Loading article…
OpenAI replaces GPT‑5.3 with GPT‑5.5 Instant, cutting response length 30%, slashing hallucinations 52.5% and boosting AIME math score to 81.2.
OpenAI switched the default ChatGPT engine to GPT‑5.5 Instant this week, promising sharper answers, far fewer hallucinations and real‑time self‑correction for all users【1】. The upgrade matters because the model powers the daily experience of hundreds of millions, so even modest efficiency gains translate into a noticeable shift in how the AI behaves across consumer and enterprise workloads.
| At a glance | |
|---|---|
| Model | GPT‑5.5 Instant (default) |
| Hallucination reduction | 52.5% fewer false claims in high‑stakes domains |
| Verbosity | ~30% fewer words per response |
| Math benchmark | AIME 2025 score 81.2 (up from 65.4) |
OpenAI says GPT‑5.5 Instant trims unnecessary verbiage, delivering answers with roughly 30% fewer words than its predecessor, which reduces “yapping” and extraneous emojis【1】. More importantly, internal tests show a 52.5% drop in hallucinated statements when the model tackles medicine, law or finance prompts【1】. The model also flags and fixes its own mistakes mid‑sentence, a capability demonstrated on a math problem where it paused, identified an error and corrected it before completing the answer【1】.
On the AIME 2025 math test GPT‑5.5 Instant scored 81.2, a jump from the 65.4 achieved by GPT‑5.3 Instant【1】. The multimodal MMMU‑Pro benchmark also rose to 76 versus 69.2 for the older version【2】. While the performance uplift is clear, OpenAI introduced a higher API price—input tokens cost $5.00 per million and output $30.00, double the rates for GPT‑5.4【3】. The company argues the new model is more token‑efficient despite the higher headline price, but the cost increase will affect developers who rely on the API for large‑scale workloads.
Paid Plus and Pro subscribers now gain “Memory Sources,” letting the model draw on prior chats, uploaded files and linked Gmail accounts to tailor responses, with full transparency on which sources were used【1】【2】. These personalization features launch first on the web, with mobile and broader tier access slated for the coming weeks. The older GPT‑5.3 Instant will stay available via API for three months before retirement, giving developers a transition window【1】.
The shift to GPT‑5.5 Instant underscores OpenAI’s focus on making its default model not just more capable but also more reliable and user‑controlled, a move that could set a new baseline for conversational AI performance while raising questions about cost and long‑term developer adoption.
Coverage is mostly measured — 103 of 125 reports stay neutral.
Every Monday — the token unlocks, Fed dates & catalysts set to move crypto and markets this week. So you’re never blindsided.
Free · 3-min read · one-click unsubscribe
AI-assisted synthesis by the TrendWatcher Editorial Desk · sourced from 4 outlets · Jun 25, 2026 · How we report
It improves conversational quality, goal understanding, handling of complex instructions, and adapts to user feedback, according to OpenAI's release notes.
Free users are expected to receive the update within a day of its rollout to paid users.
Jalapeño is a purpose‑built ASIC for LLM inference, intended to increase performance per watt and reduce dependence on GPU hardware.
Broadcom will manufacture the chip and associated server hardware, with Celestica assembling the racks.
OpenAI aims to begin deployment by the end of 2026 and expand over several years.