- → Claude Mythos Preview — How It Started (June 1)
- → Gemini 3.5 Flash GA + Gemini Spark
- → Perplexity Comet — The AI-Native Browser
- → GPT-5.5 Instant — New ChatGPT Default
- → ✅ Gemini in Chrome for Android — Now Live
- → Developer Tools: AWS MCP, Managed Agents, Glasswing
- → What's Still Coming in June
- → 🧑💻 Try These Yourself — Gemini Flash, Comet, Polaris, Foundry Local
- → 🛠️ Foundry Local — Full 5-Step Tutorial
- → June 4 — Google Flow + Veo 3 & Gemini Storybook
- → June 5–6 — ChatGPT Dreaming, Ads Manager & Go Global
- → June 9 — 🚨 Claude Fable 5: Mythos-Class for Everyone
- → June 7–9 — NVIDIA Nemotron 3 Ultra, OpenCode & Developer Alerts
- → June 10–14 — MiniMax M3, ZoomMate & Gemini 3.5 Pro Preview
- → June 15–21 — 🚨 Fable 5 Suspended, Siri AI Betas Live, Google Search Agents
- → June 22–30 — ✅ Sonnet 5 Launched, Fable 5 Restored, GPT-4.5 Gone, LongCat-2.0
- → June Verdict — What to Actually Try
🧠 Claude Mythos Preview — The Backstory (→ Now Launched as Fable 5)
The biggest AI story entering June was Anthropic's Claude Mythos. First teased in March and previewed in limited release to security researchers via Project Glasswing, Mythos was cleared for broad release "in the coming weeks." It launched June 9, 2026 as Claude Fable 5 (public) and Claude Mythos 5 (restricted partners).
Mythos is the model Anthropic previously said was too powerful to release publicly. It was originally gated because it demonstrated a "step change" in autonomous cybersecurity capabilities — specifically, the ability to find zero-day vulnerabilities in every major operating system and browser. Anthropic ran it first for defenders, not attackers: Project Glasswing used Mythos to patch thousands of critical software vulnerabilities before the model came anywhere near the general public.
What changed: Anthropic raised $65 billion at a $965 billion valuation in late May 2026 and simultaneously said it had made "swift progress" on safety safeguards. The raise and the Mythos release are directly connected — the company needed the safety story solid before it could unlock the product that justifies the valuation.
By: Anthropic. The most capable Claude model to date — a class above Opus 4.8. Exceptional at cybersecurity, complex reasoning, autonomous agentic tasks, and long-horizon planning. Initially previewed to security researchers via Project Glasswing, now moving to broad release. Benchmarks not yet fully public, but internal Anthropic tests show it outperforms Claude Opus 4.8 (which already outperforms GPT-5.5 in most categories but one).
Who it matters for: Security teams, enterprise automation, and power users running complex multi-step agents. For developers: wait for the API pricing announcement before committing pipelines — Mythos-tier models typically come at 3–5× the cost of Sonnet/Flash alternatives.
Context: Claude Opus 4.8 — What Mythos is Replacing
Before Mythos ships broadly, Claude Opus 4.8 (released May 28, 2026) remains the current best Anthropic model. It outperforms GPT-5.5 in all but one benchmark category, is four times less likely to let code bugs pass without flagging them, and is significantly harder to manipulate into deceptive outputs than Opus 4.7. If you are running enterprise workloads today, Opus 4.8 is the floor — Mythos will be the ceiling.
⚡ Gemini 3.5 Flash GA + Gemini Spark
Google's big post-I/O 2026 landing: Gemini 3.5 Flash is now generally available across the Gemini API, Google AI Studio, and Android Studio. This is the first model in Google's new "Antigravity" architecture — a design philosophy combining frontier-level intelligence with real-time action capabilities at Flash-tier speed and price.
At $0.10 per million input tokens, Gemini 3.5 Flash is the cheapest production-grade model available from a major provider that matches frontier performance on agentic and coding tasks. It supports Managed Agents — a single API call provisions a remote Linux environment where the agent can reason, plan, execute code, manage files, and browse the live web. For startups and developers building agentic products, this is the most cost-effective entry point in June 2026.
By: Google. The most capable Flash-tier model Google has shipped. GA across Gemini API, Google AI Studio, and Android Studio. Built for sustained agentic and coding tasks at speed. Managed Agents support via single API call — remote Linux environment with code execution and live web browsing baked in.
Who it matters for: Developers building high-volume agentic pipelines where cost matters. At $0.10/M input it undercuts Gemini 3.1 Flash and matches DeepSeek V4-Flash for price on mainstream cloud infrastructure — without the self-hosting overhead.
Gemini Spark — The 24/7 Workplace Agent
Gemini Spark is Google's answer to enterprise agentic AI. Available to Gemini Enterprise and Google Workspace customers, Spark is described as a 24/7 personal AI agent that autonomously acts on your behalf — booking meetings, processing emails, summarising documents, and executing multi-step work tasks under your direction. Think of it less as a chatbot and more as a fully automated junior team member that never sleeps and doesn't need Slack onboarding.
