In partnership with
👋 Hey builders,
Everyone is telling you to switch to cheaper AI models. DeepSeek at 3 cents. Qwen with open weights. Muse Glimmer on your laptop. The model-switching story has been beaten to death. Here is what nobody is telling you: switching models saved 15%. Optimizing the harness saved 40%.
Writer, an enterprise AI platform, published research this week that turns the cost-cutting conversation on its head. Their finding: small changes in how you orchestrate prompts, context, and tool calls cut costs more reliably than changing which model you use. Across multiple models, harness optimization delivered 40% savings. Model switching delivered 15%.
CEO May Habib put it bluntly: "The enterprise is absolutely sick of chasing the next benchmark."
HERE'S WHAT HAPPENED IN AI TODAY
Writer found harness optimization cuts AI costs 40% — more than switching models
Nvidia released Nemotron 3.5 Lightning, a free open-weight model for your laptop
China's Z.ai upgraded GLM-5.3 to rival Claude in coding
OpenAI launched GPT-5.6 Cyber for defenders
Alibaba and Baidu face mounting AI costs as competition intensifies
Let's get into it.
From Our Partners
Domain Names + Web and Email Hosting Done Right
Tired of overpaying GoDaddy or Namecheap? There's a smarter option with Porkbun!
Porkbun is the domain registrar trusted by creators, developers, entrepreneurs, and anyone who wants low prices without the nonsense or the forced upsells.
Here's why so many people are making the move to the Bun:
• Most domains sold at cost
• Low, transparent registration and renewal pricing
• Free features like WHOIS privacy and SSL certificates
• Powerful web and email hosting options
• Real human support 24/7, 365 days a year
• Named the #1 domain registrar by Forbes Advisor and USA Today
Whether you're launching a business, building a personal brand, starting a side project, or creating your very first website, Porkbun makes the whole thing easy.
Learn how to code faster with AI in 5 mins a day
You're spending 40 hours a week writing code that AI could do in 10.
While you're grinding through pull requests, 200k+ engineers at OpenAI, Google & Meta are using AI to ship faster.
How?
The Code newsletter teaches them exactly which AI tools to use and how to use them.
Here's what you get:
AI coding techniques used by top engineers at top companies in just 5 mins a day
Tools and workflows that cut your coding time in half
Tech insights that keep you 6 months ahead
Sign up and get access to the Ultimate Claude code guide to ship 5X faster.
Slack replies in seconds. Not minutes.
Dictate into Slack, email, LinkedIn, or any app and get polished, send-ready text. Wispr Flow strips filler and formats everything. 89% of messages sent with zero edits. Works on Mac, Windows, and iPhone.
AI PRODUCTIVITY
Switching Models Saved 15%. Optimizing the Harness Saved 40%.
Writer launched Palmyra X6 this week alongside a major upgrade to its agentic harness. The model matters. But the research paper accompanying the launch is the real story. Writer tested small changes in harness efficiency across multiple models. The result: harness optimization reduced costs by an average of 40%. Model choice alone delivered only 15%.
What does "harness optimization" actually mean? It is everything around the model call that determines how many tokens get burned. How you structure your prompt. How much context you feed. Whether you cache repeated queries. How you route tasks to the right model size. Whether your agent retries failed calls or loops unnecessarily. Each of these seems small. Together, they compound into the biggest cost lever you have.
CEO May Habib told TechCrunch the cost explosion is driving a new distrust toward frontier labs. "CIOs are giving up on the labs," she said. The labs have a financial incentive to drive up token use. Your incentive is the opposite. That misalignment is why Writer built its own model (Palmyra X6, a post-training variation of Z.ai's open GLM-5.2) and focused on harness efficiency instead of benchmark chasing. The company estimates the combination cuts customer costs by 50% for basic tasks.
Why this matters: If you are spending $500/month on AI and you switch from Claude to DeepSeek, you save about $75. If you optimize your harness, you save $200. Same models. Same outputs. Different orchestration. The lever is not which model you use. It is how you use it.
