OpenAI Unveils Its Own Chip "Jalapeño"

In today's issue

01 · Models

OpenAI unveils Jalapeño, its first custom inference chip built with Broadcom

OpenAI unveiled Jalapeño, its first custom inference chip built with Broadcom, targeting the massive cost of running ChatGPT and other models at scale. Early tests show better performance-per-watt than today's alternatives, letting OpenAI chip away at inference costs that directly hit its bottom line.

The details:

The chip is purpose-built for inference workloads like real-time coding; pre-training will continue on Nvidia hardware for now.
The Broadcom partnership launched in October, mirroring Google's TPU strategy and Amazon's custom Trainium and Inferentia accelerators.
OpenAI frames Jalapeño as part of full-stack optimization—chip design, kernels, memory, networking, and scheduling all tuned around the same goal of speed and cost.

Why it matters: OpenAI is following the playbook Google and Amazon wrote: custom silicon carves out margin on high-volume inference workloads where off-the-shelf GPUs are overkill. Jalapeño won't dethrone Nvidia overnight, but it signals OpenAI's shift from pure software player to infrastructure builder—and every workload that moves off Nvidia is revenue that stays in-house.

Read the full story →

02 · Business

Anthropic ships Claude Tag, a Slack-native virtual employee for enterprises

Anthropic launched Claude Tag, a Slack-native agent that works like a shared virtual employee—already approving 65% of code changes inside Anthropic's own product team. The move targets enterprises tired of opaque, siloed AI work and directly challenges Salesforce's Slackbot.

The details:

Claude Tag lets any team member tag the bot in a channel, watch it decompose and execute tasks, and intervene mid-workflow—all visible to colleagues for full auditability.
Anthropic now leads OpenAI in enterprise adoption for the first time: 34.4% of U.S. firms (50,000+ tracked by Ramp) pay for Anthropic services vs. 32.3% for OpenAI, driven mainly by Claude Code.
Admins can scope tool access, channel permissions, and memory per instance, plus set token-spend caps per channel to prevent budget overruns and data leaks.

Why it matters: Claude Tag is Anthropic's play to lock in its newfound enterprise lead before OpenAI and Google ship competing Slack agents. The multiplayer, auditable design solves the opacity problem that has stalled most agent pilots—visibility and control are table stakes now, and whoever ships them first in the workflow layer wins mindshare. Watch whether this accelerates or triggers competitor launches over the next two quarters.

Read the full story →

03 · Models

GPT-5 Pro helps immunologist crack a 3-year T cell mystery

GPT-5 Pro helped immunologist Derya Unutmaz solve a 3-year-old mystery about T cell behavior, supplying the mechanistic insight that had stalled his lab's research. The breakthrough has implications for cancer and autoimmune disease therapies, where understanding T cell dysfunction is central to modern treatment pipelines.

The details:

Unutmaz is an independent immunologist, not an OpenAI employee—the detail that gives weight to OpenAI's pitch that GPT-5 Pro generates testable hypotheses domain experts actually use.
T cell research underpins checkpoint inhibitors, CAR-T therapies, and most major pharma autoimmune pipelines, making any model-assisted insight into anomalous behavior a potential input to drug programs.
GPT-5 Pro is OpenAI's premium reasoning tier, priced higher than base GPT-5 and marketed explicitly at high-stakes work like immunology, drug discovery, and clinical interpretation.

Why it matters: OpenAI is moving GPT-5 Pro upmarket by collecting named-scientist endorsements in domains where a single correct hypothesis can unlock months of lab work. One anecdote doesn't prove the model works reliably across many labs, but the arithmetic is compelling: if GPT-5 Pro generates even occasional insights worth testing in immunotherapy research, the API cost vanishes against the value of a new therapeutic angle. This is the test case for whether reasoning models become standard research infrastructure.

Read the full story →

AI Box

Every AI model. One chat.

The latest models from ChatGPT, Claude, Gemini, Sora, ElevenLabs — 80+ models in a single chat. Compare answers side by side. Pick the best one every time.

● ChatGPT, Claude, Gemini, Grok, DeepSeek — in one chat

● Generate images & video with Sora, Veo, Ideogram

● Compare any two models side by side

● From $8.99/mo · 80+ models, all included

Try AI Box →

aibox.ai

Trusted by 3,000+ teams

04 · Tools

Meta drops Ray-Ban branding for $299 smart glasses line

Meta launched a $299 smart glasses line without Ray-Ban branding, undercutting its Ray-Ban Meta Gen 2 by $80 and bringing the form factor closer to impulse-buy price territory. The move signals Meta's bet that dropping premium co-branding unlocks mass-market adoption where higher price points stalled.

The details:

Three styles ship at launch—Fury, Adventurer, and Meta Glasses by Kylie—across seven colors, with the same internals as Ray-Ban Meta Optics Styles but longer battery life.
Muse Spark AI, the first model from Meta Superintelligence Labs, now supports 14 new languages including Mandarin, Hindi, Japanese, Arabic, and Korean.
Adjustable nose pads, bendable temple tips, and overextension hinges target glasses-wearers directly; prescription support ranges from -12 to +2.25 diopters without optician fitting below -6.

Why it matters: Meta's willingness to shed Ray-Ban branding for volume is a sharp move—EssilorLuxottica still makes the frames, but dropping the luxury label removes pricing friction without sacrificing manufacturing quality. The real play is Muse Spark rolling out across the entire Ray-Ban Meta installed base. That's how you build an AI platform on wearables: lock in adoption with cheaper hardware, then push intelligence gains to the whole user base at once.

Read the full story →

05 · Models

Google bakes computer use into Gemini 3.5 Flash as a native tool

Google DeepMind baked computer use directly into Gemini 3.5 Flash as a native tool, retiring its standalone model and letting developers build agents that control browsers, apps, and desktops through a single API call. This collapses what was a two-model workflow into one, unlocking agents that can stay coherent across long sequences of clicks and form fields.

The details:

The integrated version shipped June 24, 2026, and delivers better performance on agentic tasks than the previous separate Gemini 2.5 computer use endpoint.
Two optional enterprise safeguards ship with it: human confirmation gates before irreversible actions like checkouts or deletes, and automatic task halts on detected prompt injection.
Developers can test the capability in Browserbase's hosted demo sandbox before wiring it into the Gemini Enterprise Agent Platform for production deployments.

Why it matters: Google just collapsed the agent stack. Instead of bolting computer use onto a generalist model, Flash now ships with sight, reasoning, and action as a single substrate — function calling, search, maps, and screen control all in one context. That's a friction kill for enterprise builders. Anthropic and OpenAI have computer use too, but integrating it as cleanly as this raises the bar for the whole market.

Read the full story →