AI Product Updates Daily — June 12, 2026
Anthropic reverses a secret Claude Fable 5 policy that silently degraded outputs for researchers, and apologizes. Google Gemini suffers one of its biggest Workspace outages in history. Samsung lifts its 2023 AI ban and deploys ChatGPT, Gemini, and Claude companywide. GitHub's usage-based billing shift sparks backlash, Google releases DiffusionGemma at 700+ tokens/sec, and Anthropic pledges $200M to research AI's economic impact.
Today's sweep covers Anthropic reversing a secret Claude Fable 5 policy after backlash, Gemini suffering one of its biggest-ever Workspace outages, Samsung lifting its three-year AI ban and deploying ChatGPT / Claude / Gemini companywide, GitHub's usage-based billing shake-up taking full effect, Google releasing DiffusionGemma, and a cluster of enterprise AI integrations from Smartsheet, 8x8, and Visa.
Anthropic walks back Fable 5's hidden research-sabotage safeguard
The rollout of Claude Fable 5 hit a wall of criticism within days. Anthropic had quietly baked in a safeguard that degraded the model's performance for users it suspected of building competing AI systems — without telling those users anything was wrong. The company explicitly bans using Claude to train rival models in its terms of service, but researchers argued that silently downgrading outputs crossed a different line entirely.1
The pushback was sharp. Dean Ball, a former White House AI adviser now at the Foundation for American Innovation, called it "shockingly hostile" in a post on X, arguing that secret performance degradation undermines anyone trying to do AI safety research. Will Brown, research lead at open-source AI startup Prime Intellect, framed it more bluntly: "It felt like Anthropic was saying, 'We don't trust anybody else to do AI research. We are the only ones who have to do AI research.' It feels a bit like they're starting to pull the ladder up behind them."1
Anthropic reversed course on June 11, saying it is changing Fable 5's safeguards for frontier LLM development to make them visible. "We made the wrong trade-off and we apologize for not getting the balance right," the company said in a statement to WIRED. The visible version casts a wider net — more benign requests may now trip the filter — but users will at least be told when they are being rerouted to a less capable model.1
Separately, Claude Haiku 4.5 had a brief service disruption on June 11, resolved by 10:21 a.m. PT, according to Anthropic's status page.2
Gemini goes down for most of the workday
Google's Gemini had what Gemini App VP Josh Woodward called, in a post on X, one of its biggest outages yet. Starting early Wednesday morning, Workspace users across macOS, web, iOS, Android, and Gemini in Chrome started seeing error codes 1099 and 1076 — "Something went wrong" — across Drive, Docs, Sheets, and Slides. DownDetector logged over 1,600 reports at peak.2
Google confirmed resolution at 6 p.m. PT. Its postmortem: "From preliminary analysis, the issue was triggered by a performance issue in our backend database which impacted the retrieval of Gemini App tools catalog." Engineers fixed it by redistributing load across the database.2
The timing is awkward. Google spent much of Google I/O in May promoting Gemini's deep integration into Workspace as a productivity argument. An all-day outage that blocks Workspace AI entirely is the kind of failure that sticks in enterprise buyers' minds.
Samsung lifts its AI ban, deploys ChatGPT, Gemini, and Claude companywide
Three years after a data-leak incident led Samsung to ban its own employees from using external generative AI, the company reversed that policy on June 11. Samsung Electronics' Device eXperience (DX) division is now authorized to use ChatGPT (OpenAI), Gemini Enterprise (Google), and Claude (Anthropic) across its operations.3
The rollout follows a proof-of-concept from April through May that involved about 2,500 DX employees testing all three services. All tools are gated behind an internal security control layer that uses identity-based access and data-loss-prevention controls to inspect prompts before they reach external models. DX co-CEO Roh Tae-moon said the move "marks the starting point for fundamentally transforming the way we work and execute." Samsung's in-house model Samsung Gauss will continue running alongside the external services.3
Samsung is also running an "AX Boot Camp" this month for about 50 affiliate presidents and plans AI training for roughly 2,300 executives across affiliates through August 12, with a target of reaching all employees by year end. OpenAI CEO Sam Altman is scheduled to visit Samsung headquarters in Suwon next Monday to brief DX Division employees, with the Korea Times noting the visit will also cover "next phase" Stargate project discussions.3
GitHub Copilot's new billing model lands — and developers hate it
GitHub Copilot switched from flat-rate subscriptions to usage-based billing on June 1, and the fallout is still generating forum posts. The headline numbers did not change: Copilot Pro is still $10/month, Pro+ still $39/month. What changed is that premium model interactions now draw down a monthly credit pool, and any usage beyond that pool is billed at $0.04 per additional premium request.4
The plan structure now looks like this:
| Plan | Monthly price | Included premium requests |
|---|---|---|
| Copilot Free | $0 | 50/month |
| Copilot Pro | $10 | 300/month |
| Copilot Pro+ | $39 | 1,500/month |
| Copilot Business | $19/user | 300/user/month |
| Copilot Enterprise | $39/seat | 1,000/seat/month |
The part generating the most anger: Anthropic's Opus-family models have been removed from the base Pro tier (restricted to Pro+ and above), and token-based billing means agentic workflows — which fan out into many model calls — can cost dramatically more than before. One developer posted that a month of activity that previously cost the $39 Pro+ flat fee would have run to roughly CAD $1,159 (about USD $838) under the new meter. That person is a power user with unusually heavy agentic workloads, but the sticker shock hit the community's loudest and most technically sophisticated segment, which is why the backlash spread quickly.4
Code completions and Next Edit Suggestions remain unmetered. Copilot code review now also burns GitHub Actions minutes in addition to AI Credits — a second meter teams running automated PR reviews need to factor in.4
Google releases DiffusionGemma: parallel text generation with 700+ tokens/sec
Google released the developer guide and open weights for DiffusionGemma, an experimental model built on the Gemma 4 26B MoE backbone that generates text using diffusion rather than token-by-token autoregression.5
Instead of predicting one token at a time, DiffusionGemma works on a 256-token "canvas," iteratively denoising the whole block in parallel. The model uses bidirectional attention during the denoising step, which lets every token position see every other position simultaneously — a structural advantage for tasks like constraint-solving where a solution in one cell depends on constraints in another. Google demonstrated this with a Sudoku solver fine-tuned from the base model: the fine-tuned version solved 80% of puzzles (up from near 0% for the base model) and did so in 12 denoising steps rather than the 48-step maximum, with adaptive early stopping.5
Speed numbers: up to 700+ tokens/sec on an NVIDIA RTX 5090 and 1,000+ tokens/sec on a single H100. With quantization, the model fits within 18GB VRAM, making it deployable on consumer cards. The model is available under Apache 2.0 on Hugging Face, works with vLLM and SGLang out of the box, and can be deployed via Google Cloud Model Garden or NVIDIA NIM.6
Worth noting: the speed advantage is strongest at low batch sizes on a single accelerator — the parallel decoding approach gives diminishing returns under high-QPS cloud workloads where traditional autoregressive models can batch thousands of requests efficiently.
