Hybrid LLM Routing Cost Calculator
You keep a frontier model for planning and review, and send the rest of the work to a local box or a cheap API. Enter the month's tokens, the share you route, and how often the cheaper model fails. The calculator prices the month three ways: all on the frontier, routed to a cheap API, and routed to your own box.
How it works
all-frontier = API bill for every token (uncached input, cached input, output)
frontier part = bill for the work it keeps
+ bill for the routed work x fail share (done again)
+ routed output x review share x uncached input price
cheap API part = API bill for the routed tokens at the executor's prices
box part = (price paid - sale price x (1 - selling fee)) / months
+ power for the hours the routed work takes
hybrid bill = frontier part + cheap API part, or frontier part + box part
break-even fail share = the fail share at which hybrid bill = all-frontier - Enter the month as if it all ran on the frontier. Take input and output tokens from your API dashboard or agent log. The routed share is the share of that work, input and output together, that the cheaper model does.
- Failed work is paid for twice. The executor runs it, then the frontier runs it again from the start. A fail share is yours to measure: log each escalation for two weeks. No source gives one number for it.
- Review reads are uncached input. When the frontier checks the executor's work, it reads that output as new text. New text misses the cache, and Anthropic bills a cache miss at the 5-minute write price while caching is on (1.25x input).
- The frontier's cache can expire while the box works. Anthropic's default cache lifetime is 5 minutes, refreshed on each hit (platform.claude.com pricing page, read 2026-11-20). If a local step runs longer than that, the next frontier call re-reads its context at the write price. Lower the frontier cache hit share to see the effect.
- Box cost: the same math as the cost per million tokens calculator. Two resale cases (sells for what you paid, sells for 40% less), eBay's 13.6% fee, and power for the hours of work. "Already own it" counts power only.
- Speeds are yours to enter. The defaults (41 decode, 800 prefill) are the figures from a CloudYeti DGX Spark video and a viewer comment, not a measurement of your box. Read
prompt eval and eval tokens per second from your own llama.cpp or Ollama timing lines, or estimate decode with the tokens per second calculator. At the default month the box works about 1 hour a day, so speed changes the answer only when the box is close to full. - API prices, dollars per million tokens, read on each vendor's page on 2026-11-20: Qwen3.6 27B (OpenRouter, cheapest host) $0.25 input, $0.025 cached, $2.00 output; DeepSeek V4.1 Flash (off-peak) $0.15 input, $0.003 cached, $0.60 output; DeepSeek V4.1 Flash (peak) $0.30 input, $0.006 cached, $1.20 output; DeepSeek V4 Pro (peak) $1.32 input, $0.044 cached, $3.96 output; Claude Sonnet 5.5 $2.00 input, $0.10 cached, $10.00 output; Claude Opus 5.5 $4.00 input, $0.20 cached, $20.00 output; Claude Opus 4.8 $5.00 input, $0.50 cached, $25.00 output. Sources: platform.claude.com/docs/en/about-claude/pricing, api-docs.deepseek.com/quick_start/pricing, openrouter.ai endpoints API for qwen/qwen3.6-27b.
- Box prices and watts: NVIDIA DGX Spark 128GB (new) $6,950 (NVIDIA list price since October 2, 2026 (The Register, October 2, 2026)); ServeTheHome wall-meter review (October 14, 2025): idle 40-45 W, LLM inference 60-90 W, "just under 200W" with CPU and GPU loaded. NVIDIA: peak total system power 240 W. RTX 3090 24GB (used) $1,649 (resaleprices.com used average, updated November 17, 2026 (310 listings, +30.1% in 90 days)); NVIDIA graphics card power, 350 W. RTX 4090 24GB (used) $3,449 (resaleprices.com used average, updated November 17, 2026 (84 listings, +38.0% in 90 days)); NVIDIA total graphics power, 450 W. RTX 5090 32GB (used) $7,061 (resaleprices.com used average, updated November 17, 2026 (257 listings, +58.9% in 90 days)); NVIDIA total graphics power, 575 W. Mac mini M5 Pro 24GB (new) $1,699 (apple.com Mac mini M5 Pro "from $1699", read November 17, 2026); Apple support 103253: M5 Pro Mac mini maximum 145 W at the wall (idle 6 W). Mac Studio M5 Ultra 96GB (new) $5,499 (Apple Newsroom, September 22, 2026: "Mac Studio with M5 Ultra starts at $5,499"); Apple support 102027: M5 Ultra Mac Studio maximum 385 W at the wall (idle 9 W).
- Not modelled: whether the cheaper model's accepted work is as good, your time checking it, the plan tokens the executor reads at handoff, rate limits, and subscription plans (a flat fee with usage limits).
Worked example: 200M input and 4M output tokens a month on Claude Opus 5.5
80% frontier cache hits, 10% of routed work fails, the frontier reviews all routed output. Cheap API: Qwen3.6 27B (OpenRouter, cheapest host). Box: RTX 3090 24GB (used) at $1,649, kept 24 months.
The 0% row is the all-frontier bill, $312.00. A bought box that does no work still costs its monthly hardware share. Routing 70% of the work cuts the bill by 54%, not 70%.