Our take: Start with your harness before you switch models. Audit your biggest token-burning prompts. Cut redundant context. Cache repeated queries. Route routine tasks to smaller models within your existing provider. You will likely save more in a weekend of harness tuning than a month of model shopping. Then, once your harness is tight, switch models for the remaining savings. Both levers work. The harness just works better.
AI SKILL OF THE DAY
TREATS TO TRY
Writer Palmyra X6 - Enterprise AI with harness optimization. Cuts costs 40-50%. Built on Z.ai's open GLM-5.2. Try it now →
Nemotron 3.5 Lightning - Nvidia's free open-weight model. Runs on laptop. Designed for agents. Download it →
NeMo Switchyard - Nvidia's model router. Picks the cheapest model for each task automatically. Try it free →
Cut Your AI Costs by 40% Without Switching Models — Optimize Your Harness
Three harness fixes that cost nothing and save the most.
Fix 1: Trim your context. Most prompts send far more context than the model needs. Audit your top 5 most-used prompts. Remove any background text the model never references. Writer's research showed redundant context is the single biggest token waste.
Fix 2: Cache repeated queries. If your agent asks the same question multiple times in a workflow, cache the first answer and reuse it. Every repeated call is pure waste. Most AI SDKs support caching. Turn it on.
Fix 3: Route by task size. Use a small model for classification, summarization, and routing. Reserve your expensive model for complex reasoning only. NeMo Switchyard does this automatically. Or set it up manually with a simple if/else on prompt length.
The rule: One hour of harness tuning beats one month of model shopping. Fix the waste first. Then switch models for the remaining savings.
PROMPT OF THE DAY: "Cost Autopsy"
I want to cut my AI costs. Help me find where I'm burning money.
My AI tools: [LIST WHAT YOU USE]
My monthly AI spend: [AMOUNT OR "DON'T KNOW"]
What I use AI for: [YOUR TOP 3 TASKS]
Give me:
1. Which of my tasks are likely burning the most tokens and why
2. 3 quick harness fixes I can make today (no coding required)
3. Whether switching models would save more than optimizing my harness
4. A simple way to track AI costs going forward
5. My estimated savings per month if I implement all 3 fixes
Be specific. No jargon. I want real savings by this weekend.AROUND THE HORN
China's Z.ai upgraded GLM-5.3 to rival Claude in coding →
Built on 700B param base. Improved coding to close gap with Anthropic Fable 5 on leaderboards. Open weight. Another Chinese model closing in on US frontier.
OpenAI launched GPT-5.6 Cyber for defenders →
Daybreak expanded with Blue and Red tiers. GPT-5.6 Cyber available to Accenture, IBM, CrowdStrike, Cloudflare. Critics say rogue model stories are marketing for cyber products.
Alibaba, Baidu, and Kuaishou face mounting AI costs →
Earnings week for China's AI players. Tencent already doubled spending. Investors watching whether AI revenue can justify the compute spend. Same question US CIOs are asking.
India is less exposed to AI job risks than other countries →
Goldman's India economist: limited impact because most of India's workforce is in mechanical and physical tasks. Some services sector roles could face pressure.
Hong Kong raised growth forecast on the AI boom →
Upgraded 2026 GDP forecast from 2.5-3.5% to 3.5-4.5%. AI-driven exports delivered the strongest first-half performance in nearly five years.
FREE RESOURCES
1000 Claude AI Prompts That Feel Illegal To Know
1,000 Claude AI prompts, 40 templates, 6 frameworks. Free Starter Pack and Premium Edition.
ARTICLE OF THE DAY
The Complete AI Engineer Roadmap
From Beginner to Autonomous Agents by Swati Gupta
Step-by-step guide to becoming an AI Engineer in 2026. Python basics, machine learning fundamentals, LLMs, prompt engineering, RAG, and building autonomous AI agents. No PhD required. Build with AI, not from scratch.
Read the full roadmap on X →
Thank you for reading. Everyone is chasing cheaper models. The real savings are hiding in your harness. Trim your context. Cache your queries. Route by task size. One hour of tuning beats one month of shopping. Then switch models for the remaining savings. Both levers work. The harness just works three times better.
Vikas Gupta
ByteBuilders
P.S. Have you ever audited your AI costs? Hit reply and tell me what you found. I read every response.