Anthropic pledges $200M for AI economic research, Dario Amodei raises UBI scenario
Anthropic announced a $200M initial commitment to a new Economic Future Research Fund, alongside a $150M national scholarship program to help early-career professionals bring AI benefits to communities across the US.7
CEO Dario Amodei, writing in an essay published Wednesday, argued that AI's disruption to the labor market will be larger and longer-lasting than previous technology transitions. His framework for response scales with severity: improved unemployment data collection and pro-employment policies for modest disruption; universal basic income-style mechanisms if AI produces a permanent long-term decline in demand for human labor. He said UBI could be funded through taxes on AI companies or capital gains increases.7
The announcement comes as Anthropic and OpenAI are both pursuing IPOs, and competitors are making similar public commitments to ensuring AI benefits are broadly shared.
Smartsheet adds ChatGPT, Microsoft Copilot, and Gemini to its MCP server
Smartsheet expanded its MCP (Model Context Protocol) Server to support ChatGPT, Microsoft Copilot, and Google Cloud Gemini Enterprise, joining Claude — which launched as the first supported model in March.8
Since the March launch, the Smartsheet MCP Server has seen 22,000+ unique users, 3 million AI actions, and weekly active user growth from under 1,000 to more than 9,000. Tool call volume jumped from 42,000 to over 700,000 per week. The company also announced Smart Assist, an AI companion built natively into the Smartsheet interface that draws on the same live work data as the external MCP connections.
Connections to Copilot and ChatGPT are available to all US customers today; EMEA and APJ rollout follows. Gemini Enterprise and Claude connections are available globally.8
8x8 adds real-time voice translation across 13 languages to AI Studio
8x8's AI Studio platform now supports live simultaneous voice translation during customer support calls, covering 13 languages. When a caller speaks in French, Spanish, or Japanese, the agent hears a real-time AI-generated translation of the customer's voice — no interpreter, no transfer, no separate call.9
The feature is in early availability. It builds on recent transcription improvements tied to OpenAI's GPT-Realtime-2. 8x8 AI Studio also now lets organizations select between Claude, Gemini, Grok, and ChatGPT as the underlying model per agent, without any platform change. One-click connectors cover HubSpot, Slack, Stripe, Atlassian, Twilio, GitHub, Asana, Figma, Intercom, Dropbox, and ClickUp.9
Visa embeds its payment network in ChatGPT for agentic shopping
Announced June 10 at Visa's Payments Forum in San Francisco, Visa is embedding its payment network directly into ChatGPT, allowing AI agents to browse, select, and complete purchases on behalf of users across any merchant that accepts Visa — not just a curated list of partners.10
The architecture: OpenAI handles the agent interaction and purchase decision; Visa handles payment authorization, fraud monitoring, and transaction security. Users set spending limits and mandatory approval steps; most initial transactions require manual confirmation, with the option to extend more autonomous authority over time. Visa modified its token framework and data-capture processes to cover the intermediate steps in AI-mediated transactions.10
The setup replaces ChatGPT's earlier Instant Checkout, which was limited to a small number of partner merchants and drew complaints about high error rates and merchant fees.
Quick notes
- Oracle Q4 FY26: Revenue of $19.2B (+21% YoY), cloud infrastructure up 93% to $5.8B. FY27 guidance calls for $90B revenue. Shares fell after-hours as investors weighed the announced $90B CapEx plan for AI data center buildout. 11
- Anthropic "When AI builds itself" essay: Updated guidance published on Anthropic's Institute page as of June 11, expanding on the recursive self-improvement data points the company has been sharing. 12
- German court rules against Google AI Overviews: An Arstechnica report published June 11 covers a German court ruling finding that AI-generated search overviews are not necessary to internet search, potentially affecting AI search integration across the EU. 13
References
- 1
- 2
- 3Samsung Adopts ChatGPT, Gemini, Claude Companywide
letsdatascience.com
- 4GitHub Copilot Pricing 2026: From $0 to $39/mo
tech-insider.org
- 5DiffusionGemma: The Developer Guide
developers.googleblog.com
- 6DiffusionGemma model overview
ai.google.dev
- 7
- 8
- 98x8 AI Studio Delivers Wave of New Capabilities
businesswire.com
- 10
- 11Oracle Q4 FY26 Financial Analysis
linkedin.com
- 12When AI builds itself
anthropic.com
- 13Nobody needs AI to search the Internet, court says
arstechnica.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.