# OpenAI’s Jalapeño Chip Leads a Day of Agent and Infrastructure Shifts

*By AI High Signal Digest • June 25, 2026*

OpenAI’s custom Jalapeño chip led the day, alongside Gemini’s new computer-use mode and a $200M launch for MirendilAI. Elsewhere, GLM-5.2 advanced open-model benchmarks, while the first legal challenge to US AI export controls moved policy risk into court.

## Top Stories

*Why it matters: today’s clearest signals were about who controls AI compute, how far agents can act on software surfaces, and where new frontier R&D money is going.*

- **OpenAI unveiled Jalapeño, its first custom inference chip.** Built with Broadcom for ChatGPT, Codex, the API, and future agentic products, it extends OpenAI’s stack into infrastructure [^1]. OpenAI said the program went from initial design to tape-out in nine months, used ChatGPT in the engineering process, is already running GPT-5.3-Codex-Spark in the lab, and targets substantially better performance per watt, with deployment planned for end-2026 [^2][^3]. The stated goal is lower dependence on external GPUs and tighter control over compute economics [^3].

- **Google launched computer use in Gemini 3.5 Flash.** The feature lets an agent take a screen and a goal, then determine actions across browser, mobile, and desktop environments [^4]. Shared examples show it auditing docs pages by navigating, running code snippets, taking screenshots, and returning a report, with safeguards including user confirmation and auto-stop on prompt injection [^4].

- **MirendilAI launched with a $200M seed round.** The startup says it will build self-accelerating AI R&D systems, with a 20-person founding team from Anthropic, xAI, Google DeepMind, and OpenAI; the round was led by a16z and Kleiner Perkins, with a major NVIDIA investment [^5]. Its pitch is to democratize frontier AI capabilities for broader scientific use rather than concentrate them in a few labs [^5].

## Research & Innovation

*Why it matters: the most notable technical updates focused on open-model progress, learning limits, and better ways to compose agents.*

- **GLM-5.2 posted the strongest ARC-AGI-2 result yet for an open-source model.** Verified scores were 22.8% on ARC-AGI-2 at $0.25 and 77.0% on ARC-AGI-1 at $0.19, with ARC Prize saying performance is comparable to GPT-5.4 and GPT-5.5 at low reasoning effort [^6]. François Chollet called it the strongest ARC-AGI-2 performance to date by an open-source model [^7].

- **Zyphra argued that continual learning hits a deeper failure mode than forgetting.** Its new work identifies plasticity loss as models losing the ability to learn new data, shows it across 5M-314M parameter GPT-style models, and reports the same decline even in stationary pretraining [^8][^9][^10]. The team fit a scaling law for the onset, T ∝ P^0.83, suggesting scale delays the problem but does not remove it [^11].

- **AI21 topped DeepResearch Bench II by merging weak agents instead of building a new one.** It combined seven agents ranked 7-13 into a single report pipeline and reported a new #1 score of 64.38 [^12][^13].

## Products & Launches

*Why it matters: product updates kept pushing models from chat into domain tools and team workflows.*

- **OpenAI updated GPT-5.5 Instant.** The new version is described as better at understanding user intent, adapting responses, handling complex constraints, and improving shopping and local recommendations; rollout started today for paid users and tomorrow for free users [^14].

- **Perplexity launched Computer for Counsel.** The product connects legal research databases, document tools, and matter-management systems so lawyers can pull citable sources from tools including MidpageAI, LegalZoom, Docusign, and NetDocuments [^15]. It is available to Pro and Max subscribers [^15].

- **Notion and Cursor pushed agents deeper into team workflows.** Notion introduced External Agents with Claude and Cursor so teams can assign work from shared boards and @-mention agents like teammates [^16]. Cursor said the integration runs on its SDK so cloud agents can take tasks from Notion and open PRs using the same runtime as Cursor itself [^17].

## Industry Moves

*Why it matters: infrastructure control and talent concentration remain central competitive levers.*

- **Qualcomm agreed to acquire Modular.** Both sides said the deal is meant to unify accelerated compute with an open platform spanning edge to cloud and hardware from CPUs and GPUs to NPUs and custom ASICs [^18][^19][^20].

