Get tomorrow's brief in your inbox
Today: Anthropic made Claude Fable 5 permanent on Max plans after pressure from GPT-5.6 competition. China's Moonshot AI released the world's largest open model, competing with frontier labs. OpenAI pitched "Useful Intelligence per Dollar" as the new scorecard for measuring AI ROI.
Anthropic reversed course on Fable 5 access this week, making its most capable model permanent on Max and Team Premium plans at 50% of limits starting July 20. The company had planned to remove subscriber access and push Fable 5 to API-only pricing, but GPT-5.6 Sol and Kimi K3 made that plan untenable. Why pay $100-200/month for a subscription that doesn't include the best model?
For small businesses, this is excellent news if you're already on Claude Max ($200/month). You get access to the most powerful reasoning model without per-token billing eating your budget. If you're on Pro ($100/month) or Team Standard, you'll receive a one-time $100 usage credit instead of guaranteed access. That credit goes further than you'd think for focused work like contract reviews, technical documentation, or strategic analysis.
The real story here is compute capacity. Anthropic's original plan was driven by GPU constraints. Making Fable 5 permanent means they're likely dialing back training runs to serve existing customers, a trade-off that favors reliability over racing to the next model version. If you run a business that depends on consistent AI access rather than chasing the newest release, that's a meaningful shift in priorities.
Simon Willison - Claude make Fable 5 permanent
Lofty CRM launches agentic AI for real estate agents
Lofty's new AI features put it ahead of the pack for real estate CRM software in 2026, according to a comprehensive review of dozens of platforms. The system handles lead generation, automated follow-ups, and property tracking with AI that recommends next actions based on customer behavior. Pricing wasn't disclosed in the review, but competitors range from $20-100/month for individual agents.
If you run a real estate business, the key differentiator is agentic AI vs. simple automation. Most CRMs send scheduled emails; Lofty's system analyzes client engagement patterns and adjusts outreach timing and content. That's the difference between a reminder system and a tool that actually thinks about what works.
Zapier - The best CRM software for real estate agents in 2026
TikTok tests AI likeness detection tool
TikTok rolled out an opt-in tool for US creators that scans for AI deepfakes using their likeness. Creators verify identity through Jumio (selfie scan and ID check), then TikTok's system monitors for unauthorized AI-generated content. YouTube launched similar protection recently.
If you create video content for your business, this matters more than you'd think. Deepfakes aren't just a celebrity problem anymore. A competitor could use AI to create fake testimonials with your face or voice. Opting in takes minutes and gives you reporting power when TikTok's system flags unauthorized use. The tool is free, identity verification happens once, and TikTok doesn't retain your ID documents.
The Verge - TikTok is testing an AI likeness detection tool
NVIDIA NeMo Automodel integrates with Hugging Face Diffusers
NVIDIA and Hugging Face partnered to bring production-grade distributed training to any Diffusers model on the Hub. The integration supports text-to-image (FLUX.1-dev) and text-to-video (Wan 2.1, HunyuanVideo) with no checkpoint conversion required. It's open source under Apache 2.0.
This is infrastructure news, but if you run a creative agency or marketing team experimenting with custom image/video generation, it matters. You can now fine-tune models like FLUX.1 on your brand's visual style without hiring ML engineers or converting formats between training and inference. The technical barrier just dropped significantly.
Hugging Face Blog - Fine-tune video and image models at scale with NVIDIA NeMo Automodel
Databricks benchmarked AI coding tools and found harness choice matters as much as model choice
Databricks CEO Ali Ghodsi shared internal benchmarking results from managing AI costs for 3,000 software engineers. The company found that open models like GLM 5.2 handle high-difficulty coding tasks at lower total cost than Anthropic and OpenAI's proprietary models. But the bigger surprise: the agentic harness (the wrapper managing context and instructions) impacts costs just as much as model selection. Open-source harness Pi delivered low costs without sacrificing quality by managing context efficiently.