By: Google. Enterprise-tier 24/7 AI agent available to Gemini Enterprise and Workspace customers. Autonomously manages tasks — email triage, meeting scheduling, document summarisation, workflow execution — under user direction. Integrated natively across Gmail, Docs, Calendar, and Meet.
Who it matters for: Enterprise knowledge workers spending 30%+ of their week on administrative overhead. Spark isn't a productivity enhancement — it's task delegation at scale. Requires Gemini Enterprise subscription.
🌐 Perplexity Comet — The AI-Native Browser
Perplexity's Comet browser is the most ambitious product launch of mid-2026. Initially previewed for Max plan subscribers ($200/month), Comet is now available globally across iOS, Android, Mac, and Windows — and Perplexity made it free for all users. That is not a typo. A browser that synthesises live web research into cited reports, executes agentic tasks on your behalf, and upgrades Deep Research to state-of-the-art quality — free.
What makes Comet different from Chrome + an AI extension: the AI is not bolted on — it browses with you. When you open a page, Comet can immediately summarise it, compare it against other sources you've visited, extract structured data, or act on it (booking a table, filing a form, reserving a slot). The underlying Deep Research for Pro and Max users now runs on Claude Opus 4.5 and can generate presentations, spreadsheets, dashboards, and websites directly from research results.
By: Perplexity AI. AI-native browser available on iOS, Android, Mac, and Windows. Built-in assistant browses with you — summarising, comparing, extracting, and acting on web content. Deep Research (Pro/Max users) now generates presentations, spreadsheets, and dashboards directly. Runs on Claude Opus 4.5 for enhanced research quality.
Who it matters for: Researchers, analysts, and anyone spending hours cross-referencing sources. Comet doesn't just help you find information — it synthesises and acts on it. The free tier is genuinely capable; Max tier ($200/month) unlocks full Deep Research with output generation.
🎬 Video walkthrough of the biggest AI tool launches in June 2026
💬 GPT-5.5 Instant — The New ChatGPT Default
OpenAI quietly made a significant swap: GPT-5.5 Instant is now the default model across all ChatGPT tiers, replacing GPT-5.3 Instant. The upgrade is not just a version bump — internal evaluations show GPT-5.5 Instant produces 52.5% fewer hallucinated claims than its predecessor on high-stakes prompts covering medicine, law, and finance. That is a meaningful reliability jump for everyday users who don't distinguish between model versions.
It also rolls out in the API as chat-latest, meaning any application that points to the default endpoint gets the upgrade without code changes. Worth noting: GPT-4.5 retires from ChatGPT on June 27, 2026 following a 30-day sunset period. If you have any workflow or application still referencing GPT-4.5, migrate now.
By: OpenAI. New default ChatGPT model replacing GPT-5.3 Instant across all tiers. 52.5% fewer hallucinations than predecessor on medicine, law, and finance prompts. Also available in API as chat-latest. Memory sources now surface what context was used to personalise responses, with individual delete/correction controls.
Who it matters for: Everyone using ChatGPT — you are already on it. For API developers: if you reference chat-latest you get this automatically. ⚠️ Action required: migrate any GPT-4.5 references before June 27 when that endpoint retires.
📱 Gemini in Chrome for Android — Now Live
Gemini in Chrome for Android is now live — a native AI assistant embedded directly into Chrome, not bolted on as an extension. On Android, Gemini works with anything on your screen including Chrome itself, giving it full-page context as you browse. It can deliver key takeaways, clarify concepts, compare options across open tabs, and find answers based on what you're currently reading — without ever switching apps.
Two standout features: Auto Browse — Gemini autonomously completes background tasks (appointment booking, parking reservation, form fills) while you stay in control at each step. Rambler — a voice-to-text polish tool that cleans up spoken messages or builds custom browser widgets from natural language descriptions. On iOS, Gemini is built directly into the Chrome app via the omnibox. Rolling out now to eligible users in select regions; broader global rollout to follow.
By: Google. Native Gemini assistant embedded in Chrome for Android — not an extension, a first-class browser feature. Works with anything on your screen, including Chrome tabs. Summarises pages, compares content across tabs, answers questions in context. Auto Browse for autonomous background task completion. Rambler for voice message polish and natural-language widget creation. Rolling out to select regions; iOS gets Gemini in Chrome omnibox.
Who it matters for: Android users who currently have 3–4 tabs open trying to research the same thing. Auto Browse is the feature to watch — autonomous background task completion from within the browser is a meaningful UX shift. For iOS users: access is via the Chrome omnibox, slightly different experience but now available.
🛠️ Developer Tools: Managed Agents, Glasswing & More
Google Managed Agents — Public Preview
Google launched Managed Agents in the Gemini API in public preview. A single API call provisions a full remote Linux environment where an agent can plan, execute code, manage files, and browse the web — all in a sandboxed, isolated runtime. This is Google's answer to OpenAI's Code Interpreter and Anthropic's computer use — but designed from the ground up for production agentic deployments, not just demos.
By: Google. Single API call provisions a remote sandboxed Linux environment for AI agents — plan, reason, execute code, manage files, browse live web. Fully managed runtime means no infrastructure to spin up. Available through Gemini API and Google AI Studio.
Who it matters for: Developers building production agentic systems who don't want to manage their own execution environments. Pairs naturally with Gemini 3.5 Flash for cost-efficient agentic workflows.
