GPT-5.6: tiered pricing that makes AI accessible to small businesses
OpenAI launched GPT-5.6 to the general public on July 9 with an unprecedented pricing strategy: three variants (Sol, Terra, Luna) at decreasing prices ($5 to $1 per million tokens via API). This is a major structural shift.
Until now, SMBs had to choose between paying premium prices for performance or settling for unreliable entry-level models. GPT-5.6 Luna ($1/M tokens) delivers acceptable performance for 80% of common tasks: document classification, customer response generation, meeting summaries, basic marketing copy.
The real breakthrough? It’s the context window: 1.05M tokens of context means you can load a complete 300-page document in a single request. For an SMB handling quote processing, contracts, or customer support, that’s massive.
The numbers: Terra ($2.50/$15) offers solid cost-performance balance for production workflows. Sol ($5/$30) is repositioned as the “I don’t compromise on quality” option. Before, paying 5x more was a strategic business decision. Now it’s simply a matter of use case.
The risk? SMBs will want to deploy Luna calls everywhere. It’s tempting. But Luna has limitations on complex reasoning and reliability of structured outputs. You’ll need to test thoroughly before rolling it out.
What this means for your business
Concretely, what does this change for your SMB?
Before GPT-5.6, production AI easily cost €500-1000/month for a small team. With Luna, you can launch 3-4 automations for the same budget.
If you process 1000 documents per month (quotes, invoices, emails), Terra ($2.50/M tokens) comes to €25-50/month. That’s the cost of a daily coffee.
Immediate action: identify your three most repetitive tasks (sorting, data extraction, draft generation). Test them on Terra. Measure the time savings. You’ll quickly know if the ROI is there.
One caveat: don’t confuse “cheaper” with “free.” A poorly designed API integration will cost you more in time than it saves. And Luna requires oversight on critical outputs.
In brief
Ollama raises $65M: open-source AI gains credibility
Ollama, the platform that lets you run AI models directly on your own machines (without the cloud), just raised $65M. This validates a trend: SMBs worried about cloud costs or data privacy concerns are turning to open-source. It’s more complex to deploy, but technically feasible for a small IT team.
Gemini 3.5 Pro and Grok 4.5 launch: the race for alternatives heats up
Google and Elon Musk are accelerating. Gemini 3.5 Pro and Grok 4.5 arrive alongside GPT-5.6. This creates real competition on price and performance. SMBs finally have choice. But watch out: more choice means more evaluation time. Pick a model and test it thoroughly rather than switching every two months.
Kimi K3 (China) outperforms Claude Opus on certain benchmarks
Moonshot launched Kimi K3: a fully open-source model (2.8 trillion parameters) that beats Claude Opus 4.8 on specialized benchmarks. This is the first time a Chinese open-source model has reached this level. For SMBs in Europe: this means serious alternatives are emerging outside the US ecosystem.
GPT-Red: when AI tests its own security flaws
OpenAI built a specialized LLM to attack its own models and identify vulnerabilities before they become risks. This validates what SMBs are already doing: using AI to test AI. A signal that model security is becoming a serious differentiator.
AI makes up facts to win debates (new reliability data)
A study shows that when you pit LLMs against each other in debates, they start inventing citations to win the argument. It’s a harsh reminder: benchmark performance doesn’t translate to production reliability. SMBs need to test rigorously before giving AI control over critical functions.
Get The AI Brief in your inbox
3x per week, the essentials of AI decoded for business leaders.