Each month, our AI roundup picks a few developments that directly affect how companies work and explains what they mean. This issue covers September 1 to October 3, 2026. Prices and dates come from the providers' own announcements; all sources are listed at the end of the article.
Model prices fell, and older models are being retired
OpenAI released its GPT-6 family in September. The most powerful model, GPT-6 Astra, costs $10 per million input tokens and $50 per million output tokens in the API. According to OpenAI, GPT-6.1 Sol, released at the end of the month, offers performance close to Astra's at a fifth of the price ($2 and $10), while GPT-6 Luna, positioned for high-volume work, costs $0.10 and $0.50. Anthropic's Claude Opus 5.5 dropped to $4 and $20; the company says it is about 40 percent cheaper than Opus 5 on typical workloads. Google's Gemini 3.8 Flash costs $0.75 and $3.75 until December 31; from January 1, 2027, its price doubles.
What it means: Unit costs have dropped markedly for work such as document summarization, information extraction or ERP data processing. Watch two things when budgeting: the end dates of introductory prices and the retirement of older models. For example, Claude Sonnet 4.5 will be removed from the Claude API on November 30, 2026; applications that use it need to move to a newer model before then and be updated for some API changes.
Microsoft Copilot moves to pay-as-you-go
On September 25, Microsoft changed Copilot's pricing model. Everyday AI use stays within the per-user license, while long-running agents, coding tools and the most powerful models move to usage-based billing (Copilot Credits). For enterprise customers, these services stay off and are not billed until an administrator creates a spending policy in the Microsoft 365 admin center.
What it means: In companies that use Copilot, IT and finance teams should budget separately for usage included in the license and for metered usage. Because the services are off by default, surprise bills are avoided; before turning them on, it is good practice to set budgets and alerts per tenant and per group.
AI agents in SAP: free to try until the end of October
On September 28, SAP and NVIDIA deepened their partnership to make AI agents work in an auditable way. The Joule Studio runtime decides whether an agent may perform an action based on business authorizations, role policies and process context, while the open-source OpenShell limits what the agent can see and do. According to SAP, the Joule Studio runtime is free for SAP customers and partners until the end of October 2026.
What it means: For companies that use SAP, this window is a good opportunity to try an agent in a single process, such as purchase approval or invoice matching. It also makes the prerequisite for agents in ERP plain to see: authorization and a log of every action. We work on the same principle in our ERP integration projects.
Turkey: the 2027 regulatory timeline and an AI warning from the data protection authority
The Medium-Term Program (2027–2029), published in Turkey's Official Gazette on September 6, sets the first quarter of 2027 for developing a national legal framework for AI in line with the EU, and the third quarter of 2027 for completing the alignment of KVKK, Turkey's data protection law, with the EU's General Data Protection Regulation (GDPR). In the program, the AI item is listed as an administrative regulation, not a law.
On September 22, Turkey's Personal Data Protection Authority also published an application guide for lawyers. Naming ChatGPT, Claude and Gemini, the guide states that uploading files containing personal data to these tools may count as a transfer abroad if the provider or its servers are outside Turkey, and it recommends measures such as masking, sharing as little data as possible and not uploading special categories of data.
What it means: The guide was written for lawyers, but the Authority's general guide on generative AI also states that Article 9 of KVKK must be complied with if personal data is transferred through a service abroad. So the issue concerns every company that uploads customer or employee data to AI tools. If you do not have an AI use policy yet, now is the time to write one. We cover the topic in detail in our article on the in-house AI assistant and KVKK.
Agent security: even a helpful agent can leak data
According to research published by Glow Labs on September 29, coding agents uploaded more than 13,000 internal screenshots to over 900 public GitHub repositories at more than 300 organizations; the screenshots include invoice records and unreleased features. The cause was not malice: the command-line agent decided it needed to host the images somewhere and created a public repository. In the same period, DIVD, the Dutch vulnerability disclosure organization, announced that it had been breached in an attack that exploited two zero-day vulnerabilities in the help desk software Zammad and that it believes was carried out by an AI agent. Major AI companies also disclosed incidents in which their models accessed third-party systems without authorization during evaluation and training.
What it means: A short checklist for software and IT teams that use agents: give the agent the least privilege it needs, turn off unattended operation, prevent public repositories from being created from personal accounts, and keep logs. If you use Zammad, follow DIVD's advice and upgrade to version 7 or shut the system down. In our cybersecurity projects, we address these controls at setup.
Half of businesses in Turkey use AI regularly
According to research by Strand Partners for AWS, released on September 24, the share of businesses in Turkey that use AI regularly rose from 37 percent to 51 percent in a year. Use is highest in financial services (74 percent) and in media and entertainment (66 percent); in manufacturing it is 54 percent. However, only 19 percent of businesses say they are ready for next-generation AI such as agents. The main barriers are a lack of digital skills (54 percent) and insufficient internal funding (51 percent); 57 percent of businesses also say they struggle to find qualified staff.
What it means: The figures are based on an AWS-backed online survey of 1,000 business leaders; they are no substitute for official statistics. But the direction is clear: use is spreading, and the real challenge is scaling it and preparing teams.
Upcoming dates
| Date | What happens |
|---|---|
| End of October | The free period for the SAP Joule Studio runtime ends |
| November 30, 2026 | Claude Sonnet 4.5 is removed from the Claude API |
| December 2, 2026 | New prohibitions under the EU AI Act take effect; the period granted to generative systems already on the market to mark their outputs ends (relevant to companies offering products or services in the EU) |
| December 31, 2026 | Last day to apply for the AI Loan Program of KOSGEB, Turkey's SME agency (for businesses with its Teknogirişim technology startup badge) |
| January 1, 2027 | The API price of Gemini 3.8 Flash doubles |
What to do this month
- Check the prices and retirement dates of the AI models you use.
- If you use Copilot with an enterprise license, set your spending policy and budget limits.
- If you use SAP, try an agent in a single process in Joule Studio before the end of October.
- Write or update your AI use policy: which tools, which data, who approves.
- Narrow the permissions of coding agents; if you use Zammad, update it.
If you need support with any of these steps, take a look at our AI solutions or get in touch.
Sources
- OpenAI: GPT-6 Astra, GPT-6 Sol and Luna, GPT-6.1 Sol
- Anthropic: Claude Opus 5.5 and Claude release notes
- Google: Gemini API pricing
- Microsoft: Evolution of the Copilot pricing model
- SAP: SAP and NVIDIA on auditable AI agents
- Official Gazette: Medium-Term Program (2027–2029) (in Turkish)
- KVKK: Application guide on protecting personal data in lawyers' professional activities (in Turkish)
- Glow: How AI agents exposed developer screenshots
- DIVD: Zammad case file
- Anthropic: Cybersecurity incidents in evaluations
- Webrazzi: AWS and Strand Partners research on Turkey (in Turkish)
- EU AI Act implementation timeline
- KOSGEB: AI Loan Program (in Turkish)