01
OpenAI previews GPT‑5.6 Sol, Terra, and Luna
OpenAI launched a limited preview of GPT‑5.6 with three tiers and highlighted Sol for cybersecurity tasks. Access is restricted to select partners via API and Codex, with broader availability planned later.
- Use Sol/Terra/Luna to align model selection with risk, cost, and latency requirements across security, knowledge work, and high-volume automation.
- Treat “limited preview” as a delivery constraint and confirm regional availability and contract terms before committing Czech rollouts that depend on specific model IDs.
- Use the announced safeguards and access controls as inputs to your EU AI Act and internal security review, especially for cyber-related use cases.
Source — OpenAI modelssecurity 02
GPT‑5.6 pricing and caching policy changes published
OpenAI disclosed per‑token pricing for Sol, Terra, and Luna and introduced new prompt‑caching billing rules and cache behavior. The same announcement described explicit cache breakpoints and a 30‑minute cache lifetime.
- Update TCO models using the published input/output token rates and the new cache write/read economics, then re-run unit economics for long-context and retrieval-heavy workflows.
- Ask vendors and integrators how they will manage cache breakpoints and cache lifetime in production to avoid unpredictable bills and latency spikes.
- Use the pricing disclosure to benchmark against EU-hosted alternatives when data residency or procurement risk requires a second-source option.
03
OpenAI plans Cerebras-hosted GPT‑5.6 Sol for speed
OpenAI said GPT‑5.6 Sol on Cerebras will start in July for select customers, targeting inference speeds up to 750 tokens per second. The offering initially targets a subset of customers rather than general availability.
- If you run latency-sensitive Czech user experiences (agent assist, chat, real-time analysis), request benchmarks that match your context lengths and tool-calling patterns.
- Validate how the Cerebras deployment affects data handling, logging, and incident response compared with your current OpenAI deployment path.
- Plan capacity tests early because “select customers” gating can delay go-lives if your project depends on the higher-throughput SKU.
Source — OpenAI inferenceperformance 04
White House reportedly asked OpenAI to limit GPT‑5.6
CNN reported that the White House requested OpenAI restrict GPT‑5.6 access to government‑approved partners due to cybersecurity misuse concerns. The report says OpenAI is staggering rollout while vetting continues.
- Add an “access restriction” risk line to procurement and roadmap plans if your architecture depends on always having the newest OpenAI frontier model available in the EU.
- Expect tighter controls on cyber-capable models and prepare internal policy on who can use them, for which tasks, and with what monitoring.
- Reduce dependency risk by maintaining a fall-back model strategy for critical workflows, including contractual clarity on model availability and deprecation.
05
ChatGPT Enterprise adds memory upgrades, Skills, and connectors
OpenAI’s Enterprise & Edu release notes describe memory improvements, new model availability in ChatGPT, a Skills feature for reusable team instructions, a Codex app for macOS, and expanded MCP connectors. The changes extend ChatGPT’s role as an integration layer, not just a chat interface.
- Treat Skills as a governance primitive and define who can publish, approve, and audit reusable instructions to avoid inconsistent outputs across teams.
- Reassess data classification and retention controls before enabling improved memory in Enterprise/Edu, especially for regulated Czech sectors.
- Use MCP connectors to reduce custom integration work, but require a connector-by-connector security review (scopes, logging, admin controls, and least-privilege).