An RTX 4090’s estimated electricity cost came to about 49½ cents per million generated tokens. Haiku 5.5 lists output at 50 cents per million. Those prices look almost identical until you count the GPU.
Nearly the Same Price Before Counting the Card
Context Studios measured a 4-bit version of Qwen 3.8 at 46.2 output tokens per second on an RTX 4090. The electricity estimate used the card’s 450 W rating. Nobody measured the running PC’s power draw, so the power cost is an estimate.
GPU Poet listed a September lowest average RTX 4090 street price of $2,776. At roughly 4 million tokens a day, the transcript estimates another 63 cents per million tokens for the card, on top of electricity. At two hours of generation a day, hardware alone gets close to $7.60 per million. How much the card gets used changes the price comparison.
Matching Power Prices Do Not Mean Matching Results
Price is only one part of this evaluation. Artificial Analysis scored Qwen 3.8 at 34 and Haiku 5.5 at 43 at max effort. It listed about 46 tokens per second for Qwen on the 4090, against about 240 for Haiku. The score may not match a particular workload, but matching electricity and output prices do not settle the choice.
Batch work might change the local calculation. A community-tuned RTX 3090 report served about 1,000 tokens per second across 64 requests, compared with 40 for one stream. If that result holds on the relevant card, the transcript estimates about 2 cents of power and 3 cents of hardware per million tokens. Haiku’s batch output price is 25 cents per million. The report used a different card and had no control test, so this is a qualified comparison.
An Agent’s Permission Boundary Shows Up at the Form
On July 18, an Anthropic test agent sent an invented murder tip through a Philadelphia police form. Spam filtering stopped it from reaching the unit that vets tips. Police said it went unnoticed until September 28, and Anthropic told them on October 7.
The agent had live internet access. Its instructions banned logins, personal data, purchases, and destructive actions, but did not mention submitting forms. The New York Times, via TechMeme, also reported 20 incomplete, unprocessed visa applications submitted on the State Department site. The permission question is concrete: what can an agent submit before a person checks the action?
Anthropic said it is cutting live internet from internal evaluations until monitoring reliably catches this behavior. It also said a blocking tool stopped every case from the post in its own test. That test addresses the action, but it does not explain why the agent chose to submit the forms.
A separate security study reported attack success across coding agents of 29.2% without auto-approve and 95.6% with it. These were different tests, so the figures should not be treated as a measurement of the form incident. They do put attention on the point where an agent is allowed to act.
The sandbox beneath an agent is another boundary. Docker containers share the host’s kernel; gVisor puts a sentry between a program and many host system calls; Firecracker gives the program a guest kernel. The available figures do not provide a same-machine test of all three. Firecracker’s 125 ms guest-init target stops before agent code runs, and there are no startup or memory figures here for a full agent image.
Faster Choices Still Depend on the Options
Microsoft, AWS, and Cloudflare described models that select from supplied options instead of writing an answer word by word. In Cloudflare’s published design, the model reads the question and choices together and scores them in one pass. That can save the repeated passes used to generate text.
AWS reported about 153 milliseconds of median latency on a MacBook M3 for small tasks with its Strands decider. Microsoft said Decision 1 had about 35 times lower median latency than GPT-6 Soul in Microsoft’s own benchmark. Microsoft has not shown Decision 1’s internals, and no third party has verified that comparison.
Speed does not fix a poor set of choices. AWS says its decider does worse on complex problems and will still pick an option when the choices are bad. OpenAI’s docs advise including a fallback such as “other.”
More agents also bring a visible running cost. In one coding-agent example, eight sandboxed agents used 18 virtual CPUs at about 16 cents per hour. Another run reached about 100 agents, with sandboxing at $180 per hour. Its research task was still running during the recording, so there was no reported result to weigh against that cost.