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 daysInput (uncached)Cache writeCache readOutputAPI-equiv $
Claude Fable 50.1M89.8M4,868M18.2M$7,475
Claude Opus 50.4M319.3M6,599M13.1M$5,799
Claude Opus 4.80.3M40.5M1,346M7.5M$1,200
Claude Sonnet 50.2M62.1M1,326M0.8M$441
Total1.0M511.6M14,138M39.6M$14,914
02,0004,0006,0008,000Claude Fable 5Claude Fable 5: $7,475/mo API-equivalent, 18.2M output tokens$7,475Claude Opus 5Claude Opus 5: $5,799/mo API-equivalent, 13.1M output tokens$5,799Claude Opus 4.8Claude Opus 4.8: $1,200/mo API-equivalent, 7.5M output tokens$1,200Claude Sonnet 5Claude Sonnet 5: $441/mo API-equivalent, 0.8M output tokens$441

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

ProviderFree tierTop consumer tierCoding-agent limits
Anthropic$0 — Fable/Opus/Sonnet/Haiku at ~50% of weekly limitsMax 20x — $200/mo (Pro $17–20, Max 5x $100)5-h rolling + weekly caps, pooled across surfaces; token amounts unpublished.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.
OpenAI$0 — GPT-5.6 Luna, basic CodexPro — $100 (5×) or $200 (20×) (Plus $20, Go $8)Only provider publishing numbers: Pro-20x = 200–2,000 GPT-5.6-Sol Codex msgs/5h. 1 credit ≈ $0.04 (inferred).
Google$0 — Gemini 3.6 Flash; Jules 15 tasks/dayAI 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 510 / 50gpt-5.6-sol4 / 20Gemini 3.1 Pro Prev ≤200k2 / 12
Opus 55 / 25gpt-5.6-terra2 / 12>200k ctx4 / 18
Sonnet 52 / 10gpt-5.6-luna0.20 / 1.20Gemini 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

OptionPrice (Aug 2026)Best current open model it can serveIdle / load W3-yr residual (base)
2× used RTX 3090 build (48GB)~$2,900Qwen3.8-27B 4-bit40 / 50050%
Strix Halo 128GB (GMKtec EVO-X2)$3,650Qwen3.8-Flash-Next13 / 11240%
NVIDIA DGX Spark 128GB$4,699 (+$700 Aug hike)Qwen3.8-Flash-Next35 / 17040%
RTX 5090 workstation (32GB)~$8,500 (card $4.3–4.9k street, 2.2× MSRP)Qwen3.8-27B 4-bit30 / 70070%
Mac Studio M5 Ultra 256GB$10,799GLM-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)

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:

0h350h700h1,050h1,400hRTX 5090 wkstn — Qwen3.8-27BRTX 5090 wkstn — Qwen3.8-27B: base 143 h/mo (range 113–234 h across 5–25% KV-miss)143 h2× used RTX 3090 — Qwen3.8-27B2× used RTX 3090 — Qwen3.8-27B: base 341 h/mo (range 243–636 h across 5–25% KV-miss)341 hMac Studio M5 Ultra 256GB — GLM-5.3-FlashMac Studio M5 Ultra 256GB — GLM-5.3-Flash: base 692 h/mo (range 496–1,282 h across 5–25% KV-miss)692 hDGX Spark 128GB — Qwen3.8-Flash-NextDGX Spark 128GB — Qwen3.8-Flash-Next: base 731 h/mo (range 553–1,267 h across 5–25% KV-miss)731 hStrix Halo 128GB — Qwen3.8-Flash-NextStrix Halo 128GB — Qwen3.8-Flash-Next: base 1,263 h/mo (range 906–2,334 h across 5–25% KV-miss)1,263 hMac M5 Ultra 512GB — DeepSeek V4 / Kimi K3Mac M5 Ultra 512GB — DeepSeek V4 / Kimi K3: base 3,817 h/mo (range 2,784–6,918 h across 5–25% KV-miss)3,817 h →720 h = running 24/7
workable (<240 h) saturated 240–720 h: permanent queueing, minutes-long TTFT impossible >720 h even running 24/7

Bars show the base case; tooltips carry the optimistic→pessimistic range. Two findings:

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):