Project Glasswing — Open Source Security Collaboration
Anthropic's Project Glasswing is not a consumer product — but it is the most consequential AI safety deployment of 2026. Using Claude Mythos Preview, Anthropic identified thousands of zero-day vulnerabilities across every major OS and browser, then gave that access to critical industry partners and open-source developers to patch the vulnerabilities before broad model release. June marks when those patches ship publicly and Glasswing transitions from closed to open collaboration.
New image generation and editing tool from Google, integrated into Workspace. Supports natural-language prompts for both creation and in-place editing.
New $100/month tier specifically for developers, technical leads, and advanced creators. Unlocks the full Gemini model suite and priority API access.
Proactive on-device AI for Android. Surfaces suggestions, handles routine tasks, and anticipates needs across apps without requiring a prompt from the user.
Google's most advanced multimodal model. Creates from any input — text, image, audio, video simultaneously. Described as "a leap forward in world understanding, multimodality and editing."
📅 What's Still Coming in June 2026
Microsoft revealed Project Polaris — a purpose-built coding AI that will replace GPT-4 Turbo as the default engine in GitHub Copilot starting August 2026. Polaris runs on Microsoft's own Maia 200 chips, built specifically for code generation, multi-file refactoring, test writing, and autonomous bug fixing. A three-month fallback period lets teams stay on GPT-4 if needed. This is Microsoft's first step toward owning its AI stack end-to-end — separating Copilot from OpenAI's model dependency.
Microsoft open-sourced the Windows Agent Framework at Build 2026 under the MIT licence. WAF lets developers build agents that run on local Windows machines, Windows 365 Cloud PCs, and Azure Arc edge devices — with a built-in human approval queue for privileged actions. Works with any AI model (not just Microsoft's). Visual Studio 2026 ships WAF project templates today. This is how third-party agents will integrate with Windows for the next decade.
Azure Agent Mesh federates AI agent execution across Azure, AWS Bedrock, Google Cloud, on-premise systems, and edge devices — all under a single governance layer (Entra ID + Purview). It's the first Microsoft product explicitly designed as multi-cloud. For regulated industries where agents need to touch data across hybrid environments without losing audit trails, this is the enterprise answer.
Foundry Local hit GA at Build — full AI inference on-device with no cloud dependency, no data leaving the device, no per-token cost. Runs on NPU (Qualcomm Snapdragon X, Intel Lunar Lake), GPU, or CPU. Supports open-source and proprietary models through the same API as Azure AI Foundry cloud. DirectML 2.0 unifies NVIDIA, AMD, Intel, and Qualcomm GPU targets. Download it free for Windows, macOS Apple Silicon, and Linux x64.
Microsoft's own MAI models are now available to all developers for commercial use via Azure AI Foundry: MAI-Transcribe-1 (speech-to-text), MAI-Voice-1 (1 min of audio in under 1 second on a single GPU, with multi-speaker support), and MAI-Image-2-Efficient (fast, cost-optimised image generation). These are Microsoft-built — not OpenAI re-packaged. The MAI team's first public commercial release.
Anthropic filed its IPO with the SEC on June 2, following its $65B funding round at a $965B post-money valuation. It's positioned as the first major AI pure-play to go public, ahead of OpenAI's September 2026 target. Expected listing: October 2026. The S-1 will need to disclose how Anthropic's Public Benefit Corporation structure interacts with public shareholder rights — a first for the frontier AI space.
Claude Opus 4.8 and Claude Sonnet 4.6 are now first-class options in Azure AI Foundry alongside OpenAI, DeepSeek, Llama 4, and Mistral — with the same enterprise SLAs, Entra ID identity, Purview data governance, and billing infrastructure as Microsoft's own models. Enterprise developers can now use Claude through existing Azure agreements without separate Anthropic API credentials.
🧑💻 Try These Yourself — Right Now, Free
Three tools from June 2026 you can personally test today with zero cost and no waitlist. No credit card required for any of them.
🛠️ Hands-On: Try Foundry Local in 5 Minutes (Free, No Cloud)
Foundry Local is the easiest of the June 2026 launches to test right now — it's free, works on Windows, macOS Apple Silicon, and Linux, and runs entirely on-device. Here's how to go from zero to running your first local AI model in under 5 minutes.
Open your terminal and run the installer for your platform:
The installer detects your hardware (NPU, GPU, or CPU) and configures DirectML automatically. Takes ~90 seconds.
Foundry Local uses a Docker-style pull command. For a fast first test, grab Phi-4-mini (3.8B parameters, runs well on CPU):
Download is ~2.2 GB. Other available models: llama-4-scout, mistral-7b, deepseek-r1-8b
Foundry Local exposes an OpenAI-compatible REST API on localhost. Start it with:
You'll see Foundry Local running at http://localhost:11434. It's ready to accept requests.
Use curl or any OpenAI-compatible SDK. It's the same API you already know:
-H "Content-Type: application/json" \
-d '{"model":"phi-4-mini","messages":[{"role":"user","content":"Explain AI agents in 2 sentences"}]}'
client = OpenAI(base_url="http://localhost:11434/v1", api_key="none")
r = client.chat.completions.create(model="phi-4-mini", messages=[{"role":"user","content":"Explain AI agents in 2 sentences"}])
print(r.choices[0].message.content)
The api_key="none" is intentional — Foundry Local doesn't require auth on localhost.
