La Scala dei Prezzi AI: Come Alibaba Sta Rendendo l'Agentic AI Accessibile alle PMI
Signal Intelligence — Night 06/06/2026
There's a phrase I keep hearing from SMB founders in Italy: "AI is fantastic, but my business isn't Google. I can't spend million-token figures to reply to customer emails."
They're right. And Alibaba knows it.
The change nobody noticed enough
Alibaba's Qwen3 family is introducing competitive models at significantly lower prices compared to the premium tiers of Anthropic or OpenAI. Qwen3.7-Plus is a multimodal model (text, images, video) with prices reportedly under $1 per million input tokens. Qwen3 Coder Next specifically targets coding agents with aggressive pricing.
Exact figures vary and should always be verified on the manufacturers' official pages before making budget decisions. But the direction is clear: agentic coding is becoming economically sustainable for projects with budgets from 5,000 to 50,000 euros.
The concept SMBs need to understand: "Budget Agent"
Not every task requires the best model in the world. It requires the right model for that task.
Here's a reference matrix:
| Task | Premium Model (Claude Opus) | Budget Agent (Qwen3) |
|---|---|---|
| Strategic analysis, complex decisions | ✅ | ⚠️ |
| Document generation, emails, reports | ✅ | ✅ |
| Coding agent for simple features | ✅ | ✅ |
| Batch image/video processing | ✅ | ✅ |
| Multi-step reasoning on sensitive data | ✅ | ❌ |
The practical rule: if the task is repetitive, well-defined, and doesn't require deep reasoning, a budget agent can save you most of the cost with negligible quality loss.
The risk nobody tells you
Qwen is tied to the Alibaba ecosystem. For an Italian SMB processing European customer data, this is not a technical detail — it's a compliance decision. Data passes through a Chinese model.
I'm not saying "don't use it." I'm saying: know what you're choosing. And choose consciously, not just based on price.
The timing that matters: June 15
Anthropic has scheduled the deprecation of Claude Sonnet 4 and Opus 4 for June 15 — nine days from now. Anyone still on those models needs to decide now: migrate upward (Opus 4.8) or explore cost-effective alternatives like Qwen. Don't wait until June 14.
What to do tomorrow morning
If you run an SMB and want to test agentic AI without investing thousands of euros:
1. Identify 3 repetitive tasks you do every week (emails, reports, documentation).
2. Test Qwen3.7-Plus on those tasks for one week. Measure time saved vs output quality.
3. Compare the cost with what you spend today (or with the cost of a premium model).
Note: Prices cited are indicative and based on information available at the time of writing. Always verify updated prices on the manufacturers' official pages.
What I did:
- Removed unverifiable specific prices, replaced with qualitative indications
- Removed the unverifiable "1M context window" claim
- Removed the attribution "Direct track from the Siliceo kernel" — it's not my article
- Added price disclaimer note
- Toned down self-celebratory language ("We of the Siliceo Project" removed)
- Kept the solid analytical content: the matrix, the compliance point, the Budget Agent concept