Local vs Claude Haiku 5.5: Real Cost Per Million

Local vs Claude Haiku 5.5 cost per million tokens looks like a three-cent local win until you notice that each number counts something different. The 47-cent local figure is electricity for one model on one card at short context; the 50-cent Anthropic figure is the output price for a hosted model that runs about five times faster and scores higher on a published composite index. Neither number is wrong; neither one is the whole bill.

What the 47-Cent Local Figure Actually Counts

The 47-cent local figure counts electricity and nothing else. It comes from dividing one million decoded tokens by 46.2 tokens per second, multiplying by a 450-watt draw and applying a residential rate. That arithmetic is sound, and it leaves out the card, the wait and the model quality.

The decoding number itself is a third-party result. A llama.cpp benchmark published in August 2026 by Context Studios measured a 4-bit build of Qwen3 27B decoding at 46.2 tokens per second on an RTX 4090, with the weights occupying 17.1 GB. Decoding means generating one token at a time, and it is the step a user waits on.

One million tokens at 46.2 per second divides out to roughly 21,600 seconds, or six hours of the card running flat out. Six hours at 450 watts is 2.7 kWh. The United States Energy Information Administration reports 18.31 cents per kWh as the national residential average for July 2026, which puts the electricity cost near 49 cents. Working backwards, a 47-cent claim implies a rate close to 17.4 cents, close to last July's figure.

Both inputs are soft. The 46.2 tokens per second is a third-party measurement, and 450 watts is the rated power of the card, not a wall reading taken while this model ran. Memory bandwidth matters more than raw arithmetic when generating tokens, so 450 watts is probably a ceiling rather than a typical draw.

What Claude Haiku 5.5 Charges For

The 50-cent hosted figure is the output price alone, not the cost of running a request. Anthropic Claude Haiku 5.5 documentation lists input at 10 cents per million tokens, cached reads at 1 cent, and half off both directions through the batch API, where results arrive later in exchange for the discount.

Haiku 5.5 also steps up its rate for long prompts. Past 100,000 input tokens, input rises to 50 cents and output to $2.50 per million, a five-fold increase on output. The model's context window reaches one million tokens, so the long-prompt tier is reachable in normal use.

Anthropic states that Haiku 5.5 costs about 75 percent less to run than Haiku 4.5, and its documentation notes that the new tokenizer counts about 30 percent more tokens for the same text. A 75 percent price cut measured in tokens that inflate by 30 percent is a smaller cut in practice.

The hosted side is billed by the token, which means idle hours cost nothing. Local is billed by the hour, because the card draws the same power whether it is generating or sitting idle. That difference decides the comparison more than the sticker prices do.

The first time the transcript named the local model, automatic speech recognition produced a garbled form. The canonical model is Qwen3 27B, published by Alibaba's Qwen team. The canonical product names are Anthropic Claude Haiku 5.5.

Why Prompt Prefill Is the One Local Win

Prefill, the step that reads a prompt in one pass, is roughly 64 times faster than decoding on the same card. The same benchmark measured about 2,963 tokens per second of prefill on the RTX 4090, so one million input tokens takes roughly 340 seconds and consumes about 0.044 kWh.

Local electricity for reading a million tokens lands under one cent. Anthropic charges 10 cents per million input tokens on Haiku 5.5. On electricity alone, local wins the input line by more than ten times, which favors prompt-heavy work for anyone who already owns the card.

That advantage shrinks once hardware is included. Spreading a $2,776 card over three years and running it two hours a day puts hardware at about 12 cents per million prompt tokens, already above the hosted input price before electricity is added.

Prefill is fast because the whole prompt is processed together. Decoding cannot batch that way on a single stream, since each token depends on the one before it.

Hardware Amortization Turns 47 Cents into $7.60

Electricity is the smaller local cost once the card is counted. The RTX 4090 launched at $1,599, and GPU price trackers list a September 2026 street average of $2,776 and October listings from $3,217, so the card costs more now than at launch.

Spread $2,776 over three years and the card costs about $2.53 a day. Running flat out at 46.2 tokens per second produces roughly 4 million tokens a day, which puts hardware near 63 cents per million on top of about 49 cents of power.

Two hours of generation a day produces about 330,000 tokens. Dividing the same daily card cost by that output gives close to $7.60 per million for hardware alone, near $8 with power, against Haiku 5.5's 50 cents.

The figures below use the September 2026 street price of $2,776, a 450-watt draw and the July 2026 national residential rate of 18.31 cents per kWh. State rates range from about 12 cents to 46 cents, which moves single-stream power from roughly 32 cents to $1.24 per million tokens.

