01
OpenAI launches GPT-6 Sol and Luna with lower API prices
OpenAI announced GPT-6 Sol and Luna for the API, ChatGPT Work and Codex. It set API prices 50% below the promotional prices of their GPT-5.6 predecessors.
- Recalculate costs for Czech-language support and document-processing workloads using actual token volumes rather than the headline discount.
- Benchmark Czech-language accuracy, latency and failure rates before migrating production applications.
- Confirm regional availability and data-processing terms separately for API deployments and ChatGPT Work.
Source — OpenAI ModelsAPI pricing 02
GPT-6 prompt caching adds discounts and monitoring tools
OpenAI announced improved prompt caching, including discounts for eligible shared prompt prefixes reused within 30 minutes. New monitoring tools help developers inspect cache performance and diagnose misses.
- Measure prefix reuse in document assistants and internal knowledge applications to identify workloads eligible for savings.
- Add cache-hit rates and effective token costs to production dashboards before revising AI budgets.
- Check cache eligibility and retention terms against internal rules for confidential corporate and personal data.
Source — OpenAI DevelopersCost management 03
OpenAI proposes international technical standards for frontier AI
OpenAI called for US-led international work on frontier-AI technical standards. Its proposal covers common capability measurements, incident reporting and systems capable of recursive self-improvement.
- Treat the announcement as a policy proposal, not a new compliance obligation for Czech businesses.
- Ask vendors for comparable capability evaluations when assessing models for regulated or high-impact workflows.
- Specify incident-notification deadlines and escalation contacts in contracts rather than relying on future standards.
Source — OpenAI AI governanceStandards 04
Altman urges human control in UN Security Council remarks
Sam Altman addressed the UN Security Council on human control of powerful AI systems. He called for international cooperation on frontier-AI safety standards.
- Require approval steps for agents that can change business records, send external messages or initiate payments.
- Test whether administrators can stop agent actions and reconstruct decisions from audit logs.
- Distinguish OpenAI’s public oversight commitments from controls included in the specific service and contract you buy.
Source — OpenAI Human oversightAI policy 05
OpenAI publishes principles for independent AI safety assessments
OpenAI published principles for effective third-party assessments of AI systems.
- Ask suppliers to identify independent assessors and disclose the scope of their evaluations.
- Check whether assessment evidence covers the model version and deployment configuration your organisation will use.
- Request Czech-language and sector-specific testing where general evaluations do not cover your workflows.
Source — OpenAI AI safetyThird-party assessments 06
MentalHealthBench targets safety in mental-health conversations
OpenAI introduced MentalHealthBench, an open benchmark developed with more than 80 licensed mental-health experts from 22 countries. It evaluates safety, context-seeking, user agency and actionable guidance.
- Include domain-specific safety tests when evaluating assistants for employee assistance, healthcare or insurance.
- Validate performance in Czech and against local clinical practice before using benchmark results in deployment decisions.
- Keep clinical validation and human escalation requirements separate from benchmark scores.
Source — OpenAI Mental healthSafety evaluation 07
OpenAI extends Daybreak cyber-defense access to Ukraine
OpenAI announced that Ukraine’s government would receive access to its Daybreak program for civilian infrastructure defense. The announced uses include finding software vulnerabilities and developing and testing fixes.
- Do not assume this government arrangement establishes Daybreak availability for Czech enterprises.
- Limit AI-assisted vulnerability testing to explicitly authorised systems and documented scopes.
- Require human review and controlled testing before applying AI-generated fixes to critical infrastructure.
Source — OpenAI CybersecurityCritical infrastructure