Because Foundry Local is OpenAI-compatible, you can point tools like Open WebUI, Continue.dev (VS Code), or your own app at it with one config change:
-e OPENAI_API_BASE_URL=http://host.docker.internal:11434/v1 \
-e OPENAI_API_KEY=none \
ghcr.io/open-webui/open-webui:main
Everything runs locally. No tokens, no billing, no data leaving your machine.
Pick a prompt scenario below and see the recommended tool from June's launches to handle it. Useful for deciding which new tool is actually worth testing first for your workflow.
🗓️ June 4, 2026 — AI Creative Tools
Google Flow + Veo 3 — AI Filmmaking with Native Audio
Google Flow is the most complete AI video production tool available today. Built on Veo 3 (and now Veo 3.1), it combines text-to-video generation, a timeline editor (Scene Builder), camera controls, character consistency (Ingredients), mood board creation (Whisk), and text-to-image (ImageFX) — all inside one workspace. The headline feature that separates it from every competitor: native audio generation. Dialogue, sound effects, and ambient audio are produced in a single pass alongside the video. No separate pipeline, no syncing — the character's lips move with their voice.
Veo 3.1 extends this with 4K output, spatial audio, up to 8-second clips, and significantly better character consistency across scenes. For content creators, marketers, and indie filmmakers, it removes the most painful bottleneck in AI video: assembling a coherent sequence where characters look and sound the same from clip to clip.
By: Google Labs. Full AI filmmaking pipeline — text-to-video, camera controls, timeline editor, character consistency, and native synchronized audio in a single workspace. Powered by Veo 3.1 with 4K output and spatial audio. Available via Google AI Pro ($19.99/mo) and Ultra ($249.99/mo).
Who it matters for: Content creators, marketers, indie filmmakers, and anyone who has tried AI video tools and been frustrated by character drift or silent output. The camera controls alone make it feel like actual directing rather than just prompting. Full guide + real video examples →
Google Gemini Storybook — Free AI Illustrated Books
Gemini Storybook quietly became one of the most accessible AI creative tools Google has shipped. Type a prompt — or upload a child's drawing or photo — and Gemini generates a full 10-page illustrated storybook with custom art and read-aloud narration in over 45 languages. No design skills needed. No subscription. No credits. It works in the standard Gemini interface under the Storybook option.
Art styles include Watercolour, Claymation, Coloring Book, Sketch, Comic Strip, and Pixel Art. Watercolour and Claymation produce the most consistently impressive outputs. The main limitation: character appearance can drift slightly between pages — the same character may look subtly different on page 4 versus page 1. For casual use and children's stories, this is a non-issue. For professional storyboards, it's worth knowing upfront.
By: Google. Free 10-page illustrated storybook generator with AI narration in 45+ languages. Multiple art styles including Watercolour, Claymation, and Pixel Art. Works directly in the Gemini interface — no extra subscription required.
Who it matters for: Parents, teachers, children's authors, educators, and anyone who wants a beautifully illustrated story in under 2 minutes. The narration quality is genuinely impressive — real cadence, not robotic text-to-speech. Full review with real storybook example →
🗓️ June 5–6, 2026 — OpenAI Overhauls ChatGPT
ChatGPT "Dreaming" — Self-Updating Memory
OpenAI's biggest change to ChatGPT personalization since the original memory launch. Dreaming (officially Dreaming V3) replaces the manually curated saved-memories list with a background synthesis process that reads across your entire conversation history and automatically rewrites what ChatGPT knows about you — without any prompting from you. The model can shift a memory from "You're going to Singapore in July" to "You went to Singapore in July 2026" after the date passes. It tracks evolving preferences, completed goals, and changing context over months of use.
The mechanism is closer to how human memory consolidates during sleep — hence the name. A background process periodically synthesises conversation history into a compressed, updatable memory state. The old system created individual memory entries you could review and delete; the new system creates a synthesised summary that is harder to audit line by line but is significantly more accurate over long time horizons.
By: OpenAI. Background memory synthesis that auto-updates what ChatGPT knows about you — no manual curation. Rolling out to Plus and Pro subscribers in the US first; Free, Go, and international users follow in coming weeks. Includes a reviewable memory summary page.
Who it matters for: Any ChatGPT user who has found the old memory system stale or wrong after a few months. If you use ChatGPT daily, Dreaming will meaningfully improve personalisation accuracy over time — particularly for evolving preferences and completed projects.
OpenAI Self-Serve Ads Manager — ChatGPT Goes Commercial
OpenAI opened a self-serve advertising platform inside ChatGPT: the ChatGPT Ads Manager. Advertisers can now create, manage, and measure ad campaigns inside ChatGPT without going through a sales team. The platform introduces cost-per-click (CPC) bidding — the same model that made Google AdWords accessible to small businesses in 2000 — alongside expanded measurement tools. Initial testing starts with free-tier users and ChatGPT Go in the US.
