AI News June 2026: The Month Anthropic Passed OpenAI (and Everyone Else Kept Moving)
We’re back from the editorial break, and the AI world waited for no one. Between May 8 and June 6, 2026, roughly a full year of tech got compressed into four weeks: Google I/O dropped a whole new Gemini family, Microsoft shipped seven in-house models at Build, and Anthropic closed the largest private funding round in history — becoming, in the process, the most valuable AI startup on the planet, passing OpenAI.
In the middle of it all, the White House signed an AI executive order, NVIDIA open-sourced an omnimodel for robotics, Intel packed 288 cores onto a single chip, and Brazil’s Congress set a date to vote on its AI legal framework. This is the iabrief comeback edition: a monthly digest, organized by theme, with what matters and links to go deeper wherever you want.
Launches: Google I/O and Microsoft Build Own the Month
Google I/O 2026: the Gemini 3.5 family arrives
At I/O on May 19–20, Google launched Gemini 3.5 Flash as the first model of the new generation — and the bar went up. It beats Gemini 3.1 Pro on coding and agentic benchmarks like Terminal-Bench 2.1 (76.2%) and MCP Atlas (83.6%), running up to 4x faster in tokens per second than other frontier models. It’s already live in the API, the Gemini app, Search, and Antigravity (Google blog).
- Gemini 3.5 Pro is in internal testing and was promised for June — the bigger sibling tuned for the heaviest tasks (9to5Google).
- Gemini Omni is the multimodal bet: generate any output from any input, starting with video. Combine image, audio, text, and video, then edit by conversation — say what to change and the model re-renders, no timeline or layers (cybernews).
- Gemini Spark is a 24/7 personal agent that acts on your behalf under your direction, integrated with Gmail, Docs, and Slides, running on Gemini 3.5 and the Antigravity harness.
- Antigravity 2.0 became a full agentic suite: a desktop app running multiple agents at once, a CLI for those who prefer the terminal, and an SDK for custom workflows (Memeburn).
Microsoft Build 2026: the MAI family gets seven models
At Build (early June), Microsoft’s superintelligence team launched seven in-house models — a clear move to reduce its dependence on OpenAI. The highlights:
- MAI-Thinking-1, the team’s first reasoning model: 35 billion active parameters, a 256K context window, and a focus on low cost per token.
- MAI-Code-1-Flash, a 5-billion-parameter coding model already in GitHub Copilot and VS Code, solving tasks with up to 60% fewer tokens (Microsoft AI).
- Rounding out the family: MAI-Image-2.5 (text-to-image and image-to-image), MAI-Voice-2 (15+ new languages), and MAI-Transcribe-1.5 (43 languages) (Neowin).
Models: Claude Opus 4.8 Leads, MiniMax M3 Surprises
Claude Opus 4.8 sets the coding pace
Released on May 28, Claude Opus 4.8 arrived leading software-engineering benchmarks. On SWE-bench Verified it hits 88.6% (vs. Gemini 3.1 Pro’s 80.6%); on the harder SWE-bench Pro, 69.2%, more than 10 points clear of GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%). Pricing stays at $5 / $25 per million input/output tokens (Vellum). If you want to compare options before picking a tool, see our guide to the best AI model in 2026.
MiniMax M3: frontier capability at 1/20 the compute
China’s MiniMax launched M3 on June 1, the first open-weight model to combine frontier coding, a 1-million-token context window, and native multimodality. The technical novelty is MiniMax Sparse Attention (MSA): it processes only the relevant blocks of the cache, delivering per-token compute at 1/20 of the previous generation, with 9x faster prefill and 15x faster decode. On SWE-bench Pro it reaches 59.0%, beating GPT-5.5 and Gemini 3.1 Pro per the company (MiniMax). The lesson of the month: efficiency is now as strategic as raw capability. To see why that matters, read our explainer on multimodal AI.
Market & Funding: Anthropic Becomes the Most Valuable
The story of the month: Anthropic closed a $65 billion Series H on May 28 at a $965 billion post-money valuation, passing OpenAI ($852B from March) to become the most valuable AI startup in the world (Anthropic). Talks started smaller — in May, reports cited raising around $30B at a valuation above $900B — but demand inflated the round.
- Round leads: Altimeter, Dragoneer, Greenoaks, and Sequoia Capital.
- Revenue: run-rate crossed $47 billion in May, and the company projects topping $50 billion by the end of July — roughly 80x growth in two years (VentureBeat).
- IPO: on June 1, Anthropic confidentially filed for an IPO with the SEC, targeting a possible debut as early as October 2026 — ahead of OpenAI (CNBC).
On the broader backdrop: Q1 2026 had already broken every record, with ~$300 billion in global venture and over 80% going to AI (Crunchbase). Q2 marks the shift from private records to public markets: Cerebras IPO’d in May, raising $5.55B, and AI agents became the fastest-growing software category. For the backdrop to this valuation race, read our analysis of OpenAI at $852 billion.
