OpenAI’s 2025–2026 season is best read as a product architecture, not a highlight reel of model names. This consequence file separates what is confirmed, what is analysis, and what is still unknown — Brocker’s standing rule for AI coverage.
Quick takeaways
- Confirmed: ChatGPT is a routed system; the API is an explicit model family — dual surface since GPT-5.
- Confirmed: Responses + Agents SDK + built-in tools (including computer use) are the agent platform; Assistants sunsets August 26, 2026.
- Analysis: OpenAI competes on surface/tier architecture; Anthropic’s parallel bet is access twins (Fable/Mythos).
Confirmed
- August 7, 2025: GPT-5 launches as ChatGPT system + API family (gpt-5 / mini / nano / gpt-5-chat-latest), with plan-based usage differences and GPT-5 Pro for extended reasoning.
- July 9, 2026: GPT-5.6 GA with Sol / Terra / Luna tiers across ChatGPT Work, Codex, and API; chat product continues to remap which tiers free vs paid users see by default.
- March 11, 2025 → August 26, 2026: Responses becomes the agent primitive; Assistants API beta sunsets on the hard date OpenAI published.
- Agents SDK: orchestration layer for loops, handoffs, guardrails, sessions, tracing — optional when you own Responses yourself.
- Computer use & hosted tools: web search, file search, computer use (and related Responses tools) as the action surface.
Analysis
- In our view, OpenAI’s durable advantage claim is full-stack default: consumer app, developer SKUs, agent API, and UI automation in one vendor path.
- GPT-5.6’s Sol/Terra/Luna naming makes pricing/performance tradeoffs legible — similar pressure Anthropic answered with Opus-near-frontier pricing, different mechanism (Anthropic Frontier).
- The Assistants sunset mid-2026 forces ecosystem consolidation onto Responses just as model marketing accelerates — platform lock-in via integration gravity, not only via evals (cadence context).
Unknown
- How long Chat Completions remains a peer path versus soft-deprecated in practice for new agent features.
- Whether Sol/Terra/Luna naming stays stable across the next frontier bump, or resets again.
- Independent reproducibility of every agentic/coding benchmark chart OpenAI publishes.
- Enterprise readiness bars for unsupervised computer-use deployments (policy, approval UX, liability) beyond docs examples.
Our take
What actually changed is the menu builders and buyers must navigate: which surface, which tier, which API primitive, and whether the agent may act. What has not changed is the need to verify work when stakes are high — OpenAI says the same in its own GPT-5 developer materials.
Series: OpenAI Product Stack — dual surface · Responses/Assistants · Agents SDK · computer use · consequence.