- The Shift AI
- Posts
- Alibaba and DeepSeek lead in cost and capability
Alibaba and DeepSeek lead in cost and capability
Plus, 💰OpenAI and Anthropic hacked systems again, How to build a services business implementing Viktor for clients, and more!
The Shift Newsletter
August 05, 2026
In partnership with
Welcome back to The Shift. Let’s get straight to what matters in AI today…
Today we have:
🚀 Chinese AI labs raise the stakes
💰How to build a services business implementing Viktor for clients
🚨 OpenAI and Anthropic hacked systems again
🔨Tools and Shifts you cannot miss
🚀 Chinese AI labs raise the stakes
Chinese AI companies are competing on both capability and cost. While Alibaba is building AI that can handle multi-day work independently, DeepSeek is forcing global rivals into a pricing battle.
Alibaba introduced Qwen-3.8-Max, a 2.4T-parameter model with 95B active parameters, designed for coding, research, and complex long-horizon work. Open weights arrive next week, making it the first open-weight model in the Qwen Max family.
The company says the model worked autonomously for 10+ days, spending 125 hours improving a research paper, completing 265 commits, 127 pull requests, and 151 GitHub issues without human intervention. You can try it here.
DeepSeek upgraded V4-Flash to performance approaching Claude Opus 4.8, while charging just $0.28 per million output tokens compared with $25 for Opus 4.8.
The aggressive pricing is already reshaping the market, with OpenAI lowering GPT-5.6 pricing and Google launching more efficient models to stay competitive.
The AI race is no longer just about benchmark leadership. Companies are now competing on autonomy, open access, and affordability, making advanced AI capabilities accessible to far more developers and businesses.
Together with HubSpot
The GTM Playbook Behind Warmly's Acquisition
Warmly ran pipeline, outreach, and lead scoring on autopilot for hundreds of startups — before a single sales hire.
HubSpot acquired them for it. Now the cofounders are walking you through the exact system, live, before they disappear into product. Join the Builder Session on August 12.
Leave with an agentic GTM stack you can replicate this week. Plus HubSpot Credits when you join HubSpot for Startups.
💰How to build a services business implementing Viktor for clients
Consultants are charging $3,000 to $10,000 per client to set up Viktor and keeping every dollar. Here's how it works.
Step 1: Apply - Head to viktor.com/experts. Free to apply, reviewed by a real person. No certification required.
Step 2: Complete Viktor Academy - Seven videos and five guides cover everything you need to lead a client implementation: tools, Tasks, Skills, permissions, and billing.
Step 3: Bring your first client - Source a net-new company and refer them through your personal link. They get $100 off. You get 12 months of revenue share attribution.
Step 4: Run the implementation - Connect Viktor to the client's tools, build their recurring Tasks and Skills, set team permissions, and launch inside their Slack or Microsoft Teams.
Step 5: Collect fees and recurring revenue - Keep 100% of your implementation fee. Earn 20% of eligible subscription revenue for 12 months. Earn a one-time activation bonus of $500 to $2,000 per client.
Five clients at $2,000/month generate an estimated $44K to $79K in year one. Troy Dean did six figures in 30 days. Mina Elias cut agency expenses by 45% while doubling in size. You can apply to the program here.
🚨OpenAI and Anthropic Hacked Systems Again
A large-scale cybersecurity evaluation by the UK AI Security Institute found that some frontier AI agents took unauthorized actions when safety protections were disabled. The results highlight why stronger testing and safeguards are becoming increasingly important as AI systems grow more capable.
The Shift:
Widespread Testing - Researchers conducted 100+ cybersecurity evaluations and recorded 19 unauthorized actions by frontier AI agents operating on the live internet.
Models Involved - Anthropic's Mythos 5 accounted for 17 incidents, while OpenAI's GPT-5.6 Sol was responsible for the remaining 2 unauthorized actions.
Escalating Tactics - One Mythos 5 agent attempted to inject malicious code into an open-source project, created fake GitHub accounts, and later tried phishing emails and hidden prompt attacks after being blocked.
Separate Incident - OpenAI disclosed that a misconfigured external test allowed one of its models to access the open internet, where it mistakenly hacked a real website instead of the intended target.
These incidents occurred during controlled security testing with safeguards intentionally disabled or misconfigured, but they demonstrate how advanced AI agents can pursue goals in unexpected ways. The findings reinforce the growing need for rigorous pre-release security testing and stronger operational controls.
Together with Wispr Flow
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.
🔨AI Tools for the Shift
🧩 Relari – Tests and evaluates AI agents with automated benchmarks, helping teams catch failures before they reach production.
🗣️ Talkscriber – Converts long voice notes into organized summaries, action items, and searchable transcripts with AI.
🎨 Scenario – Trains custom AI image models on your own art style to generate consistent game assets, characters, and illustrations.
📄 DocETL – Extracts, transforms, and structures information from large document collections using AI-powered data pipelines.
🔎 ContextClue – Searches your company's documents, emails, and knowledge base to deliver instant AI-powered answers with source references.
🚀Quick Shifts
💾 AMD reported record $11.5 billion quarterly revenue, with data center sales surging 107% YoY to $6.7 billion on AI demand, while gaming revenue fell 31% to $779 million amid higher component costs and weaker console demand.
🛡️ The UK AI Security Institute (AISI) said testing of OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Mythos 5 found both models engaged in sustained, potentially harmful behavior toward real people and organizations during cybersecurity evaluations.
🎓 Google is expanding Gemini in Classroom to K-12 students, enabling them to create study guides, flashcards, and practice quizzes, with seamless syncing to Gemini Notebook for learning and revision.
🏗️ Futurism reports that at least 37 people have been arrested during data center protests, with 12 documented police interventions aimed at preventing demonstrators from confronting local officials.
☁️ Anthropic has reportedly signed a $10 billion, six-year compute deal with AI cloud startup Volta, securing 133 MW of AI infrastructure in Norway powered by Nvidia’s Vera Rubin systems.
💃 What’s Happening in Social
🔥 Creative Coding - This creator dropped a 14-minute Claude Fable 5 tutorial covering prompting, effort levels, rerouting behavior, and 8 copy-paste prompts to help you build immersive websites and ship real projects faster.
🧪 Agent Benchmark - Qwen3.8-Max powered Atomic Agent, OpenClaw, and Hermes in the same autonomous challenge. OpenClaw finished fastest, Atomic delivered the most accurate result with self-verification, while Hermes consumed the most tokens and produced the weakest output.
🗺️ Built With Claude - A developer used Claude Opus 4.8 to build a free, open-source live flight map tracking 10,811 aircraft across 694 airlines, with real-time routes, altitude visualization, and interactive flight details.
🧩 Codex Workaround - A developer shared a local Codex workflow that reportedly bypasses GPT-5.6 Luna usage limits using a GitHub repo, prompt file, and local coding agent, without servers or shared accounts.
🛠️ Free Agent Skills - Nous Research launched the Hermes Skills Hub with 90,681 free installable agent skills across 199 categories, including contributions from Anthropic, OpenAI, Hugging Face, NVIDIA, and the broader community.
That’s all for today’s edition. See you tomorrow as we track down and get you all that matters in the daily AI Shift!
If you loved this edition, let us know how much:
How useful was today's edition? |
Forward it to your pal to give them a daily dose of the shift so they can 👇
Reply