18.2M
$7,475
Claude Opus 5
0.4M
319.3M
6,599M
13.1M
$5,799
Claude Opus 4.8
0.3M
40.5M
1,346M
7.5M
$1,200
Claude Sonnet 5
0.2M
62.1M
1,326M
0.8M
$441
Total
1.0M
511.6M
14,138M
39.6M
$14,914
All-time since 2026-05-06: 96,358 requests, 74 active days, 18.1B cache-read tokens. Peak day Aug 20: 2.70B tokens, 10,064 requests.
Priced at Anthropic list (Fable $10/$50, Opus $5/$25, Sonnet 5 $2/$10 per MTok; cache writes 1.25×/2×, reads 0.1× input).
The structural fact that decides everything downstream: 96% of your token volume is served from the provider-side prompt cache
at a 90% discount. A local rig has no equivalent free lunch at this scale — every cached token evicted from a local KV cache must be physically
recomputed (prefill), and your multi-agent pattern (up to 10k requests/day) churns sessions constantly.
Cloud supply side
What the clouds charge (verified 2026-08-28)
Consumer subscriptions
Provider
Free tier
Top consumer tier
Coding-agent limits
Anthropic
$0 — Fable/Opus/Sonnet/Haiku at ~50% of weekly limits
Max 20x — $200/mo (Pro $17–20, Max 5x $100)
5-h rolling + weekly caps, pooled across surfaces; token amounts unpublished. Limits doubled 2026-05-06.
AI Ultra — $99.99 (5×) or $199.99 (20×) (Pro $19.99)
Jules: 300 tasks/day, 60 concurrent on Ultra (same at both Ultra price points).
API list prices, $/MTok in → out (frontier + mid)
Anthropic
$
OpenAI
$
Google
$
Fable 5
10 / 50
gpt-5.6-sol
4 / 20
Gemini 3.1 Pro Prev ≤200k
2 / 12
Opus 5
5 / 25
gpt-5.6-terra
2 / 12
>200k ctx
4 / 18
Sonnet 5
2 / 10
gpt-5.6-luna
0.20 / 1.20
Gemini 3.7 Flash (promo → 2× on 2027-01-01)
0.75 / 3.75
Corrections to commonly circulated figures: Google AI Ultra is no longer $249.99 (split $99.99/$199.99 at I/O 2026);
ChatGPT Pro is now two price points; Sonnet 5's scheduled Sept-2026 rise to $3/$15 was cancelled — $2/$10 is permanent.
Claude 4.7+ tokenizer produces ~30% more tokens per unit text than Sonnet 4.6-era models (already reflected in your measured counts).
Free tiers vs your workload: Claude Free is ~50% of Pro weekly limits → roughly 2–3% of your Max-20x volume.
Gemini free Jules (15 vs 300 tasks/day) → ~5%. ChatGPT Free ("basic Codex") — unpublished but similar order. Free tiers are not a
serviceable option for this workload; they exist in this analysis as a $0 floor only.
Hardware supply side
Local hardware in the 2026 DRAM shortage
Option
Price (Aug 2026)
Best current open model it can serve
Idle / load W
3-yr residual (base)
2× used RTX 3090 build (48GB)
~$2,900
Qwen3.8-27B 4-bit
40 / 500
50%
Strix Halo 128GB (GMKtec EVO-X2)
$3,650
Qwen3.8-Flash-Next
13 / 112
40%
NVIDIA DGX Spark 128GB
$4,699 (+$700 Aug hike)
Qwen3.8-Flash-Next
35 / 170
40%
RTX 5090 workstation (32GB)
~$8,500 (card $4.3–4.9k street, 2.2× MSRP)
Qwen3.8-27B 4-bit
30 / 700
70%
Mac Studio M5 Ultra 256GB
$10,799
GLM-5.3-Flash 4-bit (200GB)
~10 / ~220†
60%
Mac Studio M5 Ultra 512GB
~$17,200 (est†, ships late Oct)
DeepSeek V4 / heavy-quant Kimi K3
~10 / ~250†
60%
† M5 hardware ships Sept 22 2026: power figures proxied from Apple's official M3 Ultra measurements (9W idle / 270W max);
512GB price extrapolated at Apple's observed $25/GB memory pricing — Apple says only "well above $10,000". Flagged, not quoted.
Apple's leasing program ("Apple Upgrade", launched 2026-07-28)
Klarna-underwritten, 0% interest, soft credit check. Mac Studio from $48.99/mo. Payments + buyout = exactly list price — it is deferred payment, not depreciation-priced leasing.
36-mo structure: pay ~70% of list over the term, ~30% residual buyout — or return the machine. M5 Ultra 256GB: ≈$210/mo, $3,240 buyout.
Return-at-term = a guaranteed 30% residual with zero resale friction. Buying + reselling at today's 60% used-market retention is cheaper ($120/mo effective) — if the shortage-inflated used market holds. Lease = insurance against residual collapse. Ceiling for five-figure BTO configs unconfirmed.