The privacy architecture is important to understand: conversations are not shared with advertisers. Ads are triggered by topic signals derived from the conversation, not by the conversation text itself. Whether that distinction holds under scrutiny is a legitimate question — but it is structurally different from Google's keyword targeting model, which did read query text.
By: OpenAI. Self-serve advertising platform for ChatGPT with CPC bidding and campaign analytics. Ad targeting is topic-based, not conversation-text-based. Currently in beta for US advertisers; rolling out to free and Go users as the initial ad audience.
Who it matters for: Marketers and performance advertisers who want to reach users actively seeking information — a fundamentally different context than social media scroll advertising. The CPC model means low entry cost for small businesses to test the channel.
ChatGPT Go — Global Rollout at $8/month
ChatGPT Go, originally launched in India in August 2025 as a low-cost tier, completed its worldwide rollout in early June 2026. At $8/month in the US (pricing varies by country), Go sits between free and Plus ($20/month). It delivers 10× more messages, file uploads, and image creation than the free tier, running on GPT-5.5 Instant. For users who find the free tier's rate limits frustrating but don't need the full Plus feature set, Go is the practical upgrade.
By: OpenAI. $8/month ChatGPT tier (US pricing) — 10× more messages, uploads, and image creation than free. Runs on GPT-5.5 Instant. Now available in 170+ countries. Positioned between Free and Plus for users who hit rate limits but don't need advanced features.
Who it matters for: Students, casual professionals, and anyone who regularly hits ChatGPT's free-tier limits. At $8/month it is the most accessible paid AI assistant tier currently on the market from a major lab.
🗓️ June 9, 2026 — Claude Fable 5 & Mythos 5
This is the launch the AI community has been waiting for since Project Glasswing began in April. Anthropic just dropped Claude Fable 5 — a Mythos-class model made safe for general use — and simultaneously launched Claude Mythos 5 for restricted partner access. Fable 5 is Anthropic's most capable publicly available model ever. Today, as a user of Claude, this is what I'm running on.
What is Claude Fable 5?
Fable 5 is the same underlying model as Claude Mythos — Anthropic's frontier-level architecture that has been restricted to cybersecurity researchers since April. The difference: Fable has safety classifiers layered on top that catch high-risk queries (offensive cybersecurity, biology/chemistry, and model distillation attempts) and route them to Claude Opus 4.8 instead. In less than 5% of sessions on average does any fallback occur — for 95%+ of users, every session runs at full Mythos-class capability.
The name is deliberate. Fable comes from the Latin fabula — "that which is told" — akin to the Greek mythos. Same architecture, same weights, different safeguard layer.
By: Anthropic. Mythos-class model for general use. $10/M input, $50/M output — less than half the price of Mythos Preview. Model string: claude-fable-5. Available on Claude API, AWS Bedrock, Vertex AI, and Microsoft Foundry. Free on Pro, Max, Team, and seat-based Enterprise plans through June 22 — after that, requires usage credits until capacity expands.
Key benchmarks: #1 on CursorBench (Cursor), FrontierCode (Cognition), Hebbia Finance Benchmark, ViBench (Replit), and WebDev Arena. 90%+ on complex long-running analytics (10-point jump over Opus 4.8). Stripe used it to migrate a 50M-line Ruby codebase in one day — a task estimated at 2 months for a human team.
What Fable 5 does that no previous Claude model could
Software engineering at scale. Fable 5 can run complex, multi-hour agentic coding sessions without losing context or making compounding errors. Cursor confirmed it's state-of-the-art on CursorBench. GitHub says it "took on complex, long-horizon coding tasks with a level of autonomy and reliability that exceeded previous benchmarks." Cognition's FrontierCode — which tests whether a model passes difficult coding tasks while meeting production codebase standards — is topped by Fable 5, even at medium effort.
Vision that actually works. Fable 5 rebuilt web apps from screenshots alone. It beat Pokémon FireRed using a vision-only harness — no maps, no navigation aids, just raw game screenshots. Earlier Claude models needed a complex helper harness to even attempt it.
Long-context memory. In testing with the deck-building game Slay the Spire, giving Fable 5 persistent file-based memory improved its performance three times more than the same setup gave Opus 4.8. Fable reached the game's final act three times more often. This maps directly to real-world agentic tasks: complex research projects, codebase migrations, long-form writing with evolving requirements.
Scientific research. Mythos 5 (the restricted version) matched or beat skilled human protein design operators with no human assistance — nine of fourteen drug design targets yielded strong candidates now under investigation. It also conducted a week-long autonomous genomics research project, training a model that outperformed a recent Science journal paper at 1/100th the model size.
The safety model — what Fable 5 won't do
Anthropic deliberately tuned Fable's classifiers conservatively. Three categories route to Opus 4.8 instead of Fable when triggered:
- Cybersecurity: Offensive cyber tasks, exploit development, and agentic hacking. Fable scored zero on harmful single-turn cyber requests even against 30 public jailbreak techniques.
- Biology and chemistry: Broad fallback for now, narrowing as classifiers improve. A trusted access program for biomedical researchers is opening in coming weeks.
- Model distillation: Detects and blocks large-scale attempts to extract Fable's capabilities to train competing models.
