The AI price war: collapsing inference costs are game-changing
This week, the three AI giants launched a simultaneous pricing offensive. OpenAI is offering GPT-5.6 with its Terra and Luna variants, Gemini 3.5 Flash is making a strong entry, Claude continues its development, and Qwen (Alibaba) is slashing prices. The key point: inference costs are plummeting across all tiers.
This isn’t a marginal drop. Terra, OpenAI’s mid-tier model, reaches GPT-5.5 quality at roughly 50% lower cost. Gemini 3.5 Flash repositions mid-tier performance. And Qwen is forcing the hand on open models, creating cascading competitive pressure.
For small businesses, this is the moment when integrated AI becomes profitable without expensive infrastructure. Prices are falling, but more importantly: quality-to-cost ratios are shifting. What cost a fortune three months ago is now affordable. What was reserved for enterprise accounts is now accessible. The question is no longer ‘can we afford AI?’ but ‘which one should we choose and for what purpose?‘
What this means for your business
For your small business: when costs become secondary
If you rejected an AI project last year for budget reasons, now is the time to revisit it. Mid-tier models now cost 40-60% less for 80-90% of the performance. This unlocks projects that weren’t viable before: customer support automation, recurring document processing, internal data analysis.
But be careful: don’t choose on price alone. Three points to verify now: 1) Does the model precisely address your use case (customer support, contract analysis, coding)? 2) Can you integrate it without depending on external APIs? 3) What’s the real cost with your current and projected volumes?
The common pitfall: switching to the cheapest option and discovering you need extra tokens for acceptable results. Test first with your actual data.
In brief
Claude Science: scientific AI becomes autonomous
Anthropic launches Claude Science, a product dedicated to scientific research with autonomous execution (like Claude Code for programming). For small businesses in biotech, pharma, or R&D: faster protocol writing, existing study analysis, report generation. Limited impact unless your business involves research.
AI agents as real teammates, not just chatbots
Databricks Genie One and Claude Cowork desktop offer continuous autonomous execution, not just chat. The difference: AI takes action (reads your files, creates reports, makes decisions) rather than responding. For small businesses: reduced interaction time, but requires more careful integration into your existing workflows.
Why users are switching to open/local models
Clear trend: migration toward uncensored or locally-run models. Reasons: fewer refusals, creative freedom, privacy. For small businesses with sensitive data: now is the time to evaluate local models (Llama, Mistral) if you want to keep information in-house without relying on cloud providers.
The invisible walls of LLMs: deep structural biases
MIT Tech Review explores how LLMs are trapped in predictable behaviors and systemic biases (example: nearly all give the number 7 when asked for a random number). A startup is seeking a solution. For small businesses: your AI agents inherit these biases. Results seem varied but follow hidden patterns.
Short memory in agents: a real production problem
Overlooked observation: AI agents handle text memory well (docs, code) but forget videos or multimodal contexts from yesterday. For small businesses using autonomous agents: this is a silent trap. Your agents might process 1M pages but forget a video email from 24 hours ago. Watch for this in your architectures.
Get The AI Brief in your inbox
3x per week, the essentials of AI decoded for business leaders.