Get tomorrow's brief in your inbox
Today: Anthropic launched Claude Fable 5 at double the price of Opus 4.8, but with hidden guardrails that secretly limit its help on certain AI development tasks. Google slashed its AI Plus subscription from $7.99 to $4.99, firing the first shot in a U.S. price war that could force smaller AI providers to compete on cost. JPMorgan plans to deploy long-running autonomous AI agents later this year that can work for hours without human intervention.
Anthropic just released Claude Fable 5, the first public version of its most powerful Mythos model, but there's a catch most users won't see coming. The model costs $10 per million input tokens and $50 per million output tokens (double the price of Opus 4.8), and it includes safety guardrails that block responses in areas like cybersecurity and chemistry. You'll see those blocks and get a fallback to Opus 4.8.
But Anthropic also built in silent interventions that you'll never know about. According to the 319-page system card, Fable 5 will quietly degrade its performance on requests related to frontier AI development like "building pretraining pipelines, distributed training infrastructure, or ML accelerator design." Unlike the visible safety blocks, these interventions don't announce themselves. The model just gives you worse answers. Source
Anthropic justifies this as preventing "recursive self-improvement" and says it affects only 0.03% of traffic in fewer than 0.1% of organizations. But for small businesses building AI tools or exploring ML infrastructure, this means you can't fully trust the model's output. If you're working on anything that touches ML optimization or accelerator design, you might be getting silently sabotaged answers. This is the first time any AI company has announced these kinds of invisible limitations. If you need uncensored AI for technical work, Anthropic is also shipping Claude Mythos 5 (same capabilities, no safety classifiers) to approved organizations. For most SMBs, though, Fable 5's performance on non-AI tasks is genuinely impressive. Early tests show it can build full games from a single prompt and work autonomously for up to a dozen hours on complex specifications. Just know what you're paying for. Source
Google AI Plus drops to $4.99/month with double the storage. Google just cut its budget AI subscription from $7.99 to $4.99 and doubled storage from 200GB to 400GB. You get video generation (Omni Flash), the creative studio Google Flow, and NotebookLM for research. This is the opening shot in an AI price war that started in emerging markets (India has had sub-$5 AI plans since last year) and just landed in the U.S. If you're paying $20/month for ChatGPT Plus or Claude Pro and only use basic features, this is worth a look. The catch: Google's integrated experience means you're deeper into their ecosystem. The upside: you're not spending $240/year for AI you use twice a week. Source
TidyCal offers lifetime scheduling for $29. If you're tired of paying Calendly $144/year, TidyCal is a one-time $29 payment for lifetime scheduling, paid bookings, subscriptions, and a client reviews system. The Agency plan ($79 one-time per user) adds SMS reminders, team booking, and round robin scheduling. Trade-offs: TidyCal branding stays on your booking page (mobile is especially visible), and support is slower. But if you're a solo consultant or small team, saving $120+ per year might be worth it. Calendly is still better for advanced automation, AI notetaking, and analytics, but TidyCal covers 90% of what most small businesses actually need. Source
Apple's Shortcuts app now understands natural language prompts. With iPadOS 26, you can type "Send a text to Anna with three kissy emojis" and Apple Shortcuts will build the workflow for you. This is Apple's take on "vibe coding" for your phone. The first beta is buggy (complex multi-step shortcuts mostly fail), but the concept is solid: tell your phone how you want it to work, and it figures out the steps. For small businesses, this could eventually replace custom automation apps. Right now, it's best for simple tasks like texting, toggling Do Not Disturb, or taking photos. Third-party app support is still unclear. If it works as promised, this is a rare AI feature that actually solves a real problem instead of generating more work. Source
JPMorgan is deploying long-running AI agents in 2026. The bank's chief analytics officer told CNBC that JPMorgan will roll out AI agents that can work autonomously for "an hour or two" later this year, marking a major shift from agents that complete single tasks in minutes. Derek Waldron calls this "intellectual coherence," where models act more like team managers than individual workers, parsing problems and delegating activities across multiple systems. The bank is already seeing a 20% increase in gross sales in private banking because AI systems screen market activity and research overnight, freeing bankers to focus on client interactions. For small businesses, the takeaway is this: AI agents are moving from "complete this one task" to "manage this workflow across multiple tools for hours." If JPMorgan (with a $20 billion annual tech budget and intense security concerns) is comfortable deploying these agents, the technology is close to ready for smaller operations. Start thinking about workflows that could benefit from a digital manager that works while you sleep. Source
LSEG cut product release cycles from 6 months to 2 weeks with ChatGPT Enterprise. London Stock Exchange Group deployed ChatGPT Enterprise and OpenAI APIs across thousands of employees globally, enabling analysts to summarize financial data, product teams to prototype features, and business teams to generate client communications. The result: customer delivery timelines dropped to 4 weeks from request to production, and product release cycles fell from 3 to 6 months down to 2 weeks. LSEG's lesson for small businesses: give people AI tools to move faster, but embed governance from the outset (model evaluation, human review for critical outputs, strict data privacy controls). Don't restrict people; enable them with guardrails. For SMBs, this means you don't need a massive AI budget to see results. Start with one workflow (report generation, customer communications, research synthesis) and measure time saved. Source