- **Anthropic is pulling more talent from Google DeepMind.** Bloomberg reported that Jonas Adler and Alexander Pritzel, both viewed internally as key Gemini contributors, are leaving Google for Anthropic [^21].

## Policy & Regulation

*Why it matters: AI governance is shifting from abstract debate toward concrete fights over access and supply-chain alignment.*

- **The first legal challenge to the Trump administration’s AI export controls has arrived.** Legion is suing over the forced shutdown of Anthropic’s Fable 5 and Mythos 5 for foreign nationals, arguing export-control laws do not cover access to hosted AI models or text outputs and that no national emergency was declared [^22][^23]. The case turns on whether hosted frontier-model access can be treated as export-controlled technology when users only receive text outputs [^22].

- **Europe joined a US-led AI supply-chain pact.** The EU, Germany, the Netherlands, and Greece joined Pax Silica, covering chips, critical minerals, energy, and compute; Jacob Helberg explicitly positioned it against digital sovereignty built around duplicative national tech stacks [^24].

## Quick Takes

*Why it matters: these smaller updates still show where deployment patterns are heading.*

- Anthropic’s new **agent identity** model gives Claude its own credentials in shared channels, while DMs run on the user’s connectors, with one auditable identity for admins [^25][^26][^27][^28].
- **Google AI Studio** says more than **1 million Android apps** have been created since native Android app building launched in May [^29].
- **Wan-2.7 I2V** entered Video Arena at **#5**, ahead of Grok Imagine Video and every Google Veo-3.1 variant [^30].
- **Kog** open-sourced the 2B **Laneformer** model used to demonstrate **3,000+ tokens per second** [^31].

---

### Sources

[^1]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2069770172802773292)
[^2]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2069795647956373632)
[^3]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2069772454591934778)
[^4]: [𝕏 post by @_philschmid](https://x.com/_philschmid/status/2069819170477293863)
[^5]: [𝕏 post by @bneyshabur](https://x.com/bneyshabur/status/2069860934148079800)
[^6]: [𝕏 post by @arcprize](https://x.com/arcprize/status/2069845152773099854)
[^7]: [𝕏 post by @fchollet](https://x.com/fchollet/status/2069858556552298519)
[^8]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2069835009415090328)
[^9]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2069835014444142609)
[^10]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2069835020701933822)
[^11]: [𝕏 post by @ZyphraAI](https://x.com/ZyphraAI/status/2069835016763572264)
[^12]: [𝕏 post by @AI21Labs](https://x.com/AI21Labs/status/2069781244250914868)
[^13]: [𝕏 post by @AI21Labs](https://x.com/AI21Labs/status/2069781490137772502)
[^14]: [𝕏 post by @OpenAI](https://x.com/OpenAI/status/2069843083701915755)
[^15]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2069866668671766804)
[^16]: [𝕏 post by @NotionHQ](https://x.com/NotionHQ/status/2069816393395012009)
[^17]: [𝕏 post by @cursor_ai](https://x.com/cursor_ai/status/2069872515548340407)
[^18]: [𝕏 post by @Qualcomm](https://x.com/Qualcomm/status/2069758602420670572)
[^19]: [𝕏 post by @clattner_llvm](https://x.com/clattner_llvm/status/2069769232477192354)
[^20]: [𝕏 post by @clattner_llvm](https://x.com/clattner_llvm/status/2069770902523277349)
[^21]: [𝕏 post by @EdLudlow](https://x.com/EdLudlow/status/2069849445513257419)
[^22]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2069704003311567045)
[^23]: [𝕏 post by @SophiaCai99](https://x.com/SophiaCai99/status/2069577642915811441)
[^24]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2069717182905200731)
[^25]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2069895377080443271)
[^26]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2069895380154884155)
[^27]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2069895378544267619)
[^28]: [𝕏 post by @ClaudeDevs](https://x.com/ClaudeDevs/status/2069895389415874907)
[^29]: [𝕏 post by @GoogleAIStudio](https://x.com/GoogleAIStudio/status/2069920175344750957)
[^30]: [𝕏 post by @arena](https://x.com/arena/status/2069925435191054566)
[^31]: [𝕏 post by @ClementDelangue](https://x.com/ClementDelangue/status/2069839095577084364)