These safeguards also introduce a new 30-day data retention policy for Mythos-class models — all traffic is retained for 30 days for safety monitoring, then deleted. Anthropic cannot use it for training.
What this means for you right now: If you're on Claude Pro, Max, Team, or Enterprise — you have Fable 5 access at no extra cost until June 22. Use it for your hardest tasks this fortnight. After June 23, Fable 5 requires usage credits until capacity expands enough for Anthropic to include it in standard plans again.
🗓️ June 7–9, 2026 — Open Models & Developer Alerts
NVIDIA Nemotron 3 Ultra — Best US Open-Weight Model
NVIDIA shipped Nemotron 3 Ultra — a 550B parameter Mixture-of-Experts (MoE) model with 55B active parameters — to Hugging Face, NIM, and OpenRouter. It is currently the most capable open-weight model produced by a US company, running at over 300 tokens per second on current NVIDIA hardware. The architecture uses hybrid Mamba-Transformer layers for efficient long-context handling, NVFP4 quantisation for cross-architecture GPU deployment, and multi-token prediction for faster multi-turn generation.
The design target is explicitly long-running agent workflows — tasks that require planning, code execution, tool use, and multi-step reasoning over extended sessions. Within 48 hours of release it was already integrated into Perplexity, Nous Research, OpenCode, and atomic.chat. SWEBench Verified scores sit between 65–70% depending on agent framework, consistent across deployments.
By: NVIDIA. 550B MoE model (55B active params), open weights. Available on Hugging Face, NIM, and OpenRouter. Hybrid Mamba-Transformer architecture, NVFP4 quantisation for high-throughput deployment, designed for long-running agent workflows. 300+ tokens/sec on current NVIDIA hardware.
Who it matters for: Teams running self-hosted or private AI deployments who need frontier-level agent capability without sending data to a cloud API. Best US open model for coding agents, long-context reasoning, and enterprise deployments with data-residency requirements.
OpenCode — The Open-Source Coding Agent That Took #1
OpenCode crossed 160K GitHub stars and 7.5M monthly active users this week, making it the most widely adopted open-source coding agent. What separates it from GitHub Copilot and Cursor: model-agnostic access to 75+ providers via a single configuration file, full Language Server Protocol (LSP) integration for precise code intelligence, air-gapped deployment support for regulated environments, and MIT licensing that allows commercial use without royalties. You bring your own API keys; OpenCode handles the agentic scaffolding.
By: OpenCode community. Open-source coding agent with access to 75+ AI providers. LSP integration, air-gapped deployment, MIT license. 160K+ GitHub stars, 7.5M MAU. Works with Claude, Gemini, GPT-5.5, Nemotron 3 Ultra, or any model you configure.
Who it matters for: Developers who want coding agent capability without vendor lock-in, or teams that can't send code to proprietary APIs. The model-agnostic design means you can run it against whichever model has the best performance-per-dollar for your specific codebase.
Anthropic is retiring claude-sonnet-4-20250514 and claude-opus-4-20250514 on June 15, 2026. API calls using these exact model strings will return errors after that date. Migrate to claude-sonnet-4-6 (Sonnet lane) or claude-opus-4-8 (Opus lane) before the deadline. Claude Sonnet 4.6 now outperforms the original Opus 4.0 on coding and instruction-following at a lower price point — for most tasks it is the correct upgrade target.
Qwen 3.7 Max — The Price-Performance Challenger
Alibaba's Qwen 3.7 Max debuted at #4 on the WebDev Arena leaderboard this week — ahead of Claude Opus 4.6 — at $2.50/$7.50 per million tokens (input/output). For web development and frontend code generation specifically, it is currently the best value model available. Its context window and instruction-following accuracy on structured data tasks is competitive with models costing 3–5× more. Not a general-purpose frontier model, but for the specific workloads it excels at, the price gap is difficult to ignore.
By: Alibaba Cloud. #4 on WebDev Arena leaderboard. $2.50/M input, $7.50/M output — less than half the cost of comparable Anthropic or OpenAI models. Strong on web development, frontend code generation, and structured output tasks. Available via Alibaba Cloud and third-party API aggregators.
Who it matters for: Developers building or evaluating AI coding pipelines who want to benchmark cost-performance tradeoffs. Worth testing against your specific workloads before committing to a higher-cost model for production.
🗓️ June 10–14 — MiniMax M3, ZoomMate & Gemini 3.5 Pro Preview
MiniMax M3 — Open-Weight Frontier Model That Beats GPT-5.5 on Code
Shanghai-based MiniMax launched M3 on June 1 and it immediately became the most significant open-weight model drop of mid-2026. The headline: 59.0% on SWEBench Pro — higher than both OpenAI GPT-5.5 and Google Gemini 3.1 Pro on the most rigorous software engineering benchmark. The architecture behind this is MSA (MiniMax Sparse Attention), a new sparse attention mechanism that cuts per-token compute at 1M context to one-twentieth of the prior generation, with 9× faster prefill and 15× faster decoding. It's the first open-weight model to combine frontier-level coding agent performance, a 1M-token context window, and native multimodal computer use — including desktop operation — in a single model. Weights are open; API is live now on MiniMax's platform.
