curated · 2026-05-20

Mavis has one serious card: Tencent launched it on Windows, Mac, and Android at once, with local models and task routing. The pitch is clear. Move the agent out of the chat box and into files, system maintenance, and phone-app control. That is a better productivity-entry bet than another Yuanbao-style assistant feature. The proof is still thin. Tencent says Mavis can do basically everything current agents do, but the article gives no task success rate, permission model, model list, or hardware floor. “Works after unplugging the network cable” is a good hook, yet offline task coverage is not specified. Microsoft Copilot has had the Windows surface for ages; the hard part was never placement. It is permissions, error cost, and user trust. Tencent has shipped the entry point first, and the product debt lands with it.

Stitch 3.0 hit #1 on Product Hunt with two notable additions: codebase import, so you can generate UI on top of an existing project instead of starting from scratch, and dynamic UI generation that adapts to data states rather than producing static mockups. There's also a new agent called Hatter for multi-step design tasks. Both sources covering this — Product Hunt and aihot-selected — are working from the same launch page. That's fine for a product announcement, but it means no independent testing, no Google-authored technical deep-dive, and no pricing details beyond a "Free" label with unclear limits. I'd treat this as a product update signal, not a "Google just crushed v0 and Bolt" moment. The feature set overlaps heavily with what v0, Bolt, and Lovable already do. Stitch's edge looks like Google ecosystem integration and a free tier, but I haven't seen anyone run a real project through it yet to compare output quality.

OpenAI disclosed only “GPT-5 is coming soon,” and that thinness is the point. No release date, parameters, pricing, context window, or API conditions are given; even ChatGPT-first versus API-same-day is absent. For builders, this works less like product news and more like a procurement brake: teams evaluating Claude, Gemini, or Qwen now wait for GPT-5’s price, latency, and evals. OpenAI has used pre-announcements to hold market attention before, but GPT-5 has a narrower landing zone. A keynote win is not enough; it needs hard gains in coding, agent tool use, and long-context reliability.

Perplexity is betting on compression around retrieval, not bigger context windows. The concrete claim is aggressive: query-aware compression is in production, cuts context tokens by up to 70%, and improves answer quality. If that holds under live traffic, it hits the two pain points every search-RAG system feels first: latency and inference cost. A lot of teams still sell 200K or 1M context as the feature; Perplexity is saying the useful move is throwing away irrelevant text before generation. I don’t fully buy the quality claim yet. The snippet gives no eval set, latency number, recall loss, or placement in the pipeline. Compression before reranking and compression before generation are different bets. Perplexity has real search volume, which makes this more credible than a lab demo, but without an ablation, “70% fewer tokens” reads like a product metric, not a reproducible result.

OpenAI is trying to move developer context into the ChatGPT account layer, not make phone-based coding normal. The disclosed facts are narrow: ChatGPT mobile now supports Codex, users can ask on mobile, and continue the same thread on desktop. Supported platforms, app versions, rollout scope, and repo access are not disclosed. That matters because cross-device Q&A and an executable coding agent require very different permission models. I do not buy the “leave your laptop at home” framing. GitHub Copilot still lives where the code lives: the IDE. Cursor’s grip comes from local project context. OpenAI’s play is the persistent conversation thread sitting above those tools. If mobile Codex only explains code, it is an entry-point patch. If it can touch repos, PRs, and CI, it starts pressuring IDE-side agents. The snippet only proves the first case.

The Commission is moving compliance into product definition, not post-launch paperwork. Article 6 classification turns on intended purpose, provider statements, and value-chain duties; Annex III use cases and regulated products remain the hard boundary. Article 6(3) leaves exemptions for narrow procedural tasks, preparatory tasks, improving completed human activity, or detecting patterns without replacing human judgment. The painful bit for AI app vendors is profiling. The article’s FAQ asks whether profiling can be exempt, but the body cuts off before giving the answer. HR, credit, education, and risk tools will struggle to sell themselves as “just recommendation systems.” Compared with the voluntary NIST AI RMF in the US, the EU AI Act turns marketing copy, deployment context, and sales claims into classification evidence.

Kling is selling production credibility here, not just resolution. The headline says “world’s first native 4K,” and the body gives one-click true 4K plus adoption by Hollywood teams and Wonder Studios. It does not give pricing, inference cost, clip length, access limits, or any blind comparison against upscaling. For production teams, 4K pixels are cheap; stable characters, shot continuity, and predictable retry cost are the scarce parts. Wonder Studios’ point about native 4K avoiding upscaler-induced character deformation is the right hook. Runway, Pika, and Sora-class systems have all run into the same “sharp but unstable” wall. But without failure rate or generation speed, I’d treat this as Kling buying a ticket into high-end production, not proof that the pipeline is solved.

OpenAI is trading $2M in API credits for equity in each current YC startup; that looks less like investing and more like prepaid distribution. The missing terms matter: equity stake, credit expiry, model access, rate limits, and resale rules are not disclosed. Without them, $2M is a sticker price, not cash value. Early YC teams will wire prototypes, evals, logs, and user flows around OpenAI’s API before they know which parts are durable. Sam Altman’s Yuri Milner reference is a neat origin story, but cloud providers already ran this play. AWS and Google Cloud used credits to become default infrastructure. OpenAI is moving the same tactic up to the model layer. The risk for founders is not cheap inference; it is letting GPT pricing, latency, and tool-call behavior shape the product before the market does.

GitHub’s problem is not that repositories vanish; developers are moving daily work into Cursor and Claude Code. The concrete tell is Microsoft asking some Windows, Microsoft 365, Outlook, Teams, and Surface teams to stop using Claude Code by late June 2026 and move to GitHub Copilot CLI, with security workflow and FY2027 cost control as the official rationale. I don’t buy the “toolchain consistency” story cleanly. Copilot in 2021 lived inside editor autocomplete. Cursor in 2024 made the IDE a task surface. Claude Code in 2025 pushed multi-step engineering into the terminal. The repo can stay on GitHub while the decision layer moves elsewhere. The wild part: Copilot CLI still keeps access to Claude models. Microsoft is not blocking Anthropic’s intelligence; it is blocking Anthropic’s product doorway.

Model pricing is finally showing the compute bill, and app margins take the first hit. Gemini 3.1 Pro is $2/$12 per 1M tokens, Claude Opus 4.7 is $5/$25, and GPT-5.5 is $5/$30 after a short subsidy. Google is still using cloud and TPU economics as the low-price anchor; OpenAI has moved its flagship back into premium pricing. Tunguz frames this as cash getting tight, and I buy the direction, but it is too neat. Inference volume, long context, and agent loops turn output tokens into the cost sink. Teams that built 2025 gross-margin decks on falling token prices now have a bad spreadsheet, not a strategy.