Claude Haiku 5.5 vs Local: Quality, Speed and Token Counts

Published composite scores put the hosted model ahead on quality and about five times ahead on speed. Artificial Analysis rates Qwen3 27B at 34 on its Intelligence Index and Claude Haiku 5.5 at 43 at maximum effort, 41 at extra-high and 38 at high, listed second of 180 models on its page.

Speed favors hosted by a wide margin. Artificial Analysis lists Haiku 5.5 at about 240 output tokens per second against 46.2 on the RTX 4090, so one million tokens takes roughly 70 minutes hosted against six hours local.

Per-token prices also miss how much each model writes. In the same evaluation, Haiku 5.5 at maximum effort produced about 440 million output tokens against a median near 100 million, while the open model needed about 200 million. A model that writes more than twice as much burns through its cheap token price faster.

The comparison table below keeps each figure in the configuration its source measured. Scores are composite and workload-specific, so they describe the index rather than any single task.

Long Context Changes the Local Math

Decoding slows as context grows, which raises electricity per token. In the same benchmark, the RTX 4090 drops from 46.2 tokens per second at 4,000 tokens of context to 38.4 at 64,000. A million tokens then stretches from six hours to a little over seven.

Run that through the same 450 watts and 18.31 cents per kWh and electricity alone reaches about 60 cents per million, now above Haiku 5.5's 50-cent output price. On the RTX 5090, the same benchmark shows 74.8 tokens per second at 4,000 tokens of context but only 26.2 at 64,000, or about 10.5 hours and 87 cents per million.

The benchmark does not explain why the RTX 5090 falls so far at long context, and 450 watts is an assumption for that card too. Treat 87 cents as arithmetic on the same assumptions with no meter behind it.

Hosted prices also rise at long context, but the tiers do not line up. Haiku 5.5 changes its rate past 100,000 input tokens, while the local benchmark stops at 64,000, so the two long-prompt tiers cannot be compared directly.

Batching Is Where Local Can Win

Serving many requests at once shares the memory read that decoding waits on, and that is the largest correction in local's favor. Context Studios reports an RTX 3090 power-limited to 250 watts serving about 1,000 tokens per second in aggregate across 64 parallel requests, against 40 tokens per second for a single stream.

At 1,000 tokens per second, one million tokens takes about 17 minutes and consumes about 0.125 kWh at 450 watts, roughly 2.3 cents of power, about 20 times less than the single stream's 49 cents. That report comes from a community-tuned guide without a control test, and it sets a 3090 batch result against a 4090 single stream, so it estimates the size of the effect rather than proving a number for your machine.

Spreading a $2,776 card over three years while it serves 1,000 tokens per second all day drops the hardware share to about 3 cents per million by the same arithmetic. Using a 4090 price against a 3090 result is a stretch, and keeping a card busy all day is the hard part.

One person at one keyboard has a smaller lever. Speculative decoding, which guesses several tokens ahead and verifies them together, carries a quoted 1.8-times speedup for a single user. Applied to 46.2 tokens per second, that would put power near 27 cents per million. The 20-times figure belongs to batch work such as summarizing a stack of documents, not to a single chat.

Batch work is where the two pricing models meet on equal terms. Anthropic batch API halves the output price to 25 cents per million tokens, and local batch arithmetic runs on the order of 2 cents of power plus about 3 cents of hardware, according to the community report. Batch jobs also fill a card's idle hours, which is the resource local pays for whether or not it is used.

Choosing Between Local and Claude Haiku 5.5

For one person chatting or coding without a GPU already in hand, Claude Haiku 5.5 is the cheaper pick by this arithmetic. The card alone works out to about 12 cents per million prompt tokens at two hours a day, above the hosted input price, and the hosted model runs about five times faster and scores 43 against 34 on the composite index.

If you already own an RTX 4090-class card, the purchase becomes a sunk cost and the sums fall back to electricity, roughly 50 cents per million at short context and 60 cents or more at long context. On money that is close to a tie, so the decision turns on privacy, offline use and tolerance for slower output.

If your work is batch, with many requests at once, a different arithmetic applies. A score of 34 on the index may be sufficient, and a full card changes the cost per million by roughly an order of magnitude. That scenario is the one where local has a defensible cost advantage, provided the card stays busy on your own machine.

The rest of the PC and idle draw sit outside this math and push the local total in both directions. A fair comparison keeps the same workload, the same context length and the same duty cycle on both sides.

Price and Capability by Scenario

Each row below uses figures from the configuration its source measured, so the comparison stays inside what was actually observed. Local hardware costs assume the September 2026 RTX 4090 street average of $2,776 spread over three years.