What the fail share does
At 70% routed, the hybrid month costs the same as the all-frontier month when 87% of the routed work fails (cheap API executor). The token bill forgives a weak executor for a long time. Your review time does not.
Viewer claims, run through the math
These are comments on CloudYeti local-AI videos. The figures in quotes are the viewers' own. The dollar figures are this calculator's, at the default month and the assumptions above.
- "Opus 4.8 (Planning), handover to Qwen3.6-35B-A3B ... On average I am saving 72% tokens on Opus 4.8." Tokens saved and dollars saved are different numbers. At 72% routed on Opus 4.8 with a 10% fail share and full review, 63% of the Opus tokens move and the bill falls from $430.00 to $185.20 with a cheap API executor (57%), or to $180.58 to $204.33 with a bought used RTX 3090. The viewer's workflow has Opus review twice, so its real review share is above 100%.
- "Qwen 3.6 27B ... will run on a 24GB GPU ... Use them for 90% and that last 10% just rent cloud." At 90% routed the $312.00 bill becomes $88.97 to $112.71 with a bought 24GB card, and $97.08 with the same model on a cheap host. The split holds only if the local model fails on about a tenth of the work, which is a number to measure.
- "I have two old mac minis and a 20 bucks sub" and "the total out-of-pocket cost for my entire home AI cluster was just $250." Hardware you already own has no monthly hardware share. Set "Already own it" to yes: at the default month the hybrid bill is $131.26, of which power is $1.82. An old box is slow, so check the hours-a-day line in the result.
- "Only then can a box like this pay for itself within 90 days." At the default month the saving pays for a $1,649 used RTX 3090 in 9.1 months, before resale. A 90-day payback at the same mix needs an all-frontier bill of about $949 a month. The viewer's condition matches the math: the workload has to exist first.
- "The $5K price point is definitely a tough decision ... Has anyone here found a way to make that investment pay for itself in real-world projects?" At the default month a $6,950 DGX Spark as the executor makes the hybrid bill $175.51 to $275.59, against $144.84 with a cheap API executor. The saving pays for the Spark in 40 months, before resale. On cost alone the cheap API wins at this volume; a larger box needs a larger bill behind it.
- "The best economic benefit of local AI is selective and not full replacement of frontier models. Find something you do repeatedly that is expensive to do in the frontier models." The table above is that sentence in dollars. Each extra tenth of the work you route takes about $23.88 off the default bill, and the work the frontier keeps is the floor. "There is no product that Apple sells today that is suitable for frontier level agentic coding" is the same point from the other side: in a hybrid split the box does not have to be.
- "What is your take on the 200 dollar subscriptions ... How does a hobbyist get ROI?" A subscription is a flat fee with usage limits: Claude Pro $20 a month, Max $100, Ultra $200 (claude.com/pricing, read 2026-11-20). Routing work off a flat plan saves no dollars. It buys room under the limit, so the money question is whether a local executor lets you stay on a smaller plan. A hobby does not have to return money.
- "DeepSeek is already cheap and at least as smart as any model that can run on a consumer-grade machine (under $10k)." On cost, the cheap API row agrees at low volume. At a tenth of the default month (14M routed input tokens), the cheap API route costs $14.48 and a bought box $22.47 to $46.22. The quality half of the claim is not something this page measures. "Even the cheapest model deepseek is doubling their token rate during peak hour" is also true: DeepSeek's off-peak rates are half its peak rates (api-docs.deepseek.com, read 2026-11-20).
Questions
How much does hybrid LLM routing save?
Less than the share of work you route. In the default example (200M input and 4M output tokens a month on Claude Opus 5.5, 80% cache hits), routing 70% of the work to a cheaper executor moves 62% of the frontier tokens and cuts the $312.00 bill to $144.84 with a cheap API executor, a 54% saving. The frontier still pays for the work it keeps, for the 10% of routed work that fails and is done again, and for reading the executor's output back.
Is LLM routing cheaper than using one frontier model for everything?
On the token bill, yes, as long as the cheaper model's fail share stays under the break-even. In the default example the break-even is 87%: the routed work would have to fail that often before the all-frontier month is cheaper. The bill does not include your time checking the cheaper model's work, and that is usually the larger cost.
Should I route to a local model or to a cheap API?
On cost alone they are close when the box is small and the volume is high, and the cheap API wins at low volume. At the default month a used RTX 3090 at $1,649 makes the hybrid bill $140.61 to $164.35, against $144.84 with Qwen3.6 27B on a cheap host. At one tenth of that volume the cheap API route costs $14.48 and the bought box $22.47 to $46.22. A box you already own costs only its power.
Does a local model pay for itself if I keep a frontier model for planning?
It can, when the all-frontier bill is large. At the default month the saving pays for a $1,649 used RTX 3090 in 9.1 months, before any resale. To pay it back in 90 days at the same mix, the all-frontier bill has to be about $949 a month.
Does a $20 or $200 AI subscription change the routing math?
Yes. A subscription is a flat fee with usage limits, so routing work away from it saves no dollars. It buys room under the limit. Claude Pro is $20 a month, Max $100 and Ultra $200 (claude.com/pricing, read 2026-11-20). This calculator prices API tokens. Use it when you pay per token, or when you would have to upgrade a plan to fit the work.
Related