BMW prices the electric 3 Series $4,400 under its petrol M350 in the US

BMW opened US orders on 30 September for its first electric 3 Series, the 2027 i3 50 xDrive at $61,500. That undercuts the petrol M350 xDrive by $4,400. Both figures exclude a $1,350 destination charge, according to The Next Web, which reported the pricing the same day.

TechCrunch put the gap at 6.7% in the electric car's favour. It is an odd shape for a launch. BMW is not asking a premium for the battery, at least not at the top of the range.

The comparison only works because BMW now sells the eighth-generation 3 Series with a choice of drivetrains under one body. The all-wheel-drive 330 xDrive costs $51,900. The four-cylinder petrol 330, at $49,900, is the cheapest way into the range, which leaves the electric i3 $11,600 dearer than the entry model. There is no electric equivalent of that car yet. All four launch models reach US showrooms in the first quarter of 2027. An i3 M60 xDrive follows later in 2027, with its price due nearer launch.

What the money buys

The i3 50 xDrive runs two electric motors for a combined 463hp and 476lb-ft of torque and reaches 60mph in 4.5 seconds, according to BMW. The M350 xDrive uses a 3.0-litre six-cylinder mild hybrid making 437hp and gets to 60mph in 3.9 seconds. The petrol car is quicker, the electric car is cheaper. Five years ago most buyers were not offered that trade.

Range is the other half of the pitch. BMW estimates 446 to 468 miles (718 to 753km) on preliminary tests using EPA procedures, and says a 400kW charger adds 191 to 208 miles in 10 minutes. The i3 is the first 3 Series on BMW's Neue Klasse architecture and comes from the Munich plant. The petrol models use an upgraded body and chassis and will be built in Dingolfing from November. BMW's maps can plan charging stops that favour IONNA's network, its preferred US fast-charging partner.

"As a fully electric model in a high-volume segment, the BMW i3 is a hugely significant car for the BMW Group," said Jochen Goller, the board member responsible for sales.

BMW has already sold two million electric cars. That is the context for why it can price this one aggressively and still call it a launch. The caveat is that not every EV has a like-for-like petrol model, which makes price comparisons hard, as TechCrunch noted. US EV prices rose in July for the first time this year as discounts shrank, so BMW is moving against a rising market.

The other side of the inference bill

The same week brought a smaller, cheaper-looking bill on the software side. Magnitude, a Y Combinator S25 company, launched an open source inference engine for agents on 30 September that compiles and tunes its kernels on the user's own device. The GitHub launch post claims open models run up to 2x faster than llama.cpp, with 92% faster decode on Metal and 19% on CUDA, and 27% less memory per agent. It works on Apple Silicon, NVIDIA, AMD, or just a CPU, and connects to Pi, OpenCode, Hermes, Codex and others. Apache 2.0, no token costs, nothing leaves the machine.

Those are vendor benchmarks, published by the company shipping the software, and should be read that way.

Cost planning is getting its own tooling. Flavio Copes published an inference cost calculator on 23 September that ranks 27 models from OpenAI, Anthropic, Google, xAI, Mistral, OpenRouter and Workers AI on the same assumptions: daily active users, calls per user, input and output tokens per call, and an optional prompt-cache hit rate. It runs entirely in the browser. Copes is explicit that the figures come from published API rates and ignore batch pricing, enterprise discounts, image or tool surcharges, and any caching layer of your own. The output is a planning estimate, not a vendor quote.

That distinction matters more than the calculator's interface. Inference cost is now the dominant line item in an AI app, and almost nobody can estimate it before wiring up billing.

Memory, and the cost of remembering

FastRecall launched on 28 September as what its makers call an OpenRouter for memory, aimed at routers and multi-agent systems. It stores context on disk and charges for storage rather than for LLM-based compaction, which it argues introduces latency, cost and retrieval errors. The published estimate starts at $7.00 a month for 100 contexts and 500 messages each, with 1M recalls included. It is designed to sit alongside existing provider caching rather than replace it.

The pitch is that context should move between models and providers without the user paying a model to summarise it every time. Whether that holds up depends on retrieval quality, which the page does not quantify.

Two other launches on 28 September point the same direction: people are rebuilding the parts of the stack they do not want to rent. Caspian, a desktop app for building and launching small projects, matches each change to a model and shows cost, tokens and time before anything runs, with a free tier capped at five published projects. A separate project, local_roborock_server, emulates Roborock's cloud on a local network so a vacuum keeps live maps and local controls without internet access, no rooting or hardware modification required. It is unsupported on the entry-level Q series, such as the Q7, because of differences in certificate validation, and it credits firmware research by Dennis Giese and Valetudo creator Sören Beye.

None of these are GPU pricing stories in the datacentre sense. They are the demand side of it. If a developer can run a model on a laptop, tune kernels for that laptop and keep context on a local disk, part of the inference bill simply does not get billed.

Where the two halves meet

BMW's pricing and Magnitude's launch are not connected, but they rhyme. Both are bets that the expensive part of the stack can be made cheap enough to stop being the deciding factor. BMW is absorbing the cost of a battery to make the electric 3 Series the better-value fast car, and pricing the petrol M350 at $65,900 against the i3's $61,500 says as much. Magnitude is absorbing the cost of kernel tuning on the device to avoid paying a provider per token.

The difference is evidence. BMW's numbers are list prices, published by the company and checkable against a dealer. Magnitude's 2x figure is a benchmark run by the people selling the engine, on hardware they chose. The calculator from Copes is the closest thing to a neutral referee in this dossier, and it says plainly that its own output is an estimate.

One date to watch: the i3 M60 xDrive arrives later in 2027, with pricing closer to launch. If BMW keeps the electric car under its petrol sibling across the range, the 6.7% gap stops being a launch stunt and becomes the shape of the line-up.