Logo
Bono AI
3 min de lecture Bono AI Team

News of the Day — April 11, 2026

Daily AI watch: Anthropic Advisor Strategy, Shopify AI Toolkit, Perplexity Computer for Enterprise, NVIDIA Robotics Week (Isaac GR00T, Newton 1.0), Florida AG probe into OpenAI, and Sam Altman admits ChatGPT Voice runs on an older model.

News of the Day — April 11, 2026

Daily AI watch for bonoai.org. Topics selected for novelty and relevance to the site’s core themes: open-source AI, in-browser AI, LLM developments, regulation, and notable launches.

1. Anthropic launches the Advisor Strategy: let Sonnet or Haiku consult Opus in a single API call

Summary — On April 9, Anthropic unveiled a new server-side tool, advisor_20260301, which lets a cost-efficient executor model (Sonnet or Haiku) consult a higher-capability advisor model (Opus) on the fly, inside a single /v1/messages call. The executor runs the task end to end — calling tools, reading results, iterating — and only pings the advisor when it hits a blocker. Result: Sonnet with Opus as advisor gains 2.7 percentage points on SWE-bench Multilingual while costing 11.9% less per agentic task; Haiku jumps from 19.7% to 41.2% (more than double its solo score).

Why it matters — This is a direct architectural answer to OpenAI’s compute argument. Instead of stacking more GPUs, Anthropic is selling a “smart routing” pattern that pools collective model intelligence at inference time. The pattern itself isn’t new in the literature (cascades, API mixtures of experts) — what’s new is that the provider is formalizing it as a standardized server-side tool. For the open-source world, this definitively legitimizes small-to-large model routing strategies (vLLM, LiteLLM, RouteLLM).

Suggested angle — Reproduce the advisor pattern with open-source models: route between a Llama 3.3 8B and a GLM-5.1 or Qwen2.5-72B via LiteLLM. Benchmark quality/cost gains on a typical agentic task (SWE-bench lite).

Sources

2. Shopify opens its platform to coding agents: official AI Toolkit for Claude Code, Cursor, Codex and Gemini CLI

Summary — On April 9, Shopify released the Shopify AI Toolkit, an open-source plugin that wires the leading AI coding agents (Claude Code, OpenAI Codex, Cursor, Gemini CLI, VS Code) directly into the Shopify platform. The toolkit gives agents live access to Shopify’s documentation, API schemas, code validation, and CLI — letting them execute real changes on a production store from a plain-English instruction. Installation is two commands in Claude Code or a single click in Cursor.

Why it matters — This is the first time a major e-commerce platform hands full operational control of a live store to external agents. For the MCP / standard-agents movement, it’s industrial validation: Shopify isn’t reinventing a protocol, it’s adopting the existing agents. The IDE/application boundary dissolves — the IDE becomes the admin interface.

Suggested angle — Map the platforms that now accept coding agents in production (Shopify, Vercel, Stripe, Sentry) and compare the safety rails they ship (dry-run, rollback, audit logs). What can an open-source stack learn from this?

Sources

3. Perplexity Computer goes enterprise: 20 orchestrated models, native Slack/Snowflake, $40/seat Enterprise Pro tier

Summary — At its inaugural Ask 2026 conference, Perplexity announced general availability of Perplexity Computer for Enterprise. The multi-model agent orchestrates up to 20 AI models (OpenAI, Anthropic, Google) and connects natively to Slack, Snowflake, Salesforce, HubSpot and hundreds of other platforms. The enterprise tier adds SOC 2 Type II compliance, SAML SSO, audit logs and isolated sandboxing per query. Pricing: Enterprise Pro at $40/seat/month, Enterprise Max at $325/seat/month. Perplexity says more than 100 enterprise customers requested access in a single weekend.

Why it matters — Perplexity is explicitly going after Microsoft Copilot and Salesforce turf. The company’s pivot from “search engine” to “multi-model enterprise action agent” is now locked in. The fact that a single product orchestrates 20 competing models is a breath of fresh air for multi-vendor strategies — and a direct counter-signal to the single-vendor lock-in pushed by OpenAI or Anthropic.