$0$300$600$900$1,200$1,5002× used RTX 3090 (~$2.9k)2× used RTX 3090 (~$2.9k) — hardware amortisation: $40/mo2× used RTX 3090 (~$2.9k) — capital charge (5%/yr): $12/mo2× used RTX 3090 (~$2.9k) — electricity: $34/mo2× used RTX 3090 (~$2.9k) — labor (setup+maint+babysit @ $75/h): $1,025/mo$1,111● saturatedStrix Halo 128GB ($3.65k)Strix Halo 128GB ($3.65k) — hardware amortisation: $61/moStrix Halo 128GB ($3.65k) — capital charge (5%/yr): $15/moStrix Halo 128GB ($3.65k) — electricity: $15/moStrix Halo 128GB ($3.65k) — labor (setup+maint+babysit @ $75/h): $1,025/mo$1,116● impossibleDGX Spark 128GB ($4.7k)DGX Spark 128GB ($4.7k) — hardware amortisation: $78/moDGX Spark 128GB ($4.7k) — capital charge (5%/yr): $20/moDGX Spark 128GB ($4.7k) — electricity: $22/moDGX Spark 128GB ($4.7k) — labor (setup+maint+babysit @ $75/h): $1,025/mo$1,145● impossibleRTX 5090 workstation ($8.5k)RTX 5090 workstation ($8.5k) — hardware amortisation: $71/moRTX 5090 workstation ($8.5k) — capital charge (5%/yr): $35/moRTX 5090 workstation ($8.5k) — electricity: $22/moRTX 5090 workstation ($8.5k) — labor (setup+maint+babysit @ $75/h): $1,025/mo$1,153● workableMac M5 Ultra 256GB ($10.8k)Mac M5 Ultra 256GB ($10.8k) — hardware amortisation: $120/moMac M5 Ultra 256GB ($10.8k) — capital charge (5%/yr): $45/moMac M5 Ultra 256GB ($10.8k) — electricity: $28/moMac M5 Ultra 256GB ($10.8k) — labor (setup+maint+babysit @ $75/h): $1,025/mo$1,218● saturatedMac M5 Ultra 512GB (~$17.2k)Mac M5 Ultra 512GB (~$17.2k) — hardware amortisation: $191/moMac M5 Ultra 512GB (~$17.2k) — capital charge (5%/yr): $72/moMac M5 Ultra 512GB (~$17.2k) — electricity: $33/moMac M5 Ultra 512GB (~$17.2k) — labor (setup+maint+babysit @ $75/h): $1,025/mo$1,321● impossible$200 = Max 20x / ChatGPT Pro / AI Ultra
hardware amortisationcapital charge (5%/yr)electricitylabor (setup+maint+babysit @ $75/h)
Scenario ($/mo, base KV-miss)2×3090StrixSpark5090Mac 256Mac 512
Hobbyist — your time worth $0$86$91$120$128$193$296
Base — light babysit, $75/h$1,111$1,116$1,145$1,153$1,218$1,321
PG&E marginal 40¢/kWh$1,140$1,133$1,172$1,178$1,251$1,360
Heavy babysit (40 h/mo)$3,361$3,366$3,395$3,403$3,468$3,571
Senior time $150/h$2,136$2,141$2,170$2,178$2,243$2,346
Conservative residuals$1,127$1,131$1,165$1,200$1,263$1,392

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:

010203040506070Claude Opus 5 (your main driver)Claude Opus 5 (your main driver): AA Intelligence Index 63 (cloud-served, full precision)63Claude Fable 5Claude Fable 5: AA Intelligence Index 62 (cloud-served, full precision)62Kimi K3 — open, 2.8T paramsKimi K3 — open, 2.8T params: AA Intelligence Index 60 (cloud-served, full precision). ≫1TB: not desk-runnable60GLM-5.3 — open, 753BGLM-5.3 — open, 753B: AA Intelligence Index 60 (cloud-served, full precision). ~400GB+: not desk-runnable60GLM-5.3-Flash @ 256GB MacGLM-5.3-Flash @ 256GB Mac: AA Intelligence Index 57 (cloud-served, full precision); ≈53 effective after 4-bit local quantization57 · ≈53 local≈53 after 4-bit quantizationQwen3.8-Flash-Next @ 128GBQwen3.8-Flash-Next @ 128GB: AA Intelligence Index 56 (cloud-served, full precision); ≈52 effective after 4-bit local quantization56 · ≈52 local≈52 after 4-bit quantizationQwen3.8-27B @ 32GB GPUQwen3.8-27B @ 32GB GPU: AA Intelligence Index 52 (cloud-served, full precision); ≈48 effective after 4-bit local quantization52 · ≈48 local≈48 after 4-bit quantizationgpt-oss-120b (2025 gen)gpt-oss-120b (2025 gen): AA Intelligence Index 24 (cloud-served, full precision). superseded24
closed frontier (cloud only)open weightseffective score after 4-bit local quantization
Signal (independent where possible)FrontierBest local-servableGap
AA Intelligence IndexOpus 5: 63GLM-5.3-Flash @256GB: 57 → ≈53 quantized−10 pts (−16%)
SWE-bench VerifiedGPT-5.6 Sol 96.2% · Opus 5 96%DeepSeek V4-Pro 80.6% (vendor, unservable locally)−16 pts
Terminal-Bench 2.1 (agentic)GPT-5.6 Sol 89.5 · Opus 5 89.1GLM-5.3-Flash 84.3 (cloud-served)−5 pts
LMArena EloFable 5 ~1525 (#1)best open ~1450−55–75 Elo ≈ 54–58% win rate

The paper gap is the floor, not the ceiling. Three multipliers hit local deployments specifically:

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$/moQuality (AA)Serves your volume?
Claude Max 20x (current)$20063yes — proven by 3 months of logs
ChatGPT Pro 20x / Google AI Ultra 20x$200≈62–63likely yes (limits unpublished/lower)
RTX 5090 workstation, honest labor$1,153≈48yes (143 h/mo)
Mac M5 Ultra 256GB (buy or lease $210/mo)$1,218≈53saturated
API pay-as-you-go$14,91463yes
Free tiers (all providers)$062–63~2–5% of your volume

When local genuinely wins

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

Cross-checks

Key uncertainties (direction of bias)

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