AI news for June 13, 2026

Researchers (including Tsinghua) release Count Anything, a model designed to count objects reliably across diverse image types.

• Intended domains: crowds, satellite imagery, medical scans, bacterial colonies, etc.
• Method: merge a bounding‑box detector for large objects with a dot‑based counter for dense small objects; resolve duplicates by confidence.
• Builds on Meta’s SAM3 with adapter modules; introduces CLOC dataset (~220,000 images, 619 categories) for text‑guided counting.

Accurate cross‑domain counting reduces manual inspection costs in medicine, agriculture and infrastructure and makes automation of quantitative visual tasks more practical.

For you If you need automated counts (inventory, inspections, monitoring), pilot Count Anything on a sample of your images and compare against existing tools.

A coalition of US state attorneys general has launched an investigation into OpenAI and served a subpoena.

• New York served a subpoena seeking documents on advertising, user engagement, model sycophancy, handling of consumer and health data, and treatment of minors and seniors.
• OpenAI says it will engage constructively and highlights existing safety features for minors and parental tools.
• OpenAI faces multiple ongoing lawsuits on copyright, safety and other liabilities.

Investigations by state AGs signal growing regulatory scrutiny that could force changes in product features, data handling and legal exposure for AI vendors and their customers.

For you Monitor legal developments and ensure your AI contracts and data practices meet rising regulatory expectations; add clauses covering vendor cooperation in legal inquiries.

Apple offers a developer preview of Siri AI in macOS 27 Golden Gate; it shows improvements but limited desktop capabilities.

• Siri AI appears in the macOS 27 developer beta with noticeable improvements over prior versions.
• On Macs the assistant’s limitations are clearer: keyboard/mouse remain faster for many desktop tasks.
• Can launch apps but cannot perform in‑app actions; Shortcuts automations are currently limited.
• File/folder indexing on test units appears incomplete and lacks a visible progress indicator.

Indicates potential productivity gains for Apple shops but also shows Siri AI currently cannot replace deeper desktop automations or integrations.

For you Pilot Siri AI on representative Mac workflows and don’t rely on it for in‑app automation yet.

Microsoft and Meta caution against indiscriminate use of frontier models and implement controls after large internal AI spending increases.

• Satya Nadella criticises uncritical use of powerful models and warns token overuse won’t drive real growth.
• Meta reports exponential internal AI use and projected billions of dollars in costs; plans budgets, allocations and a central ‘AI Gateway’ dashboard starting 2027.
• Meta will steer employees toward internal tools and implement alerts for unusual cost spikes; leadership says token usage isn’t a proxy for impact.

Unchecked consumption of expensive models creates substantial operating costs and misaligned incentives; cost governance is a practical necessity for businesses using cloud AI.

For you Introduce usage monitoring, allocate budgets for AI consumption, and prefer cheaper specialized models for routine tasks to control costs.

A user builds a functional yard‑management app with Gemini via a single detailed prompt, but encounters runtime bugs requiring manual fixes.

• Gemini produced a preview window app within minutes from a long prompt.
• Runtime errors appeared (e.g. 'Channel is unrecoverably broken'), requiring the user to click a fix button.
• After manual intervention, Gemini reported success in 233 seconds and used technical terms like 'race conditions'.
• Project moved from simple automations to a more ambitious prompt‑driven app build.

Shows LLMs’ power to rapidly prototype apps, while highlighting persistent needs for human debugging and oversight in production‑grade projects.

For you Use generative models for fast prototyping but allocate developer time for debugging and code review before deploying.

Google Research debuts Gemini‑SQL2, built on Gemini 3.1 Pro, which achieves 80.04% execution accuracy on the BIRD benchmark — the current leader according to Google.

• Gemini‑SQL2 posts 80.04% execution accuracy on BIRD
• Comparative scores: GPT‑5.5‑xhigh ~72.8%, Claude Opus 4.6 ~70.9%
• Translates natural language into executable SQL; no public release date or paper announced

Stronger text‑to‑SQL reduces friction for non‑technical staff to query business databases, enabling faster ad‑hoc analysis and potentially lowering BI costs for SMEs.

For you Evaluate whether natural‑language querying can speed up your BI workflows: run pilots with current text‑to‑SQL offerings on representative datasets before committing to integrations.

Microsoft and academic partners introduce SkillOpt, a method that treats skill documents as trainable state and reportedly boosts GPT‑5.5 by over 20 points on procedural tasks.

• SkillOpt uses a separate LM optimizer to iteratively edit a Markdown skill document for a frozen target model
• Edits are accepted only if they improve performance on a held‑out validation set
• Tested across six benchmarks (search, spreadsheets, document analysis, math, embodied action) with notable gains

SkillOpt offers a practical alternative to model fine‑tuning: improving agent behavior via editable instruction artifacts can lower cost and increase maintainability for production agents.

For you Consider implementing structured skill documents for your AI agents and run validation‑driven optimization cycles before paying for model retraining.

Meta is introducing tighter token controls, budgets and a central 'AI Gateway' after internal AI usage surged and reportedly risks costing billions by 2026.

• Internal memo cites exponential AI usage growth and projected billions in internal costs
• From 2027: budgets, allocations and a central dashboard to track AI usage and spending
• Previous 'tokenmaxxing' behavior produced 73.7 trillion tokens in just over 30 days

This underlines how unchecked model consumption can create major operational costs; companies need governance to keep AI usage aligned with business value.

For you Put token governance in place now: allocate budgets, monitor API consumption centrally, and prefer cost‑efficient internal tooling for heavy workloads.

Moonshot AI releases Kimi K2.7 Code as an open‑weights model for programming and agentic coding workflows, touting far lower price‑per‑token than GPT‑5.5 and Claude.

• K2.7 Code available on Hugging Face; successor to K2.6 aimed at long‑running software engineering tasks
• MoE architecture: 1 trillion parameters total, 384 experts, 8 active per token; 256k token context length
• Benchmark gains over K2.6; still trails GPT‑5.5 on several coding benchmarks but outperforms Claude on some agentic tests
• Moonshot claims up to ~12x cheaper price per token versus GPT‑5.5/Claude

An open, coding‑focused model with huge context and lower cost is appealing for software SMEs and tooling vendors that run expensive agentic developer workloads.

For you Pilot Kimi K2.7 Code for heavy, agentic development tasks (CI agents, automated testing/refactoring) and benchmark cost versus closed models before production rollout.