By: MiniMax (Shanghai). 1M-token context, MSA sparse attention architecture — 9× faster prefill, 15× faster decoding vs. prior gen. Native multimodal: text, image, video, and desktop computer use in a single model. 59.0% SWEBench Pro — beats GPT-5.5 and Gemini 3.1 Pro. Available via MiniMax API and MiniMax Code. Weights and technical report releasing within 10 days of launch.
Who it matters for: Teams running self-hosted coding agent pipelines who want frontier-level performance without a proprietary API. The 1M context + desktop computer use combination is unique at open-weight level — nothing else can both browse the web and operate your desktop in the same inference call.
ZoomMate — AI That Finishes the Work Your Meetings Start
Zoom launched ZoomMate at $20/user/month for North America — a genuinely agentic AI that operates inside your live meetings, not just after them. The key distinction from every other meeting AI tool: ZoomMate monitors the conversation in real time, identifies action items and commitments as they are made, and immediately initiates downstream workflows in Salesforce, ServiceNow, Workday, Jira, Google Workspace, Microsoft 365, and Zendesk — without requiring anyone to manually log what was discussed. It also generates presentations, documents, spreadsheets, and project plans directly from meeting transcripts combined with connected enterprise data.
The $10/month AI Productivity Suite (meeting summaries, whiteboard AI, AI document creation) is bundled into the ZoomMate subscription, so the effective combined price is $20 total. EMEA and APAC rollout is planned later in 2026.
By: Zoom. Agentic AI that monitors live meetings and initiates downstream actions without manual follow-up. Connects to Salesforce, ServiceNow, Workday, Jira, Google Workspace, Microsoft 365, Zendesk. Generates deliverables (decks, docs, project plans) from meeting context. Includes AI Productivity Suite ($10/mo value). North America GA; EMEA/APAC later 2026.
Who it matters for: Sales teams, customer success, and project managers who spend 30–60 minutes after each meeting manually logging CRM updates and creating follow-up documents. ZoomMate collapses that into zero manual work — the Salesforce opportunity updates while you're still in the call.
Gemini 3.5 Pro — In Limited Preview, GA Before June 30
Google's Gemini 3.5 Pro is in limited Vertex AI preview and targeting GA before June 30, 2026. Announced at Google I/O May 19 — "give us until next month to get it to you" (Sundar Pichai, on stage) — it's the model that absorbs the use cases Gemini Ultra used to handle: frontier reasoning, deep multimodal understanding, and the largest context window Google offers. Key specs: 2 million token context window (the largest in the Gemini 3.5 family), Deep Think reasoning (extended thinking mode for hard problems), and native multimodal handling of text, image, audio, and video in a single request. Expected pricing is approximately $15/M input and $60/M output, following the Flash-to-Pro ratio of prior Gemini generations — not yet confirmed at GA.
By: Google. 2M-token context window. Deep Think reasoning mode for complex tasks. Native multimodal: text + image + audio + video in one request. Replaces Gemini Ultra use cases. In limited Vertex AI preview now; GA expected before June 30. Pricing unconfirmed — likely ~$15/$60 per M tokens input/output.
Who it matters for: Enterprise teams that were using Gemini Ultra for frontier tasks and want the upgraded 3.5-generation reasoning at potentially similar or lower cost. The 2M-token window opens use cases like full codebase analysis, large document review, and long-running research that Flash's 1M context can't handle.
🗓️ June 15–21 — Fable 5 Suspended, Siri AI Betas, Google Agents
Claude Fable 5 & Mythos 5 — Suspended by US Government
On June 12, the US Commerce Department issued an export control directive ordering Anthropic to suspend all access to Fable 5 and Mythos 5 for any foreign national — inside or outside the US, including foreign national Anthropic employees. Because Anthropic cannot gate access by nationality in real time, it disabled both models for all users globally. Claude 3.5, Opus 4.6, Sonnet 4.6, and all other Claude models remain fully accessible.
The stated reason: a jailbreak technique was demonstrated to the government that allegedly bypasses Fable 5's safety classifiers for cybersecurity tasks. Anthropic publicly disputed this characterisation, noting the same jailbreak works on GPT-5.5 Instant — which faces no similar restrictions. On June 18, a bipartisan group of House members sent a letter to the Trump administration demanding explanation for the selective enforcement and asking whether competing models would face equivalent treatment.
Apple rebuilt Siri from the ground up — officially renamed Siri AI — now powered by Google's Gemini models through a multi-year deal reportedly worth ~$1B/year. With 1.5 billion daily Siri requests, Gemini is now the reasoning backbone behind the world's most widely used voice assistant.
Key new abilities: full on-screen awareness (Siri reads what you're looking at in real time), cross-app actions in a single command, device-wide personal context (emails, messages, photos), and an expanded standalone Siri app. A three-tier privacy architecture routes simple tasks on-device, moderate ones to Apple's Private Cloud Compute, and only heavy reasoning to Google Cloud.
Ships with iOS 27 / macOS 27 / visionOS 27. Developer betas dropped June 9 — public beta arrives July 2026, full release this fall with new iPhone hardware. Not available at launch: EU (regulatory) and China.
