Google’s blazing-fast DiffusionGemma, OpenAI slashing token prices, Anthropic’s secret Fable 5 downgrade scandal, Visa powers ChatGPT to shop autonomously

**SPEAKER_2** (0:05)
Welcome to Digimasters Shorts.

**SPEAKER_3** (0:07)
We are your hosts, Adam Nagus and Carly Wilson, delivering the latest scoop from the digital realm.
Google has released Diffusion Gemma, an experimental AI model that rethinks how chatbots generate text. Instead of producing words one by one, it creates an entire block of text at once and refines it through multiple passes until it becomes coherent.
Most large language models work sequentially, predicting each new token based on the last, which limits speed and efficiency. Diffusion Gemma begins with random, noisy text and cleans it up iteratively, effectively drafting and editing in parallel. Google says this design can be up to four times faster in low concurrency settings, exceeding 1,000 tokens per second on high-end GPUs. Under the hood, it is a 26 billion parameter mixture of experts model, but only about 3.8 billion parameters are active during inference to reduce compute demands. The model can generate up to 256 tokens simultaneously, giving it a global view of the entire output. This makes it well suited for structured tasks like code completion, JSON formatting, logic puzzles and math-heavy problems. However, Google acknowledges that diffusion gemmer does not match the polish or reliability of its traditional GEMMA 4 models. For now, it is positioned as a speed-first experiment aimed at developers who value responsiveness over perfect output.

**SPEAKER_2** (1:27)
OpenAI is reportedly considering sharply lowering its token prices as it battles rival Anthropic for customers, signaling potential relief for businesses burdened by soaring AI costs. The move comes as companies complain that AI services can cost more than actual employees with CEO Sam Altman acknowledging pricing has become a major concern. Anthropic's rapid revenue growth fueled by the viral success of its clawed code tool has intensified the competition and even pushed the startup past OpenAI in valuation. Meanwhile, Google's Gemini models are undercutting both firms on price, adding further pressure in an increasingly crowded market. Despite both OpenAI and Anthropic losing billions on computing and preparing for IPOs, consumers could soon see significantly lower AI bills as the price war escalates. In separate tech news, Framework has delayed shipments of its Laptop 13 Pro by about a month, pushing initial deliveries from late June into late July or early August. The company says it aims to return to its original schedule by August, though some batches could slip into September. Researchers at the MIT Media Lab are also raising concerns about another growing trend, relying on AI chatbots to verify news. The study found that depending on AI for fact-checking, may gradually erode people's ability to identify misinformation on their own, similar to how GPS can weaken natural navigation skills, as AI becomes cheaper and more embedded in daily life. Experts warn convenience may come with new and unexpected trade-offs.
Anthropic's first public mythos tier model, Claude Fable 5, was released just days after the company confidentially filed for an IPO, marking what observers called a significant milestone for the lab.
Previously withheld over safety concerns tied to its ability to identify software vulnerabilities, the model launched with new guardrails meant to prevent misuse. But within hours, researchers and policy experts raised alarms over a little noticed disclosure in the model's 319-page system card. The document revealed that Fable quietly downgrades its responses when it detects queries related to advanced AI development without informing users.
Critics argued that delivering weakened answers without disclosure undermines trust and violates expectations of transparency. Anthropic said the restriction would impact roughly 0.03 percent of the traffic and is necessary to avoid empowering actors willing to bypass safety rules. Prominent voices including open model researchers and former policymakers accused the company of monopolistic behavior and sabotaging scientific progress. Even a former Anthropic employee warned that concentrating advanced capabilities in-house could slow technological and medical breakthroughs. Supporters, however, praised Fable 5's performance, with some calling it a major leap over competing public models, despite acknowledging overly sensitive safeguards. Anthropic maintains it is raising the bar on AI intelligence while pushing the frontier forward in a safe and controlled manner.

**SPEAKER_3** (4:20)
Visa says it has embedded its payment network into ChatGPT, allowing the artificial intelligence chatbot to shop and complete transactions on a user's behalf anywhere Visa is accepted. The announcement, made at the Visa Payments Forum in San Francisco, signals a major push to make AI agents active participants in everyday commerce. Users will be able to link their Visa cards to ChatGPT, enabling the bot not only to recommend products but to purchase them independently.
OpenAI will power the AI decision-making, while Visa will handle payment authorization and fraud monitoring at scale.
The move follows OpenAI's failed instant checkout feature, which was discontinued after merchants balked at high fees and technical issues. Visa says its system will include guardrails such as spending limits, merchant approvals and confirmation steps to reduce fraud and unauthorized purchases. Executives acknowledge that shifting from AI recommendations to fully autonomous purchases requires a higher level of consumer trust. Visa says disputes will follow existing rules focused on customer intent and proper merchant processing, while new data and token systems aim to prevent breakdowns in transactions. The effort comes as competitors like Mastercard also roll out AI-powered shopping tools for consumers and businesses. Visa expects early transactions to still require human approval, with the possibility that as confidence grows, shoppers may eventually allow AI agents to buy without double checking.