ChatGPT、新モデル「GPT 5.4」公開–ハルシネーション削減、事実誤認3割減に – CNET Japan https://www. yayafa.com/2775154/ #
agents gpt-5 openai
| Source: Mastodon | Original article
OpenAI unveiled its latest large‑language model, GPT‑5.4, on 8 March 2026, rolling out two flavours – GPT‑5.4 Thinking and GPT‑5.4 Pro. The company says the “Thinking” variant is tuned for coding, AI‑agent orchestration and complex reasoning, while the “Pro” version targets high‑throughput professional workloads. Both models boast a 1 million‑token context window, native computer‑operation APIs and a new “Tool Search” layer that lets the model invoke external utilities on the fly.
The headline claim is a 30 percent cut in factual errors and a marked drop in hallucinations, measured against GPT‑4‑Turbo in OpenAI’s internal benchmark suite. Early testers report that the model now surfaces its reasoning plan before answering, a feature that makes its output more transparent and easier to audit. By reducing spurious statements, GPT‑5.4 narrows the gap that has allowed rivals such as Anthropic’s Gemini to claim superior reliability in enterprise settings.
Why it matters is twofold. First, the lower error rate makes the model viable for mission‑critical tasks – legal drafting, financial analysis and software development – where misinformation can be costly. Second, the expanded context window and built‑in tool execution push ChatGPT further toward true agentic AI, enabling it to manage multi‑step workflows without external prompting. This evolution dovetails with the growing ecosystem of AI‑enhanced services, from Claude Code’s terminal‑based coding partner to ZOZO’s app‑linking experiments, and could accelerate adoption of AI agents across Nordic enterprises.
What to watch next are the rollout details: OpenAI plans a staged release to ChatGPT Plus users in April, followed by API access for developers in May. Industry analysts will be scrutinising real‑world error rates, pricing tiers for the Pro model, and how quickly third‑party platforms integrate the new tool‑search capabilities. The next few months should reveal whether GPT‑5.4 can deliver on its promise of more trustworthy, agentic AI at scale.
Sources
Back to AIPULSEN