Evening AI stack watch: models, pricing, and autonomy

Apple shipped SpeechAnalyzer with no accuracy numbers. We measured it against the API it replaces and against Whisper, on 5,559 standard test utterances, and released every transcript.

Apple's new SpeechAnalyzer is the most accurate on-device speech engine we tested. It beat every Whisper model we ship, including Whisper Small, on both the clean and the noisy half of LibriSpeech, while running roughly three times faster than Small. And the API it replaces, SFSpeechRecognizer, came last on clean speech: behind even Whisper Tiny, a 40MB model.

Lower is better: WER is word error rate, the percentage of words an engine substitutes, drops, or invents. LibriSpeech test-clean is 2,620 utterances of clean read speech; test-other is 2,939 harder, noisier utterances. Every engine ran fully on-device on an Apple M2 Pro (32GB, macOS 26.5.1).

With iOS 26 and macOS 26, Apple replaced SFSpeechRecognizer with a new API, SpeechAnalyzer and SpeechTranscriber. It published no accuracy figures for either one. So every developer deciding whether to migrate, and everyone comparing Apple's built-in recognition against Whisper, has been guessing.

We ship both Apple engines and three Whisper models side by side in Inscribe , a private on-device AI workspace, which puts us in an unusual position: we can run all five through identical production code paths on the same machine and the same audio. So we did.

blog links /now reviews about Building and Shipping Mac and iOS Apps Without Ever Opening Xcode Saturday, 11 Jul 2026 Lately, I’ve heard several Apple related podcasters talk about how bad Xcode is, and how Apple needs to make vibe-coding Mac and iOS apps better by making Xcode less inscrutable. They’re not wrong, but also I don’t understand why they’re even opening Xcode in the first place.

And if you’re ever in doubt about how to make any of the following work, point Claude Code or your LLM coding tool of choice to this blog post, and let it figure it out. That’s literally its job, figuring out things you don’t want to have to.

The one-time setup is the only part with any friction, so let’s get it out of the way first.

You do have to have Xcode installed, there’s no getting around that, because build depends on tools that live inside Xcode.app.

Once Xcode is installed, make sure it’s the selected command line toolchain, and not /Library/Developer/CommandLineTools . If the output of the check is /Applications/Xcode.app/Contents/Developer , you’re in good shape:

Anthropic has started localizing Claude’s pricing in India, its biggest market outside the U.S., as global AI companies increasingly tailor their offerings to win users in the world’s most populous nation.

Local pricing has begun to appear for some users in India on Claude’s website and mobile apps. However, Anthropic has yet to enable payments via the Unified Payments Interface (UPI), India’s widely used instant payments network. Users still need to pay by card or through Apple’s and Google’s app store billing systems. This is unlike OpenAI, which rolled out Indian rupee pricing for ChatGPT in August with UPI support.

Claude users in India have long sought rupee-denominated subscriptions, with dollar pricing and currency conversion adding friction to accessing the service. The move is particularly significant, as India accounts for 5.8% of global Claude usage , making it the service’s second-largest market after the U.S., according to Anthropic.

On Claude’s website in India, Anthropic is listing Claude Pro at ₹2,000 (about $21) a month when billed annually, compared with $17 a month in the U.S. Claude Max starts at ₹11,999 (around $125) a month in India, versus $100 in the U.S., while Team plans start at ₹2,399 (around $25) per seat a month, compared with $20 in the U.S. The India prices include local taxes.

The Indian rupee pricing comes amid Anthropic’s growing focus on India. The Claude maker opened an office in Bengaluru in February, after announcing the move in October, and in January appointed former Microsoft India managing director Irina Ghose to lead its business in the country.

Techmeme summarizes Valarian’s funding announcement from Fortune.

Valarian raised a $50M Series A with NEA leading the round.

The startup is positioned around running AI workloads in U.S. clouds while keeping customer data control intact.

The announcement is primarily notable as proof of demand for data-sovereignty patterns in AI operations.

The linked Q&A centers on Nvidia’s automotive strategy under Xinzhou Wu.

It discusses Nvidia's chips alongside AI model choices for self-driving systems.

Lidar’s usefulness for Level 4 autonomy is treated as a concrete design choice, not a broad marketing claim.

The piece also touches deployment questions beyond chips, including market-specific rollout concerns.