The takeaway: if you're paying for AI coding tools (GitHub Copilot, Cursor, Claude Code), don't just compare model pricing. Ask how the tool manages context window usage. A cheaper model with a wasteful harness costs more per completed task than an expensive model with tight context management.
TechCrunch - Databricks hits $188B valuation
Patreon blocks AI scrapers after they ignored robots.txt
Patreon partnered with Cloudflare to actively block AI training bots, not just ask them politely via robots.txt files. Testing showed individual AI crawlers went from thousands of weekly attempts to zero after the block. Translation: they were ignoring the "please don't scrape" requests all along.
If you run a membership site, paid newsletter, or any paywalled content, this is your playbook. Cloudflare's AI Crawl Control technology is available now. Robots.txt doesn't work because AI companies ignore it. Active blocking does. Patreon still allows indexing bots that send users back to the platform, they're just blocking training scrapers.
TechCrunch - Patreon stops asking AI bots not to scrape
Claude Fable 5 becomes permanent on Max plans at 50% limits
Starting July 20, Claude Max ($200/month) and Team Premium users get Fable 5 access permanently, capped at 50% of their regular usage limits. Pro ($100/month) and Team Standard users receive a one-time $100 usage credit instead. The original plan to remove Fable 5 from subscriptions and make it API-only is canceled.
If you're deciding between Claude Max and API access: Max now makes sense for consistent high-value work (research, analysis, complex writing) where you need the best model regularly but can't predict exact token usage. API access still wins for high-volume automated tasks where you can optimize prompts and use cheaper models for routine work.
Simon Willison - Claude make Fable 5 permanent
OpenAI released GPT-5.6 family with three pricing tiers
GPT-5.6 launched last week with Sol (flagship), Terra (balanced performance/cost), and Luna (fastest, most affordable). OpenAI framed the release around "Useful Intelligence per Dollar," a new metric for CFOs measuring AI ROI. The calculation: full cost of completing work divided by tasks that met quality standards.
The strategic shift here is away from "cost per token" toward "cost per successful outcome." A frontier model with expensive tokens that completes tasks in one pass beats a cheap model that requires retries and human review. If you're budgeting AI spend for 2026, this is the framework to use.
OpenAI Blog - A scorecard for the AI age
China's Moonshot AI released the world's largest open model, competing with Anthropic and OpenAI
A Chinese startup shipped the world's largest open AI model this week, competing directly with frontier models from Anthropic and OpenAI. The launch sent AI and semiconductor stocks sliding as markets processed the implications. Xi Jinping pitched China as an AI partner to the developing world at the World Artificial Intelligence Conference in Shanghai, making it clear China plans to lead on both AI technology and standards.
For small businesses, the immediate impact is more affordable access to frontier-class performance through open models. If you're using Claude or GPT via API, Chinese open models offer comparable quality at lower cost, but you need to evaluate data privacy policies before switching. This competition also pressures US labs to keep subscription pricing reasonable.
MIT Technology Review - The Download
Databricks raised $3B at $188B valuation, fifth round in 18 months
Databricks closed another funding round at $188 billion valuation, its fifth in 18 months. The company successfully repositioned from "yesteryear SaaS sensation" to AI infrastructure provider by rolling out products like Lakebase (database for AI agents), Unity (AI gateway), and Omnigent (multi-agent orchestrator).
The lesson for established software companies: AI doesn't have to be your core product to be your growth story. Databricks sits on enterprise data and added AI governance, agent infrastructure, and cost optimization tools. If you run a SaaS business, your existing customer data and workflows are assets AI features can leverage.
TechCrunch - Databricks hits $188B valuation
Apple sued OpenAI in splashy litigation as Siri AI ships to public beta
Apple launched a lawsuit against OpenAI this week while simultaneously shipping Siri AI to public beta. The Vergecast analyzed whether Apple is genuinely worried about competition or capitalizing on OpenAI's weak moment. The timing suggests strategy: sue while shipping your own AI product.