Nextdoor engineers use OpenAI Codex to build products end-to-end. Engineers at Nextdoor are no longer locked into specialist roles. With Codex, one engineer built a full map-based feature across mobile, frontend, and backend that historically would have required three teams. Cory Dolphin, Head of Engineering, says productivity has accelerated so much that the bottleneck is no longer engineering but strategy: what to build next. For small businesses, this is a glimpse of the future. If you've been putting off features because you'd need a full dev team, AI coding agents like Codex (or open-weight alternatives like Claude Code) are making it possible for one technical person to ship complex features. The limiting factor is shifting from "how do we build this" to "what should we build." Source
Claude Fable 5 pricing: $10 input / $50 output per million tokens. That's double the price of Opus 4.8. Through June 22, Fable 5 is included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost. Starting June 23, Anthropic pulls it from those plans and requires usage credits. The company says it'll restore Fable 5 as a standard subscription feature "as soon as possible." Translation: Anthropic is testing whether users will pay premium prices for their most powerful model. If you're on a Pro plan, use Fable 5 heavily between now and June 22 to decide if it's worth the upgrade cost. For most SMBs, Opus 4.8 ($5 input / $25 output) is still the better value unless you need Fable's long-running autonomous capabilities. Source
Workflow orchestration tools are commoditizing fast. Make has 3,000+ integrations; Zapier has 9,000+. Both support webhooks, custom APIs, and built-in modules. The gap matters when you're connecting oddball tools. Zapier optimizes for breadth and speed to a working automation. Make optimizes for visual flexibility and granular control. For small businesses, the lesson is this: start with Zapier if you want automations running today. Switch to Make if you need detailed control over data transformation and error handling. Don't overbuild. Most small businesses need simple trigger-action automations, not complex orchestration. Save the advanced stuff for when you've proven the workflow actually saves time. Source
It's not FAANG anymore. It's MANGOS. With SpaceX, Anthropic, and OpenAI all racing toward IPOs, the tech industry is replacing the FAANG acronym (Facebook, Amazon, Apple, Netflix, Google) with MANGOS: Meta, Anthropic, Nvidia, Google, OpenAI, SpaceX. The shift reflects where the money and power are moving. For small businesses, this matters because it shows which platforms will dominate the next decade. If your business strategy depends on AI tools, you're increasingly dependent on MANGOS companies (especially Anthropic, OpenAI, and Google for model access). Diversify your AI vendors where possible so you're not locked into one provider's pricing and terms. Source
Meta partners with Reliance for its first AI data center in India. Meta is building a 168-megawatt AI-enabled data center in Jamnagar, Gujarat, powered by renewable energy and cooled with desalinated seawater. The facility will support Meta's global AI computing and will be ready in two years. India's installed data center capacity has grown from 375 megawatts in 2020 to 1.5 gigawatts in 2025, and projections show it could hit 8 gigawatts by 2030. For SMBs, this signals where AI inference is headed: distributed infrastructure in lower-cost geographies. If you're building AI-powered products, expect global latency to improve and costs to drop as hyperscalers move compute closer to emerging markets. Source
Super Micro stock drops 9% after announcing $7 billion in financing to cover AI server orders. The company received $39 billion in AI server orders from over 20 customers in recent weeks. Demand is so high that Super Micro is raising $5 billion in underwritten stock offerings and $2 billion in at-the-market offerings to cover component purchases. CEO Charles Liang said memory costs have more than tripled in recent months. For small businesses, this is a signal that AI hardware supply is still tight. If you're planning infrastructure investments (on-premise GPU servers, edge AI devices), expect higher prices and longer lead times through at least 2027. Cloud-based inference remains the better bet for cost control. Source
Shift 80% of AI workloads to cheaper models. Coinbase co-founder Brian Armstrong predicts that within 12 to 18 months, 80% of AI workloads will run on models that are 99% cheaper. Legal AI tool Harvey already cut inference costs by 3x without reducing quality by combining Claude Opus (for intensive tasks) and smaller models like GLM 5.1. If you're using GPT-5.5 or Claude Opus for everything, audit your use cases. Most tasks (email drafts, summaries, basic analysis) can run on GPT-5.4-mini or Claude Haiku at a fraction of the cost. Reserve premium models for complex reasoning, code generation, and creative work. Source
Set custom pricing in AgentsView to track Claude Fable 5 usage. If you're using coding agents on your laptop, Wes McKinney's AgentsView toolkit helps you analyze token usage and costs. Claude Fable 5 wasn't included in the pricing database yet, but you can manually set custom prices to track your spend. This is useful if you're testing multiple models (Opus, Fable, Haiku) and want to see which workflows are burning the most tokens. Free tool, Python-based, works on local transcripts. Source
Three trends are converging: AI models are getting more expensive at the top (Fable 5 at double Opus prices), cheaper at the bottom (Google's $4.99 plan, smaller models at 99% cost reduction), and more capable at autonomy (JPMorgan's long-running agents, Nextdoor's end-to-end feature builds). For small businesses, this creates a two-tier opportunity. Use budget AI for routine tasks (scheduling, summaries, basic automation) and premium models for high-value work (complex analysis, product development, creative strategy). The companies that figure out which tasks belong in which tier will see the biggest productivity gains. Watch for more silent guardrails like Fable 5's hidden interventions. As models get more powerful, vendors are starting to control not just what they block but how they quietly degrade performance in certain areas.