Research & Hardware: NVIDIA Opens Cosmos 3, Intel Stacks Cores
NVIDIA Cosmos 3: an open omnimodel for physical AI
On June 1, NVIDIA launched Cosmos 3, billed as the first fully open omnimodel for physical AI. Built on a mixture-of-transformers architecture, it understands and generates text, images, video, ambient sound, and actions with leading physics accuracy, cutting robot and autonomous-vehicle training and evaluation cycles from months to days. It was trained on 20 trillion multimodal tokens and launched at Computex/GTC Taipei (NVIDIA Newsroom). It comes in two sizes: Super (32 billion parameters) and Nano (8 billion).
Intel Xeon 6+: 288 cores on 18A
Launched in early June at Computex 2026, the Xeon 6+ “Clearwater Forest” is Intel’s first data-center CPU on the 18A process, with up to 288 E-cores, 576 MB of cache, and an explicit pitch for the agentic-AI boom (DataCenterDynamics).
Policy & Regulation: A US Executive Order
On June 2, the White House signed the executive order “Promoting Advanced Artificial Intelligence Innovation and Security.” The practical points:
- It creates a voluntary benchmark to assess models’ “advanced cyber capabilities” and define what counts as a “covered frontier model.”
- It asks for government access to models up to 30 days before public release — on a voluntary basis (CNBC).
- It establishes a cybersecurity clearinghouse coordinated with industry.
- Importantly: the order explicitly bans any mandatory licensing or pre-clearance to develop and release models (White House).
If you build or run a business on AI, watch this closely — that “early access” detail could become a norm. For the full picture, see our overview of US AI regulation in 2026.
Brazil: Congress Sets a Vote on the AI Legal Framework
In Brazil, lower-house speaker Hugo Motta confirmed that AI regulation will be voted on in June. Rapporteur Aguinaldo Ribeiro is expected to present the text around June 9, with a committee vote before the July 18 recess (Diário Carioca).
- The basis is Bill 2,338/2023, approved by the Senate in December 2024.
- The government’s proposal leans on a risk matrix — regulating by each application’s potential impact rather than a single rigid law (Agência Brasil).
- Because the lower house will amend the text, it has to return to the Senate — so final sign-off is unlikely before the second half of the year.
For creators and companies operating in Brazil, the risk classification is worth reading closely: it will define concrete transparency and governance obligations.
Curiosity of the Month: the “Round-Trip” Goes Mainstream
Anthropic’s raise laid bare the pattern that defines 2026: capital circulates in a loop. Investors put money into the labs, which buy NVIDIA chips and Amazon/Microsoft cloud, which in turn invest back into the labs. It’s “ecosystem integration” to some, “financial round-tripping” to others. It works as long as real revenue grows — and for now, with run-rates jumping from $9B to $47B in months, it is. The risk is that everyone discovers, at the same time, that part of the demand was just their own money coming back.
Frequently Asked Questions
Is Anthropic really more valuable than OpenAI now?
Yes, on private valuation. The Series H closed May 28 at $965 billion post-money, above OpenAI’s $852 billion (March round). Valuation isn’t the same as revenue or profit, but it’s the yardstick the market uses to rank the labs.
Is Gemini 3.5 Pro available yet?
As of this edition’s close (June 6), no. Google confirmed 3.5 Pro was in internal testing with a June launch promised. Gemini 3.5 Flash, however, is already generally available in the API and Gemini products.
What changes with Microsoft’s MAI models?
Microsoft now has its own family of seven models — reasoning, code, image, voice, and transcription — reducing its dependence on OpenAI inside Copilot. For developers, MAI-Code-1-Flash already shows up in GitHub Copilot and promises to do more with fewer tokens.
What’s the best coding model right now?
In public May/June benchmarks, Claude Opus 4.8 leads on SWE-bench Verified (88.6%) and SWE-bench Pro (69.2%). But “best” depends on your case: cost, speed, and context matter as much as the top of a benchmark. MiniMax M3, for instance, delivers a 1M context window at a fraction of the cost.
Does the US executive order create mandatory AI licensing?
No. The June 2 order explicitly bans any mandatory licensing or pre-clearance. What it creates is a voluntary benchmark and a request for early access (up to 30 days) to frontier models.
When does AI regulation take effect in Brazil?
There’s no effective date yet. The lower-house vote is set for June 2026, but the text must return to the Senate before presidential sign-off — which pushes the timeline to the second half of the year at best.
What to Watch in the Coming Weeks
Three fronts will dominate June and July: the actual launch of Gemini 3.5 Pro (and how it stacks up against Opus 4.8 and GPT-5.5), the market’s read on Anthropic’s October IPO, and the outcome of Brazil’s AI legal-framework vote. On the hardware side, watch whether MiniMax M3’s radical efficiency pressures API prices worldwide.
To follow upcoming updates, subscribe to the iabrief newsletter — one email per week, no hype, only what matters.