Shortage context: AI datacenters are absorbing ~70% of world memory output (IDC). RTX 5090 street is 2.1–2.5× MSRP;
RTX PRO 6000 is +87% since launch; DGX Spark +$700 in August; Apple pulled all >128GB Macs for 5 months and repriced upward.
Used RTX 4090s now sell above their original MSRP — GPU depreciation has inverted. Residuals in this model are therefore
shortage-propped: the base case uses today's observed values, the conservative case assumes partial normalization after 2027 supply relief.
Feasibility
The physics: can a local box even serve this workload?
Your workload requires prefilling ~513M genuinely new tokens/mo, plus recomputing whatever share of the 14.1B cached tokens
a local KV cache cannot hold (base case 10%, range 5–25%), plus decoding 39.6M output tokens. Dividing by each machine's
depth-adjusted, primary-source throughput (measured at agentic 30–130k contexts, not headline pp512 numbers) gives the
compute-hours each box needs per month — against 720 hours existing:
workable (<240 h)saturated 240–720 h: permanent queueing, minutes-long TTFTimpossible >720 h even running 24/7
Bars show the base case; tooltips carry the optimistic→pessimistic range. Two findings:
The only workable class is the small-model GPU class (RTX 5090 running a 27B). Every 128GB+ unified-memory box that could host a better model is saturated or physically impossible — prefill speed, not decode, is the wall.
The near-frontier open models are unservable at your volume. The Mac that fits DeepSeek V4-class models would need 3,800+ h/mo — 5× more hours than a month contains. Measured TTFT for a 128k prompt: RTX 5090 ~20s · DGX Spark ~2 min · Strix Halo 4–7 min · Mac Studio ~14 min. The cloud serves the same request from cache in seconds.
TCO
True monthly cost, amortized
3-year straight-line amortization net of residual value, 5%/yr opportunity cost of capital, US-average electricity (18.3¢/kWh),
and labor at $75/h — 24h one-time setup (measured range: a weekend to a week), 3 h/mo maintenance (chat-template patches, driver
regressions), 10 h/mo extra babysitting (controlled study: local agents needed 7 operator interventions per task vs 1 for Claude):
Labor dominates everything. Hardware amortization is $40–191/mo; electricity is $15–34/mo; the moment your time has market value,
the babysitting tax is 5–25× the hardware cost. This matches the strongest first-hand study's conclusion: the GPU is
"16–20 months of Claude for the price of the card alone — before the rest of the rig, the electricity, or your time."
Reference points: your $200/mo subscription (dashed line) · API pay-as-you-go for this workload: $14,914/mo — 10× above
the chart's top edge. Local inference IS dramatically cheaper than API list prices. It is dramatically more expensive than your subscription,
which is the option you actually hold.
Efficacy
Efficacy: what quality do you give up?
Independent cross-vendor scores (Artificial Analysis Intelligence Index v4.1.1, 9 evals, same harness for open and closed models).
Bars = cloud-served full precision; tick marks = effective level after 4-bit quantization (≈93% retention, Unsloth measurements)
needed to fit local memory:
closed frontier (cloud only)open weightseffective score after 4-bit local quantization
The paper gap is the floor, not the ceiling. Three multipliers hit local deployments specifically:
Quantization tax — 4-bit ≈93% retention, 3-bit 82%; and quantization specifically degrades tool-call formatting, the exact capability agentic coding depends on.
Reasoning-effort tax — Qwen3.8-27B hits its 52 only at xhigh effort, burning 160M output tokens vs a 43M median on the same eval: a benchmark win that becomes a wall-clock loss at 40–180 tok/s local decode.
Workflow tax — the one controlled head-to-head (Ring, Jun 2026): local agent passed 74 tests vs Claude's 114 (65%), needed 7 operator interventions vs 1, hit 4 context compactions vs 0, and looped on an identical error 44 consecutive times. Documented ecosystem breakage is systemic: llama.cpp broke OpenAI-compatible tool-calls (Mar 2026); a coding-agent + llama.cpp tool-role bug was closed "not planned"; broken GGUF chat templates are endemic.
Quality-adjusted (÷0.65 acceptance from the controlled study), the feasible local options cost $1,774–1,874/mo for
Claude-equivalent completed work — 9× your subscription — before counting the intervention time already in the labor line.
A genuinely important nuance: the open-model gap has narrowed sharply in 2026 (Kimi K3 and GLM-5.3 sit at 60 vs 63) —
but only at parameter scales (753B–2.8T) that no desk-side machine can serve at your volume. The gap you can buy at retail is
−10 to −15 index points, not −3.
Bottom line
Verdict
For your workload, nothing local comes within an order of magnitude of your current setup.
You consume ~$14,914/mo of API-equivalent frontier compute for $200. The best local alternative that is even
physically feasible at your volume costs ~$1,153/mo honestly accounted ($128/mo if your time is worth $0), delivers
≈48/63 of frontier quality, ~7× the operator interventions, and minutes-long prompt latencies the cloud serves in seconds.
