This is Robin, bringing you Salty Talk's Daily AI Stew.
Today starts with what it costs to call a model, then looks at how data moves between apps, how a phone's system boundaries get broken, and one update to a communications protocol library. Put these items side by side and the focus lands in the same place: what happens once technology leaves the release notes — how you use it, how far it actually gets, and what it costs.
Kimi K3.1 Pro Lands on the API, with Speed and Price Both on the Table
Kimi K3.1 Pro is now live on the Kimi API. Per Simon Willison's write-up, the standard tier charges 2 US dollars per million input tokens and 10 dollars for output; the Highspeed tier is 2 to 3 times faster at double the price, though the price stays the same on an input cache hit. For developers wiring the model into a product, that makes the choice concrete: what lower latency costs, and which part of the bill a cache can take back. [Source: Simon Willison](https://simonwillison.net/2026/Oct/23/kimi-k31-pro)
Long-document handling is the other thing worth watching in this release. The AIHOT daily reports that Kimi K3.1 Pro is in the 3T size class, supports a 1 million token context, and is slated to have its weights opened by November 4. Two things need keeping apart here: the API is callable today, while the full weights are still a future delivery the company has promised. If the weights do open on schedule, long-document work and self-hosted deployment gain one more option. [Source: AIHOT Daily](https://aihot.news/daily/2026-10-24)
Beyond the scores, there is also an image that sticks. Simon ran the model at three reasoning levels — low, medium, and max — asking it to generate an SVG of a pelican riding a bicycle; the max-level sample is usable overall, but both of the pelican's legs ended up on the same side of the frame. The write-up lists an Artificial Analysis score of 55, while the vendor's own benchmarks claim it sits roughly level with Opus 5.5 and GPT-6 Astra. Leaderboards give you coordinates for comparison; that cycling pelican leaves behind one concrete flaw in the output. When you pick a model, price, speed, and what it actually generates all belong on the same table. [Source: Simon Willison](https://simonwillison.net/2026/Oct/23/kimi-k31-pro)
Google's Cross-App Intelligence Brings the Ad Data Boundary to the Front
When a model sets out to understand your personal context, which data does it reach for? Google's Personal Intelligence puts exactly that question in front of us. According to Ars Technica, Android Authority noticed that Google updated its privacy policy on October 16, revising the earlier line stating that it would not show you personalized ads based on your content in Drive, Gmail, or Photos. What this change touches is the written commitment that used to sit between private content and ad personalization. [Source: Ars Technica](https://arstechnica.com/gadgets/2026/10/updated-google-privacy-policy-says-your-gmail-might-be-used-for-advertising)
The technical connection point is Personal Intelligence. According to related reporting from The Verge, the feature entered beta testing in early 2026 and lets Gemini share information across services such as Gmail, Google Photos, Search, and Google Drive in order to improve recommendations. Once a user turns it on, the report says, that data may be used for broader personalization purposes, ad targeting included. Cross-app data can fill in context for an AI, and it also makes where the data goes next an unavoidable part of product design. The word "may" in the reporting sets the limits of this item: the material does not confirm that email and photos are already being used for ad serving across the board. [Source: The Verge](https://www.theverge.com/tech/1017600/google-personal-intelligence-privacy-policy-ads-update)
iPhone Security Research Shows Its Demos, but the Scope Still Needs Stating Precisely
Moving from data flows to the inside of the device, a research team says it has successfully broken into the iPhone and showed the results through a series of demos at the Hexacon conference in Paris. The project says the full write-up will be published in due course. The available material gives no supported models, iOS versions, or exploit chain, so what this item actually pins down is the breakthrough the team claims and the effects they have already demonstrated. [Source: Calif Blog](https://blog.calif.io/p/iphone)
The technical questions it stirs up are whether iOS's system restrictions can be bypassed to gain higher privileges, and whether research like this can help devices that are no longer supported stay in use. Some in the community have asked whether a PDF parsing library is involved, and others are hoping for a practical jailbreak for older devices; those are open questions in the discussion, not technical conclusions to write on the team's behalf. There is still distance between a single demo and a reusable jailbreak tool, and the material at hand is not enough to describe the result as any iPhone being remotely compromised. [Source: Calif Blog](https://blog.calif.io/p/iphone)
OpenSSL 4.1.0 Adopts DTLS 1.3 as Old Build Targets Exit
The last item lands at the plumbing of communications software. OpenSSL 4.1.0 has been released, adding support for DTLS 1.3 — that is, RFC 9147 — along with an ossl-guide-dtlsv13(7) man page to help software built on OpenSSL adopt it. DTLS targets connectionless datagram transport; compared with DTLS 1.2, the new version streamlines the handshake and record protection, with an eye on lower overhead and simpler implementation. This update gives the software around it an implementation entry point for the new protocol version. [Source: OpenSSL Release Notes](https://github.com/openssl/openssl/releases/tag/openssl-4.1.0)
The upgrade also carries compatibility changes: Configurations has dropped the VC-WIN64I target for Windows-on-Itanium and the VC-CE target for Windows CE, so software maintaining those platforms needs to plan a migration, a compatibility path, or staying on the older release. The release notes also state that some performance regressions introduced in OpenSSL 3.0 have been fixed, though the available material gives no specific fixes or performance figures; the new DTLS 1.3 implementation also has 4 known issues. A new protocol version, performance fixes, and the tradeoff around old platform support all land in the same release. [Source: OpenSSL Release Notes](https://github.com/openssl/openssl/releases/tag/openssl-4.1.0)
Today's items can be strung together along the line where technology meets the limits of its use: Kimi puts model capability into API pricing and latency choices, Google's cross-app feature brings personal context and data usage face to face, the iPhone research touches system privileges, and OpenSSL advances protocol support while adjusting its compatibility range. Reading these developments means looking at what was announced and also at how far it has shipped, and which boundaries moved along with it.