Suggested angle — Head-to-head: Perplexity Computer Enterprise vs Microsoft 365 Copilot vs Claude Cowork vs ChatGPT Enterprise. Which depends on which model, which exposes its connectors, which lets the user pick? Where does an open-source project like Oh my AI! fit in?

Sources

4. NVIDIA National Robotics Week: new Isaac GR00T and Cosmos models, Newton 1.0 physics engine reaches GA

Summary — NVIDIA used National Robotics Week (April 10) to drop a salvo of releases for robotics developers. Highlights: new open Isaac GR00T vision-language-action models that let robots understand natural language and run multi-step manipulation tasks with VLA reasoning; new Cosmos world models for synthetic training data; general availability of the open-source Newton 1.0 physics engine (co-developed with Google DeepMind and Disney Research); and GA releases of Isaac Sim 6.0, Isaac Lab 3.0 and Omniverse NuRec. Standout result: Mimic Robotics reports 10× better sample efficiency and 2× faster convergence on real-world manipulation tasks.

Why it matters — The open-source robotics ecosystem is standardizing around NVIDIA: models on Hugging Face (LeRobot), simulator (Isaac Sim), and physics engine (Newton) all open-source or openly accessible. For physical-AI researchers this is now a coherent stack that rivals MuJoCo/Gymnasium. The message: physical AI has entered its product phase.

Suggested angle — Can Isaac GR00T N1.6 run locally on a Jetson Orin Nano — or even in WebGPU for a browser demo? Analysis of open VLA model sizes and how accessible they actually are for makers.

Sources

5. Florida Attorney General opens investigation into OpenAI: ChatGPT tied to the FSU shooting

Summary — On April 9, Florida Attorney General James Uthmeier announced a formal investigation into OpenAI. The grievances: risks to minors, public safety, and — most prominently — ChatGPT’s alleged role in the April 17, 2025 Florida State University shooting (2 dead, 5 wounded), where the gunman had exchanged more than 200 messages with ChatGPT, including questions about a shooting at FSU. Attorneys for one of the victims have already filed suit against OpenAI. The probe also references child sexual abuse material, self-harm encouragement among minors, and concerns about data flows to foreign actors.

Why it matters — This is the first US state AG investigation squarely focused on product-liability for a consumer LLM in a mass-violence event. Unlike Europe (AI Act), the US has no federal framework — so state AGs are becoming the de facto regulators of frontier models. For AI labs, this formalizes a new legal risk: a post-deployment duty of care.

Suggested angle — A map of ongoing US legal actions against LLMs (Character.AI, OpenAI, Google) and what it means for open source — who carries liability when an open model is involved in an incident?

Sources

6. Sam Altman admits ChatGPT Voice runs on an older, weaker model — fix “about a year” away

Summary — On April 10, Sam Altman publicly acknowledged that ChatGPT’s Advanced Voice Mode runs on a GPT-4o-era model (April 2024 knowledge cutoff), significantly less capable than OpenAI’s current text models. The CEO confirmed a viral user example: Advanced Voice Mode cannot reliably measure a time interval or run a real timer — it fabricates the result instead. Estimated fix: “about another year.” The gap is all the more striking given that Codex, OpenAI’s top-tier coding model, can coherently refactor entire codebases on the other side of the product line.

Why it matters — The admission cuts two ways. First, multimodality is not yet unified at OpenAI: dedicated per-modality models still exist, and capability gains don’t propagate uniformly. Second, it reopens the case for modular voice pipelines (STT → LLM → TTS) — the open-source default (Whisper + Llama + Voxtral / F5-TTS / Kokoro) — which benefits directly from improvements to the central text model. For Oh my AI!, the takeaway is simple: a transparent, composable pipeline beats a year-old multimodal black box.

Suggested angle — Build and demo a fully in-browser voice pipeline, 100% WebGPU: Whisper-Web for transcription, Llama/Qwen via WebLLM for reasoning, Kokoro-JS or Voxtral-TTS for synthesis. Compare quality and latency to ChatGPT Advanced Voice Mode.

Sources

Daily watch compiled on April 11, 2026 by bonoai.org’s AI agent.