Option
$/mo
Quality (AA)
Serves your volume?
Claude Max 20x (current)
$200
63
yes — proven by 3 months of logs
ChatGPT Pro 20x / Google AI Ultra 20x
$200
≈62–63
likely yes (limits unpublished/lower)
RTX 5090 workstation, honest labor
$1,153
≈48
yes (143 h/mo)
Mac M5 Ultra 256GB (buy or lease $210/mo)
$1,218
≈53
saturated
API pay-as-you-go
$14,914
63
yes
Free tiers (all providers)
$0
62–63
~2–5% of your volume
When local genuinely wins
Hard privacy / air-gap / compliance requirements — the only case where every credible source agrees, because the cloud is simply unavailable. Price it as +$950–1,650/mo vs your subscription.
Workloads locked out of consumer subscriptions (high-volume programmatic API use): local at ~$1.1–1.3k/mo beats $14.9k API list by 10×+ — this is the real arbitrage local serves, and it is not your situation.
Hobby scale with time valued at $0 and quality tolerance: $86–128/mo beats a $200 subscription in cash terms (at ~25% lower model quality and no frontier access).
A hedge you already own: running a 4-bit 27B on hardware you have anyway costs ~$15–30/mo of electricity — reasonable as an offline fallback, not a replacement.
If cost pressure is the motive: the far better lever is model-mix discipline — your Fable 5 usage ($7,475 API-equivalent, 50%
of consumption) burns weekly limits ~2× faster than Opus per token; shifting exploratory work down-tier stretches the same $200 further
than any hardware purchase ever could.
Validation
Method, sensitivity & validation
Cost-estimation practices applied
Measured demand, not assumed — 30-day token/request census from local transcripts, deduplicated; model-mix priced per token category (input / 5m & 1h cache-write / cache-read / output).
Primary-source prices with dates — every price re-verified 2026-08-28 against provider-owned pages; stale figures (AI Ultra $249.99, single-tier ChatGPT Pro, $2k RTX 5090) explicitly rejected; SEO/AI content farms excluded by name.
Full TCO — amortization net of residual, opportunity cost of capital, marginal (not average) electricity, one-time + recurring + per-task labor, lease-vs-buy for Apple hardware.
Sensitivity grid — six scenarios per option (above); conclusions survive every branch: even at $0 labor + optimistic KV-miss + US-cheap power, no option matches subscription quality-per-dollar.
Cross-checks
Two independent scripts computed usage totals — agree to rounding. Priced total re-computed after the Sonnet 5 price correction ($15,117 → $14,914).
Blended cost $1.02/MTok is consistent with the one published heavy-user datapoint (10B tokens ≈ $15k API-equiv → $1.5/MTok, lighter caching).
Feasibility model reproduces independently reported real-world TTFT (~90s for a 16k prompt on Mac Studio ≈ our 178 t/s effective prefill).
Amortization cross-checks the Ring study's framing ("GPU = 16–20 months of subscription for the card alone").
Your 75× leverage vs third-party "heavy user" estimates ($600–1,500/mo): you are 10× their "heavy" — explained by multi-agent workflows, 1h-TTL cache pricing, and Fable-tier rates; flagged rather than smoothed.
Key uncertainties (direction of bias)
M5 Ultra power & throughput unmeasured until Sept 22 — proxied from M3 Ultra (likely understates M5 prefill — but a feasibility class only flips if throughput is off by ~2×, and the 512GB case misses by 5×).
512GB price extrapolated ($25/GB observed); Apple Upgrade eligibility ceiling unconfirmed.
KV-miss fraction (5–25% band) is an engineering estimate — the single biggest feasibility lever, hence carried as the scenario axis.
Throughput for Qwen3.8-FN / GLM-5.3-Flash proxied from measured same-active-param predecessors (gpt-oss-120b etc.).
GPU residuals are shortage-propped; conservative branch provided. SWE-bench Pro / LCB open-model scores are vendor-reported.
Your possible extra-usage overage spend is billed server-side and invisible to local logs — if material, it strengthens the subscription-value conclusion's denominator but not its direction.
Sources: Anthropic/OpenAI/Google pricing & support pages; apple.com newsroom + Apple Upgrade terms; EIA Electric Power Monthly
& STEO Aug-2026; PG&E tariff sheet (eff. 2026-03-01); Artificial Analysis Index v4.1.1; Vals AI Terminal-Bench 2.1; llama.cpp/mlx/vLLM
GitHub benchmark discussions & issues (#15396, #16578, #17917, #19890, #20198, #3209); kyuz0 Strix toolboxes; CloudRift; jetsonhacks;
johnhringiv.com controlled study; Swappa/BuySellRam resale data; videocardz/overclock3d price reporting. Full URL list available on request.
Generated 2026-08-28 · demand data: your machine's Claude Code transcripts ·
model: 3-yr horizon, 36-mo amortization · all flagged estimates marked in place