- AI Fire
- Posts
- 🥊 Grok 4.6 Matches GPT-5.6 for $2
Plus: Masterclass Replay + Complete Resource Pack: Build Your First Working App with Base44 & Claude Code
Grok 4.6 just landed, and somehow it matches GPT-5.6 on major intelligence tests while costing less. If that holds up, xAI may have a serious pricing weapon right?
What's on FIRE 🔥
IN PARTNERSHIP WITH SECTION
AI spend is rising, so CFOs want proof. Join the AI:ROI Conference to get a methodology for driving measurable value without token “wildfire.”
One day, 10 sessions, enterprise-level insights. Join free on September 17!
AI INSIGHTS
Grok 4.6 is SpaceXAI’s new flagship reasoning model for coding, autonomous agents, technical research, and professional knowledge work. Core points:
500K-token context window
Supports text, images, web search, X search, code execution, and function calling
Starts at $2/M input tokens and $6/M output tokens
Pricing doubles once prompts reach 200K tokens
Matches GPT-5.6 Sol at 61 on Artificial Analysis’ Intelligence Index
Beats GPT-5.6 Sol on several professional-agent benchmarks, including GDPVal-AA v2 and AA-Briefcase
Still trails leading models on DeepSWE and Terminal-Bench
Artificial Analysis measured roughly 67.6 output tokens/sec, but about 42 seconds to first answer token
Its biggest advantage is cost. Grok 4.6 reaches roughly the same composite intelligence score as GPT-5.6 Sol while charging substantially lower headline API rates.
Grok 4.6 looks especially strong for long-running agents and professional workflows where cost matters, but slower initial responses, uneven coding performance, and some safety regressions keep it from being the clear best model for every task.
PRESENTED BY HUBSPOT
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
AI SOURCES FROM AI FIRE
1. FREE: Grok Bot Just Killed 1,000 Startups? An Insanely Easy AI Agent Anyone Can Use 24/7. Grok Bot is bringing AI agents closer to everyday users. In this honest review, we test what Grok Bot can do, how it handles real tasks, and whether it can truly work like a 24/7 AI employee.
2. Stop Overpaying for Claude Sonnet 5. Here's the Exact Migration Plan (Full Guide). Sonnet 5 is powerful, but I bet many of you are paying premium prices for this. It’s time to build a smarter routing system, and get more value from your setup.
3. Multi Agent System: Use Graph Engineering to Make Claude & Codex 10x Better. Build a cleaner AI workflow that plans, researches, checks, and gets human approval before Claude or Codex makes the final call.
LIVE APP-BUILDING MASTERCLASS RECAP
💪 Masterclass Replay + Complete Resource Pack: Build Your First Working App with Base44 & Claude Code
Click to watch our full masterclass
The masterclass turned out even more active than we expected. People kept asking questions, sharing ideas, helping each other, and bringing up real problems while we were building the app live. So we saved everything for you.
If you joined live, you’ll have everything organized in one place so you don’t need to search through the recording again.
If you couldn’t make it, you can still watch the full masterclass and get the same prompts, notes, and practical resources we shared with everyone live.
🌟 Here’re all the resources we promised during the session, including every prompt we used, Base44 + Claude Code tips and tricks, troubleshooting notes from problems we ran into, and shortcuts from the build.
|
TODAY IN AI
AI HIGHLIGHTS
🎉 DeepSeek-V4-Pro is officially out with stronger agent capabilities, API support, and Codex integration. It’s available on web, mobile, and API, so you can test it now.
🖋 If you’re building something and still don’t have a clear go-to-market plan, save this viral guide on how YC startups approach GTM, then recreates it with Claude.
😮 Google packed Made by Google ’26 with the Pixel 11, Pixel Watch 5, Pixel Tag, and a wave of new Gemini features. There’s a lot to catch up on, here's the recap.
😅 Find a great notebook? Steal, I mean, make it yours! You can now make a copy of a notebook, including all sources, artifacts, and keep building. Ctrl+C, Ctrl+V energy.
👀 Grok Bot is offering a 7-day free trial on desktop for some users, even without SuperGrok Heavy or Cursor Ultra plan. Usage is limited, but enough to test it here.
🕵️ I can already see this becoming controversial. Some users are furious, worried schools and employers may now detect when Claude helped write something.
🚨 Twitch will now let Amazon use streamer content for generative AI training by default. You have to opt out manually, so check your settings if you stream there.
💰 Big AI Fundraising: Today was INSANE with:
1/ Cognition is seeking a $40B valuation after raising $1B at $26B in May. Its annualized revenue has already jumped from $492M toward $1B.
2/ Lovable raised $400M after reaching $500M in annualized revenue. It now hosts 60M projects and gets 900M monthly visits.
3/ OpenAI-backed Thrive Holdings raises $2B. Its AI agents processed 7,000+ returns at 98% accuracy, while help desk resolution became 36x faster.
4/ Anthropic will buy Decart AI for about $6 billion, its biggest known acquisition yet. Decart helps AI chips run more efficiently and also builds real-time video models.
NEW EMPOWERED AI TOOLS
📈 Agorapulse's new Free guide breaks down LinkedIn’s new rules, showing why saves, sends, and real comments now matter more than likes for reach, and how to adapt. Download the guide.
🖥️ Unsloth Desktop lets you run leading open models locally, including Muse Glimmer, DeepSeek-V4, and Kimi K3.
🎙️ Dograh is a VAPI alternative with 30+ model integrations, telephony, human transfers, and voice agents you can self-host for free.
📊 Lettertrace tracks how often Claude, ChatGPT, and Gemini mention your company, with free tracking using your own API keys.
AI BREAKTHROUGH
Researchers found a way to recover parts of hidden AI reasoning from encrypted reasoning blocks used by OpenAI, Anthropic, and Google. Main findings:
Encrypted reasoning blocks from powerful models could sometimes be replayed into weaker sibling models from the same provider.
With jailbreak prompts, those weaker models sometimes converted the encrypted blocks back into readable reasoning without the encryption key.
Across 315,320 public reasoning blocks, researchers recovered 367 pieces of personal information and 182 credentials, including passwords and API keys.
Kimi K3 sometimes produced reasoning that closely resembled hidden traces from frontier models, although similarity alone does not prove how it was trained.
The attack became possible because encrypted reasoning blocks can travel through APIs, apps, sessions, users, and different models so a model can later resume its previous reasoning.
An application cannot normally read those blocks, but another model from the same provider could sometimes interpret them.
OpenAI, Anthropic, and Google were notified before publication and reportedly changed their systems.
A simple way to picture it: Claude Opus seals its private reasoning inside an envelope that an app cannot open. Researchers found that they could sometimes hand that same envelope to Claude Haiku and persuade Haiku to read its contents aloud.
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