🤯 DeepSeek V4 Flash Shocks Opus

Gemini catches robot mistakes. .

Free AFIRE Guide | AI Academy | Advertise | AI Mastery A-Z

Plus: Prompting Has Changed. Anthropic’s New Rules with 7 Old Habits You Must Stop!

DeepSeek’s Post-Training Shock: How V4 Flash climbed close to Claude Opus 4.8 without a bigger model, and why its ultra-low pricing could reshape the AI coding race.

What’s on FIRE 🔥

IN PARTNERSHIP WITH STORECLAW

Auto-Generate Free SEO Audit Report

Most e-commerce sellers have no idea why their listings aren’t ranking. Wrong keywords, missing metadata, weak product descriptions — the problems are there, but no one’s told you where to look.

StoreClaw runs a full SEO audit across your Amazon and Shopify stores automatically. In minutes, you get a clear score, a breakdown of what’s hurting your rankings, and exactly what to fix.

No manual review. No SEO agency. No guesswork.

Connect your store and StoreClaw surfaces every issue that’s costing you search visibility — then tells you how to fix it.

Free to start. No credit card required.

Get FREE SEO report today

AI INSIGHTS

🤯 DeepSeek V4 Flash Nearly Matches Opus 4.8 at a Fraction of the Price

This is insane. I’m not kidding at all. DeepSeek just pushed V4-Flash-0731 into public beta, and the agent upgrade is much bigger than the “Flash” name suggests.

DeepSWE: 54.4%, beating GLM-5.2’s 44% and landing surprisingly close to Claude Opus 4.8 at 59%.
Terminal-Bench 2.1: 82.7%, alongside major gains across CyberGym, Toolathlon Verified, NL2Repo, and other agent benchmarks.

DeepSeek says the architecture and model size haven’t changed. This jump came entirely from additional post-training. V4 Flash also gets native Responses API support and dedicated Codex compatibility. And the pricing is ridiculous:

Input: $0.14 per 1M tokens
Output: $0.28 per 1M tokens
Cached input: $0.0028 per 1M tokens

For comparison, Opus 4.8 is listed at $5 input and $25 output per million tokens. Is it DeepSeek’s answer to the recent price pressure from OpenAI’s Luna & Terra models? Near-frontier coding performance at budget-model pricing is an insane combination.

10x the context. Half the time.

Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you’d cut when typing. Free on Mac, Windows, and iPhone.

Try Wispr Flow free

AI SOURCES FROM AI FIRE

1. Highly Recommend Bookmark This Super Detailed Guide: How to Build a Viral Faceless YT Channel with Claude from Zero (Full Ranking-Video System). You can use AI to plan, script, generate visuals, edit, and build a repeatable process without starting from zero every time.

2. Prompting Has Changed. Anthropic’s New Rules with 7 Old Habits You Must Stop! Anthropic engineer Thariq Shihipar explains which popular habits no longer work, what has replaced them, and how to adjust your workflow to get better results.

3. Claude Code’s Creator Reveals the Engineering Skill Every Coding Agent Needs. Learn how tests, lint rules, CI checks, documentation, and Skills can stop repeated mistakes and help Claude Code work better across your entire codebase.

TODAY IN AI

AI HIGHLIGHTS

🤖 Google DeepMind launched Gemini Robotics 2, giving humanoid robots full-body control from feet to fingertips. They can walk, crouch, handle objects, complete long tasks, and even work with other robots.

🍎 Heavy Siri users may have to pay extra. Tim Cook says Apple is considering paid iCloud+ upgrades for people who use the new AI-powered Siri more often. Google Gemini helps power parts of it.

🌍 Google added Nano Banana 2 to Google Earth, then pulled it just one day later. People were sharing misleading AI images built on real satellite views, so Google is adding stronger safety controls.

🛡️ Google says Gemini helped fix 1,072 Chrome security bugs across two June updates. That’s more than Chrome fixed across its previous 23 versions released over two years.

🎬 MiniMax H3 just launched, generating up to 15-second 2K videos with native stereo sound from text, images, audio, or video. It already ranks #2 for text-to-video with audio, close to Google’s Gemini Omni Flash, with open weights planned soon.

💰 Big AI Fundraising: Fish Audio raised $52M in seed funding led by Coreline Ventures and Capital Today, just one year after starting as a hobby project on a single NVIDIA 4090. The 22-person startup hit $21M ARR and 8M+ users, showing massive demand for expressive AI voice models.

HOT PAPERS OF THE WEEK

1/ Kimi K3 pushes open models closer to the frontier
Kimi K3 from the Kimi Team and Moonshot AI is a 2.8T-parameter MoE model with native vision and a 1M-token context window. It improves scaling efficiency by around 2.5× over Kimi K2 and performs strongly across coding, agents, reasoning, and vision. Big shift: open models are getting closer to systems like Claude Fable 5 and GPT-5.6 Sol, while still releasing full model weights.

2/ NYU turns chemistry papers into searchable claims
AskChem from New York University and Matterstack organizes chemistry research around evidence-backed claims instead of full documents. It currently indexes 2.4M claims from 147K papers, each linked to a DOI and source evidence. Key result: grounding GPT-5.5 with AskChem gives 100% resolvable DOIs, helping scientists and AI agents build more reliable cross-paper answers.

3/ Alibaba brings Qwen agents closer to real device control
Qwen-UI-Agent from Alibaba Group is a foundation GUI agent for mobile, desktop, web, and DeepSearch tasks. It combines GUI actions, CLI commands, long-horizon RL, and training across more than 10,000 concurrent environments. Big impact: it beats models like Claude Opus 4.8, GPT-5.6 Sol, and Gemini 3.1 Pro on several mobile and browser benchmarks.

NEW EMPOWERED AI TOOLS

🎬 Dreamina Seedance 2.5 creates cinematic AI videos up to 3 minutes long, with precise editing, lighting & production tools.
🎬 MiniMax H3 creates and edits high-quality AI videos from text, images, videos, or audio, #1 best video generation models now!
🧭 Nautis is an AI operating system for founders that connects fundraising, finance, CRM, hiring, meetings, and daily operations in one workspace.
🤝 Humalike gives AI agents stronger social skills and proactivity through behavioral APIs, models, and benchmarks.
🔎 Lev8 uses parallel AI agents and live web search to find, research, enrich, and contact the right people and companies.
🔥 Fuzzy AI warms prospects through content and thoughtful engagement before launching personalized LinkedIn and email outreach.

AI BREAKTHROUGH

🤖 Google Gave Robots a Smarter Brain with Gemini Robotics ER 2

Google just launched Gemini Robotics ER 2, a new high-level reasoning model that helps robots understand their surroundings, plan tasks, coordinate tools, and monitor their own progress through continuous video. Main upgrades:

Detects spills, slips, misalignment, and incomplete steps while a task is happening.
Tracks completion across five progress stages and finds exact moments in video.
Reached 57.4% accuracy for progress tracking and 91.3% accuracy for moment detection.
Lets multiple robots divide tasks and work together.
Reads ten instrument types, including digital displays, rulers, thermometers, and gauges.
Stops robots when a person enters the working area and resumes once it is safe.

ER 2 acts as the robot’s brain, while separate robot models or APIs handle physical movement. Developers can test it now through the Gemini API and Google AI Studio.

Key takeaway: Gemini Robotics ER 2 helps robots watch what is happening, recognize mistakes, and adjust their plans instead of blindly following a fixed sequence.

We read your emails, comments, and poll replies daily

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