Get tomorrow's brief in your inbox
Today: n8n and competitors battle over AI workflow deployment control while hyperscalers bet billions on natural gas that could triple in price. OpenAI and Anthropic slash API costs 80% as Chinese rivals force a price war, and Google lets you hide AI watermarks (but keeps invisible ones).
The AI automation platform wars are heating up, and deployment control is the new battleground. n8n published a detailed comparison positioning itself against Make, Zapier, and Temporal, arguing that self-hosted deployment and fair pricing matter more than ever as AI agents become production infrastructure. The timing is strategic: businesses are realizing that workflow automation platforms lock you in for 3-5 years, and picking the wrong one means either living with vendor pricing changes or migrating hundreds of workflows.
Here's what matters for small businesses: if you're building AI into your operations now, the platform you choose determines whether you control costs at scale or get squeezed later. Zapier charges per task and caps AI agent use at 400/month on free tiers. Make uses per-operation pricing that drains budgets fast. n8n requires you to manage infrastructure but gives you fixed costs and no usage caps. The analysis also reveals that leading AI adopters only use AI for 18% of workflow steps, with the rest handled by rules and logic. If you're routing every step through a model, you're likely overpaying by 71%.
Action: Before committing to any automation platform, calculate what 10,000 workflow runs would cost at next year's scale, not today's free tier. Self-hosting isn't free either (you're paying in time), but it's predictable.
Kog claims it can run LLM inference 30x faster on standard datacenter GPUs like the AMD MI300X and Nvidia H200 by optimizing software at the assembly-level, bypassing inefficiencies in standard GPU libraries. The French startup hit Hacker News in May with a demo hitting 3,000 tokens per second on a 2B parameter model. Now they're promising the same approach works for full-scale LLMs, targeting customers tired of waiting hours for Claude Code results. Anthropic already charges a premium for Fast Mode because speed equals money.
The catch: this only works on datacenter GPUs, not your laptop, and Kog hasn't yet proven it scales to production LLMs (their demo used a small custom model). But if they deliver, businesses running AI workflows on existing hardware could see massive cost savings without buying new chips. Early demand came from 200 business leads after the demo.
Worth watching. If you're evaluating dedicated inference providers, wait to see Kog's full-LLM benchmarks before committing to specialized silicon.
OpenAI just slashed GPT-5.6 Luna pricing by 80%, and Anthropic cut Opus 5 to half the cost of Fable 5. This isn't charity - it's a price war driven by Chinese AI labs like Moonshot and DeepSeek offering open models that companies like DoorDash and Airbnb are already testing to cut costs. Average token prices across leading US models dropped 25% since mid-July.
For small businesses, this is great news short-term: AI API costs are falling fast. But it also signals that flat-rate subscriptions are disappearing in favor of usage-based billing, which means unpredictable bills if you don't monitor consumption. Companies are responding by capping AI usage or switching to cheaper Chinese models.
Action: If you're on a usage-based AI plan, set hard monthly spend limits now. Track token consumption weekly. If your bill jumped when providers migrated you off flat pricing, test cheaper alternatives like open models or regional providers.
Zapier analyzed its top AI adopters and found four distinct roles AI plays in workflows: Communicator (84% use it), Clerk (79%), Analyst (44%), and Coordinator (25%). The Communicator drafts emails and summaries. The Clerk extracts structured data from unstructured input like call transcripts. The Analyst makes yes/no decisions that trigger downstream actions. The Coordinator creates tasks in project management tools.
The key insight: 44% of all AI workflow runs happen outside business hours, mostly the Analyst role making automated decisions. If you're using AI to check if a number is above a threshold, you're wasting money - a conditional rule does that for free. Reserve AI for steps that require judgment or language interpretation.
Quick win: Audit your current automations. If you're calling an AI model for something a spreadsheet formula could handle, replace it. You'll cut costs 71% on average.
Hyperscalers are betting billions on natural gas, and it could backfire hard. Meta, Microsoft, Google, and Amazon collectively announced plans for over 20 gigawatts of natural gas power plants to run AI data centers. They're locking in today's cheap prices ($2-4 per million BTUs), but energy research firm Noreva forecasts prices could triple to over $10 as hyperscaler demand collides with declining supply growth and rising LNG exports.
Here's the risk: fuel represents about half the cost of electricity from a gas plant. If prices double or triple, "bring your own power" data centers get much more expensive to run, which could drive up AI token costs or push hyperscalers back onto the grid (driving electricity prices higher for everyone). One investor was "surprised" by how much gas price risk these companies are taking on.
For small businesses: if you're betting on API prices continuing to fall, this is a reminder that infrastructure costs could reverse that trend. Natural gas futures don't anticipate big changes yet, but neither did anyone in West Texas before new pipelines connected them to global markets.
Google now lets you hide visible watermarks on AI-generated images, videos, and music, but still embeds invisible SynthID and C2PA metadata. The change acknowledges that visible watermarks make content less useful for professional work, and many users were editing them out anyway. You can still verify if something is AI-generated by asking Gemini or Search.
Anthropic made the opposite move this week, adding visible watermarks to text and images to comply with EU regulations. Users are unhappy.
The takeaway: Invisible watermarking is becoming the standard. OpenAI and Meta already skip visible marks. If you're creating AI content for clients, check whether your contract requires disclosure - some industries care, most don't yet.
Use chain-of-thought prompting to cut AI hallucinations by nearly half. A study found standard prompts hallucinate 34.5% of the time, while chain-of-thought ("think step-by-step") drops that to 18.1%. For complex multi-step tasks, add "walk me through this in manageable parts" to your prompt. It forces the model to show its work, making errors easier to catch.
Try this: Next time you ask an AI to solve something complex (pricing strategy, break-even analysis, debugging logic), append "show your reasoning step-by-step" to the prompt. You'll get better answers and spot mistakes faster.
This week's stories point to a maturing AI market where infrastructure costs, deployment control, and pricing transparency are replacing raw capability as competitive factors. The price war signals oversupply in API intelligence and undersupply in differentiation. Businesses that locked in flat-rate pricing are getting migrated to usage-based billing, and those that assumed costs would keep falling are discovering energy economics work both ways. Smart buyers are testing multiple providers, setting hard budget caps, and reserving AI for tasks that actually need judgment.