Google is beginning to roll out Search Agents — AI agents that run in the background 24/7 monitoring topics you care about: concert tickets, property listings, product price drops, sports scores, news on specific topics. You create and manage multiple agents directly inside Google Search. When something relevant changes, you get a notification.
Shopping agents are rolling out alongside: track a product's price history, set a target price, and have the agent watch for deals and notify you. Booking agents can surface live availability and direct booking links. All agents run on Gemini 3.5 Flash. Rollout starts with Google AI Pro and Ultra subscribers this summer — standard users follow.
Why it matters: This marks Search moving from query-response to persistent background intelligence. You stop checking — the agent checks for you.
⚠️ Developer Alerts — Action Required This Week
OpenAI is retiring GPT-4.5 from ChatGPT and the API on June 27. Any API call referencing gpt-4.5 will fail after this date. Migrate to gpt-5.5-instant or chat-latest before June 27.
If you're on Claude Pro, Max, Team, or Enterprise — today is your last day to run Fable 5 at no extra cost. After June 22, it moves to usage credits. The model remains available — it just costs. Bank your usage today. API model string: claude-fable-5.
OpenAI announced June 11 that o3 will be removed from ChatGPT on August 26 (90-day sunset). API snapshots deprecate December 11, 2026. Not urgent today but plan your migration to o4-mini or GPT-5.5 Instant for reasoning pipelines currently on o3.
Gemini 3.5 Pro (2M-token context, Deep Think reasoning, multimodal flagship) remains in limited Vertex AI preview as of June 21. Google confirmed general availability before June 30 — but the same was said in mid-June. Pricing unconfirmed (~$15/$60 per M tokens). Watch this space — if it drops before June 30, it changes the frontier comparison entirely.
🗓️ June 22–30 — Month End: Sonnet 5, LongCat & GPT-4.5 Gone
Anthropic launched Claude Sonnet 5 on June 30 — now the default model on Claude.ai. It achieves near-Opus 4.8 performance on most tasks at dramatically lower pricing. Benchmarks: 63.2% SWE-bench Pro (vs Fable 5's 95%), 80.4% Terminal-Bench 2.1 (beating Opus 4.8's 74.6%), 84.7% BrowseComp agentic search. Most agentic Sonnet yet — can plan, use tools (browser, terminal), and run autonomously at a level that previously needed much larger models.
Also went GA on GitHub Copilot on June 30. 1M token context window. 128K max output. Intro pricing through August 31.
Claude Fable 5 Export Controls Lifted — Restored July 1
The US government lifted the June 12 export control order on June 30 after Anthropic trained an improved safety classifier blocking the jailbreak technique in over 99% of cases. Fable 5 and Mythos 5 came back online globally July 1, after a 19-day suspension. Both now carry updated cybersecurity safeguards developed in collaboration with the government. See full details in the July 2026 tracker →
As of June 26, GPT-4.5 is no longer available in ChatGPT or via API. All existing conversations migrate automatically to GPT-5.5 Instant. If you still have gpt-4.5 API calls, they are now failing — migrate to gpt-5.5-instant immediately.
OpenAI previewed the GPT-5.6 family (three models: Sol, Terra, Luna) on June 26, gated to roughly 20 trusted partner organisations. Stronger coding, scientific reasoning, long-horizon planning, and agentic workflows. No public availability date confirmed — wider API and ChatGPT access "soon." Not yet on the open waitlist.
Chinese food-delivery giant Meituan open-sourced LongCat-2.0 — a 1.6 trillion parameter MoE model trained entirely on a 50,000-card cluster of domestic Chinese ASICs (no NVIDIA hardware). 48B parameters active per token, 1M context window, agentic coding focus. Geopolitically significant: China's largest model trained without restricted chips, beating the export control blockade at model-training level.
Benchmarks: 59.5% SWE-bench Pro (narrowly ahead of GPT-5.5's 58.6%), 70.8% Terminal-Bench. Leads OpenRouter's coding leaderboard. Weights available on Hugging Face.
🎯 June Verdict — What to Actually Try This Month
June 2026 is not a month of incremental updates. Three things stand out as genuinely worth your time:
1. Comet browser (free) — Download it regardless of whether you use Perplexity for search. The Auto Browse and real-time synthesis features are meaningfully different from any browser extension. The free tier is not crippled.
2. Gemini 3.5 Flash via API — If you are building agentic systems and haven't benchmarked Flash against your current stack, do it this month. At $0.10/M input with Managed Agents support baked in, it has the best price-to-capability ratio for production agentic workflows of any model currently available.
3. Claude Fable 5 — ⚠️ Currently suspended — Fable 5 launched June 9 but was suspended June 12 following a US government export control directive. All other Claude models (3.5, Opus 4.6, Sonnet 4.6) remain fully operational. See the full story on the suspension →
4. Google Flow (free to test on Pro) — If you have a Google AI Pro subscription, spend 20 minutes with Flow before writing it off as "just another AI video tool." The Camera Controls panel and native audio are meaningfully different from Kling, Runway, or Sora. The two-minute learning curve is worth it. See full guide with real output videos →
5. Gemini Storybook (completely free) — No excuses not to try this. Open Gemini, type any story idea, pick Watercolour. You'll have a 10-page illustrated book with narration in 90 seconds. Best creative freebie Google has shipped in years.