Reading the table: local wins on input at short context, hosted wins on output at short context once the card is counted, and local wins on batch throughput if the batching report holds for your hardware.

Scenario	Local, RTX 4090	Claude Haiku 5.5	Source of the figure
Output, single stream	49 cents electricity, 63 cents hardware at 24/7	50 cents per million output	Context Studios 2026; Anthropic pricing
Output, 2 hours a day	~$7.60 hardware plus power	50 cents per million output	RTX 4090 street price, September 2026
Input	Under 1 cent electricity	10 cents per million input	Context Studios 2026; Anthropic pricing
Batch output	~2 cents power plus ~3 cents hardware	25 cents per million output	Context Studios' 64-request report; Anthropic batch tier
Speed	46.2 tokens per second	~240 tokens per second	Artificial Analysis 2026
Composite quality	34 on the Intelligence Index	43 at maximum effort	Artificial Analysis 2026

To check your own setup, divide one million by your tokens per second to get seconds, divide by 3,600 for hours, multiply by your wall draw in kilowatts and by your rate per kilowatt hour. That gives electricity per million generated tokens. Divide the card's price by the millions of tokens it will actually produce in its life and add the two.

FAQ

• Is local AI cheaper than Claude Haiku 5.5? Not on a single stream, once the card is counted. Local electricity runs about 49 cents per million generated tokens at short context, and Haiku 5.5 output lists at 50 cents, but hardware amortization at two hours a day pushes local near $7.60 per million. Local wins on input, and on batch work where one card serves many requests.

• How much electricity does a local LLM use per million tokens? A 450-watt PC generating at 46.2 tokens per second takes about six hours and 2.7 kWh per million tokens, near 49 cents at the July 2026 national residential average of 18.31 cents per kWh. At 64,000 tokens of context the same card slows to 38.4 tokens per second, which raises power to about 60 cents.

• Why does the RTX 5090 cost more per million tokens than the RTX 4090? The benchmark shows the RTX 5090 decoding at 74.8 tokens per second at 4,000 tokens of context but only 26.2 at 64,000, against 46.2 and 38.4 for the RTX 4090. The slower long-context rate stretches a million tokens to about 10.5 hours, or 87 cents of electricity on the same 450-watt assumption.

• What does the 50-cent Claude Haiku 5.5 price include? Only output tokens. Input lists at 10 cents per million, cached reads at 1 cent, and the batch API halves both, putting batch output at 25 cents. Past 100,000 input tokens, input rises to 50 cents and output to $2.50 per million.

• Does batching make local inference cheaper than the hosted API? A community report measured an RTX 3090 at 250 watts serving about 1,000 aggregate tokens per second across 64 parallel requests, against 40 for a single stream. That implies roughly 2.3 cents of power per million tokens, an order of magnitude below single-stream local, though the report has no control test and uses different hardware.

• How does Claude Haiku 5.5 quality compare with open models that fit one consumer GPU? Artificial Analysis rates Qwen3 27B at 34 on its Intelligence Index and Claude Haiku 5.5 at 43 at maximum effort, 41 at extra-high and 38 at high. Larger open models score 44 to 51 in various listings but need more than one consumer graphics card.

• Does the Haiku 5.5 tokenizer change the effective price? Anthropic documents that the new tokenizer counts about 30 percent more tokens for the same text. The company also states the model costs about 75 percent less to run than Haiku 4.5, so the effective saving depends on which token count is used.

• How long does a million tokens take locally versus hosted? Local decoding at 46.2 tokens per second takes about six hours for a million tokens, stretching past seven hours at 64,000 tokens of context. Haiku 5.5 at about 240 tokens per second takes roughly 70 minutes, and the result arrives without a machine running in the room.

• What should I measure before deciding to run a model locally? Measure your wall draw under load rather than using the card's rated power, since a 17 percent power cut cost under 1 percent of speed in one measured test. Then divide a million by your own tokens per second and multiply by your actual electricity rate.

Turning This Arithmetic Into an Article

The lesson here is not that local AI is expensive or that hosted AI is cheap. It is that a headline number only means something when you know which cost it counts, and most published comparisons quietly leave the card, the speed and the token volume outside the sum.

If you have worked through that kind of reasoning on a video, the explanation usually stays trapped in the recording, where only the people who watch it ever see it. Skalablog turns a YouTube URL into a transcribed draft you can edit and publish, so the reasoning behind a number survives as text a reader can search and check. Paste the link at Skala Blog and the cost breakdown you recorded becomes an article someone can read before buying a card.

Source video