- AI Fire
- Posts
- 😬 Fable 5 Falls Flat? Why?
Plus: Everyone Says Kimi K3 is 'Open'. I Checked What It Truly Takes to Run It (What to Use Instead)
Remember when Fable 5 was so powerful the U.S. gov briefly paused it? Businesses seem less impressed. Now, it accounts for just 6% of Anthropic token usage? Why?
What's on FIRE 🔥
IN PARTNERSHIP WITH RUBRIK
Attackers have agents. So do you. Both act through identity, and neither waits for a human.
Join security, identity, and AI leaders to master the operating model for the agentic era: defend against AI attacks, govern agentic identities, and recover any IdP at machine speed.
AI INSIGHTS
A person representing themselves in a Connecticut court tried to hide a prompt injection inside an official legal filing, telling any AI reviewing the document to favor their side. Main details:
The hidden instructions were written in tiny 3-point white text.
They told an AI system to make its output agree with the plaintiff’s filing.
Court staff noticed unusual white space and discovered the concealed text.
The judge said Connecticut courts were not using AI to review the filing, but still treated the attempt as improper.
Later filings reportedly included more hidden messages and jokes.
The plaintiff argued that the prompt injection was an “audit” designed to test whether the court secretly used AI.
The judge allowed the case to continue but removed the plaintiff’s ability to file documents electronically.
Prompt injection could become a serious problem once courts, law firms, or government systems use AI to process documents automatically.
Key takeaway: Documents themselves can become attacks against AI systems. Any organization using LLMs to review legal, financial, or business files needs protection against hidden instructions inside the content it processes.
PRESENTED BY BELAY
Learn How to Stay Visible in the AI Era
AI is changing how customers discover businesses. If your SEO strategy is built for yesterday's search, your visibility is already slipping. Learn how to optimize your content for today’s AI search results with BELAY’s latest report..
AI SOURCES FROM AI FIRE
1. FREE: Elon Musk's NEW Grok 4.6 Completely Made It Feel Like Fable? Here's My Full Test. People say Grok 4.6 feels different. So I tested across real workflows to see if it actually brought back the Fable experience, if yes, at a huge cheaper cost.
2. How to Manage AI Agents Like a Real Team (BEST #1 Skill Anyone Needs in 2026 & 2027). You’ll be able to build your own AI team that researches, codes, analyzes, and gets real work done, and honestly, this is a skill you definitely want to learn now.
3. [Google AI Ecosystem Playbook] Lesson 2: Master the Gemini App. Turn an unclear request into an approved direction and execution-ready work without letting Gemini make the important decisions for you.
HOW TO RUN KIMI K4 LOCALLY
“Open” sounds simple right? After looking deeper, I found the hardware demands, practical challenges, and alternatives that could make more sense for most users.
Moonshot AI calls Kimi K3 an open-weight model, and this time “open” isn’t just a nice label.
They shared the model weights, configuration, and inference code, so you can download Kimi K3 and run it on your own machine.
But for me, the word “open” sounds a little funny. You’re allowed to self-host Kimi K3, but only if you have enough GPUs and you’re ready to pay the cloud bill.
So what should we do now? What does it actually take?
Today, I’ll show you what Kimi K3 really needs, how much self-hosting can cost, and what makes more sense if you just want to use the model, like easily.
TODAY IN AI
AI HIGHLIGHTS
🚽 The Compiss is a vibe-coded app (here) that points you toward the nearest public bathroom. It’s going viral on Reddit. The name is ridiculous, but I’d absolutely use it.
⏳ Hit your usage limit in Claude? There's now an auto-continue checkbox that picks up where you left off after your limit resets. Turn it on & let Claude work without you.
😅 OpenAI apparently has its own corporate SOS button. Employees have a special “friction” email to reach Sam Altman & tell him 'all is not well'. Here's how it works.
🧬 In a long & super viral post on Saturday, Anthropic CEO says AI’s biggest problem is a “crisis of trust." He thinks AI could cure most diseases within 5-10 years.
🔥 Anthropic appears to be testing a Model Comparison interface inside Claude. You could switch from normal chat into side-by-side model testing whenever you want.
🤝 Stripe is reportedly buying OpenRouter for $7B+, giving it access to a platform with 400+ AI models. Interestingly, OpenRouter was valued at just $1.3B in May.
🌍 The U.S. is preparing to push dozens of countries to pick its side or China’s in AI race, warning they will be excluded from a U.S.-led AI coalition if they choose China.
🤯 OpenAI’s annualized revenue just topped $40B as it prepares for an IPO, roughly 2x its late-2025 run rate. Massive revenue is still coming alongside massive losses.
💰 Big AI Fundraising: Wall Street giants are lining up $500B to finance Nvidia AI infrastructure, while Nvidia may backstop up to $125B itself. The bet: GPU fleets will keep generating enough value to repay the debt.
NEW EMPOWERED AI TOOLS
📚 Blume turns plain Markdown into fast, AI-ready docs with built-in search, SEO, and one-command deployment using Astro or Vite.
🤖 HarnessRouter Community Edition lets you run Codex, Claude Code, and Hermes through one API, handling all on your infrastructure.
🌍 Expeditione turns learning into interactive 3D adventures, letting you explore handcrafted worlds directly in your browser with no login.
🧬 Vidaya combines your wearables, lab results, and DNA and personalized longevity plan that tells you what to do next.
AI BREAKTHROUGH
Ramp’s August AI Index suggests Fable 5 is struggling to justify its premium price, even while Anthropic remains highly popular with businesses. One month after Fable 5 returned to market, adoption was still surprisingly small:
Fable 5 accounted for only 6% of Anthropic API tokens purchased.
It represented 11.4% of total Anthropic model spending.
GPT-5.6 Sol captured 25% of OpenAI tokens and 23% of OpenAI model spend.
In absolute spending, Fable 5 generated only about 75% as much spend as GPT-5.6 Sol in July.
Fable also competes with Anthropic’s own cheaper models. Opus 5 costs $5/$25 and Sonnet 5 costs just $2/$10, so businesses have strong alternatives for workloads that don’t require Fable’s highest-end reasoning.
Anthropic restored global access on July 1, so Ramp’s July data effectively captures Fable’s first full month back.
Ramp found Anthropic was used by 43.5% of U.S. businesses in its dataset, compared with OpenAI at 39.7%. So the issue is pricing. Companies still prefer cheaper models like Sonnet 5 and Opus 5 unless Fable’s extra capability clearly matters.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply