Anthropic has put a capable small cloud model on sale for 10 cents per million input tokens and 50 cents per million output tokens, provided the prompt stays at or below 100,000 tokens. The company says Claude Haiku 5.5 costs about 75% less per task than Haiku 4.5, and it is available now through Anthropic, AWS, Google Cloud and Microsoft Azure.[1]
For someone considering a home-AI machine, that changes the arithmetic. The cloud is no longer automatically the expensive option, even before counting the rest of the PC, the electricity or the evening lost to persuading a driver that it does, in fact, contain a GPU.
It does not make local AI pointless. It makes the honest case for local AI less about saving pennies on tokens and more about privacy, control, offline access and owning the model you depend on.
The cheap model is genuinely cheap
Haiku 5.5’s lower price applies to the short and medium prompts that Anthropic says made up about 90% of requests to the previous Haiku model. Above 100,000 prompt tokens, the rate rises to 50 cents per million input tokens and $2.50 per million output tokens. Anthropic also notes that the new tokenizer uses slightly more tokens per task, which is why a 90% cut in the short-prompt list price becomes a claimed 75% reduction in average task cost.[1]
The rest of the Claude range shows the same direction, though not the same bargain. Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens; Anthropic says improved token efficiency can make a task cost up to 30% less than on Sonnet 5.[2] Opus 5.5 is $4 in and $20 out, while Anthropic claims a 40% reduction in typical workload cost against Opus 5.[3]
Those percentage savings are company measurements, not HighVRAM tests. The capability is not mere bargain-bin positioning, however. Artificial Analysis ranked the max-effort versions of Opus 5.5 and Sonnet 5.5 as its two highest-intelligence models when checked on October 25.[4] Haiku sits lower in the range, aimed at repetitive, high-volume jobs such as summaries, classification, database queries and subagent work.[1]
That product ladder is the business strategy in miniature: use an expensive model where judgment matters, then push routine work to a cheaper one. The customer spends less per task, but can afford to run many more tasks. Anthropic gets a path from occasional chat to software that calls a model all day.
A deliberately unfair fight with an RTX 5090
A home GPU does not sell tokens, so comparing it with an API requires assumptions. Here is a transparent scenario rather than a benchmark.
Nvidia lists the 32GB RTX 5090 from $1,999.[5][6] Spread that price over three years and the GPU costs about $55.53 a month before electricity. Nvidia rates it at 575 watts, and the US Energy Information Administration’s preliminary August 2026 residential average was 18.16 cents per kilowatt-hour.[6][7]
Now assume a quantized 32-billion-parameter local model generates 50 tokens per second. HighVRAM did not measure that speed for this article; it is an illustrative workload. At that rate, one million output tokens takes about 5.56 hours. Running the GPU at its full 575-watt rating would add roughly 58 cents of electricity. The rest of the computer, prompt processing, software overhead and downtime are excluded.
For a workload that sends four input tokens for every output token, one million output tokens would cost about 90 cents on Haiku 5.5, $18 on Sonnet 5.5 and $36 on Opus 5.5 at list rates.[1][2][3] Caching can lower the API bill further.
Under those assumptions, the local GPU crosses Sonnet’s monthly cost at roughly 3.2 million output tokens and Opus at about 1.6 million. Haiku is the awkward result: its theoretical crossing point is about 174 million output tokens a month, more than the assumed local system could generate in a 30-day month at 50 tokens per second.
This is not an apples-to-apples victory for Haiku. A local 32B model is not Claude Haiku, and the article makes no claim that their quality is equal. A 32GB card can comfortably hold the raw 4-bit weights of a 32B model, about 16GB before cache and runtime overhead. A dense 70B model needs about 35GB for 4-bit weights alone, so it will not live entirely in 32GB. Local software such as Ollama, LM Studio or a llama.cpp-based runtime also needs setup and model-specific support; Nvidia itself lists several of those local-AI paths for RTX systems.[5]
The comparison is useful because it shows how low the API floor has fallen. Even an expensive GPU running hard does not automatically beat the cheapest useful hosted model on operating cost.
What the spreadsheet leaves out
Cloud AI asks for a different payment. Your prompts and documents are sent to a remote service. You depend on an account, network connection, pricing policy, usage limits and whatever model behavior the provider ships next. Anthropic’s announcement lists Haiku on hosted platforms and gives no option to download its weights for local use.[1]
A local model can work offline, keep sensitive material inside the house, accept modifications and continue running after a vendor changes direction. The hardware can also generate images, transcribe audio, train adapters, play games and run models from competing developers. Those uses make a GPU easier to justify than a single token comparison suggests.
The people who benefit most from Haiku 5.5 are light and medium users, small software projects, and agent builders with large volumes of narrow tasks. They can start for pocket change instead of buying a machine. Heavy users still have a case for local hardware when their chosen open model is good enough and their workload is steady. Privacy-sensitive users may decide the cost comparison is beside the point.
Cheap APIs make home AI more accessible because they let newcomers discover which workloads matter before buying hardware. They also raise the bar for the purchase. If the only argument for a $1,999 GPU is “the cloud bill will be worse,” Haiku 5.5 has made that argument difficult to defend. Buy local for the control, the capability and the work you know you will run. Rent first when you are still finding out.