For businesses using Apple devices and considering AI assistants, this creates uncertainty. If the lawsuit escalates, Apple might pull OpenAI integration from iOS entirely. If you're building workflows around ChatGPT on iPhone/Mac, have a backup plan. Native Siri AI is available in beta now.
The Verge - Apple's plot to crush OpenAI
General Compute landed $400M loan for inference chips, using them as collateral
General Compute raised $400 million from Upper90 to purchase SambaNova inference chips. This appears to be the first deal using inference-specific chips as collateral, not training GPUs. The SN50 chips provide 16x faster inference than GPU-based clouds and don't require expensive water cooling.
For businesses evaluating AI infrastructure: the shift from training chips to inference chips signals where the market is heading. Most companies don't need to train models; they need to run them cheaply at scale. Inference neoclouds like General Compute target that use case with purpose-built silicon that's faster and cheaper than renting Nvidia GPUs from AWS.
TechCrunch - Why the first GPU financiers are turning to inference chips
AI memory demand hit India's smartphone market hard, shipments down 10%
India's smartphone market fell 10% year-over-year in Q2 2026, the steepest decline in six years. Memory chip makers shifted production capacity toward high-bandwidth memory for AI data centers because it's more profitable than standard phone memory. That reduced supply and drove up handset prices, hitting price-sensitive markets hardest.
The sub-$150 segment dropped 45%. Chinese brands took the worst hit because they focus on entry and mid-tier phones. OnePlus responded by exiting Europe and North America while maintaining its India business.
If you're budgeting company phones or considering device upgrades, expect higher prices and longer replacement cycles. Premium brands like Apple and Samsung are less affected because buyers finance expensive devices anyway. Budget options are becoming scarce.
TechCrunch - AI-driven memory crunch jolts India's smartphone market
Agility Robotics opened 60,000 sq ft training facility in Fremont, CA
Agility Robotics opened a facility to train its Digit humanoid robots for manufacturing and warehouse deployments. The company secured $300 million in contract orders and already generates revenue from customers like Amazon, GXO, and Toyota. Digits have moved 100,000 totes at GXO logistics facilities.
Agility's approach differs from competitors: no in-home robots, just industrial work. The company takes a practical view of autonomy, keeping safety systems separate from generative AI. Digit v5 launches this fall with the ability to sense humans and work outside robot-only zones.
If you run a warehouse or manufacturing operation, humanoid robots are closer to viable than you think. Agility is already deployed and generating revenue, not pitching vaporware. The use case is bins, totes, picking, and kitting, not general-purpose labor.
TechCrunch - Agility Robotics plants its flag in Tesla's backyard
Block AI scrapers on your website now
Don't rely on robots.txt. Patreon's testing showed AI companies ignore it. Use Cloudflare's AI Crawl Control technology to actively block training bots while allowing indexing bots that send traffic back to your site. This is especially critical if you run a membership site, paid newsletter, or create proprietary content.
Use Simon Willison's LLM cliché highlighter tool
Simon Willison built a browser-based tool that detects common LLM-generated phrases in text: "no fluff, no filler," "sit with that," "you already know," etc. Paste content before publishing to catch AI writing patterns. The tool runs locally in your browser and saves settings to localStorage. Useful for editing your own AI-assisted writing or detecting lazy AI-generated content from vendors.
Simon Willison - LLM cliché highlighter
Three themes cut across today's stories: subscription models are winning over pure API pricing (Anthropic's Fable 5 reversal), open models are closing the gap with frontier labs (China's Moonshot AI, Databricks' GLM 5.2 benchmarks), and AI infrastructure is shifting from training to inference (General Compute's $400M raise for inference chips). Watch for more frontier labs to make their best models available on subscriptions rather than API-only, and expect open model quality to continue pressuring proprietary pricing. If you're evaluating AI tools this month, focus on total cost per completed task, not advertised token pricing.