Every AI price cut this week looks better on the slide than on your bill

Microsoft says its new streaming transcription model is 60% cheaper than ElevenLabs. Run the numbers from the published list prices and it comes out roughly 38% more expensive. That gap is the whole story of AI pricing right now.

Everyone is reading this as a price war that hands buyers an easy win. The short version: the war is real, but the headline numbers are a sales pitch. The number that matters is cost per finished task, at the setting you will actually run.

Start with Microsoft. MAI-Transcribe-2-Streaming costs $0.54 per hour of audio at its introductory rate. A thousand minutes is about 16.7 hours, so that works out to roughly $9. Artificial Analysis lists ElevenLabs Scribe v2 Realtime and Deepgram Flux at $6.50 per 1,000 minutes.

I'll be real: I do not know what baseline sits behind the 60% claim. It could be a different product or a different rate. But a percentage with no named baseline is a claim, not a price, and the public list points the other way.

To be clear, the model itself looks strong. It ranks first of 38 on the Artificial Analysis streaming benchmark with a 2.5% word error rate, against 2.73% for the runner-up. Pay for that if you need it.

Also know what you are paying for. The streaming version costs more than five times Microsoft's own $0.10 batch rate, because live results take more compute. If your audio can wait, batch is the deal.

Now the best deal of the week, which still needs an asterisk. Anthropic cut Claude Haiku 5.5 input pricing 90%, to $0.10 per million tokens for requests under 100,000 tokens. Anthropic itself says the average cut across all request lengths is about 75%, because its new tokenizer uses more tokens per request.

Same story on benchmarks. Haiku 5.5 scores 39.2% on Terminal-Bench 4.0 at maximum effort, and about 20% at the default medium setting. Max effort burns more tokens, so the best score and the cheapest bill do not live in the same request.

Here's the thing, though. Even with both asterisks, this is a massive upgrade. Haiku 4.5 scored 0.0% on that same test, and on OSWorld the jump runs from 15.7% to 72.4%.

It also matches GPT-6 Luna's list price while beating Luna on every benchmark in Anthropic's launch table. That table is Anthropic's own, so treat it as a claim to test. A real 75% cut on a far smarter model is still a gift.

Funny enough, the sharpest cost move came from a release with no price at all. Amazon put out Strands Decider 2B free under Apache-2.0. It does not write text. It picks from a list of options and returns a confidence score, with a median latency of 106 milliseconds on an RTX 3090.

The job is simple. Before your agent fires an expensive generative call, a tiny model on your own hardware decides whether that call should happen at all. AWS Distinguished Engineer Marc Brooker framed it as "Given where I am right now, what should I do next?" That question does not need a frontier model.

A 90% discount on a call you did not need is still 100% waste. On unseen short classification tasks, Decider answers scored at 0.9 confidence or higher were right about 95% of the time, and AWS says to confirm or escalate anything below that line. One warning: the bundled server ships with no authentication, so add your own before production.

So who wins? Buyers who measure. Who pays? Teams that budget off launch posts, then run everything at max effort through the biggest model they can find.

My rule of thumb: if a price claim has no named baseline, or a benchmark has no named effort level, assume your real number is worse until your own invoice proves otherwise. Rerun your actual workload and divide the bill by finished tasks. And if a step in your pipeline is a multiple-choice question, hand it to a small model.

See you next week, geeks.