Skip to content
Source · Daily Brief

AI Daily Brief — 27 February 2025

Thursday delivered OpenAI’s largest model ever and the most controversial pricing decision of the year. GPT-4.5 — code-named Orion — shipped as a research preview to ChatGPT Pro ($200/mo) subscribers and paid API tiers; Plus and Team rollout was scheduled for the following week. OpenAI’s technical paper initially stated “GPT-4.5 is not a frontier model” before deleting the line hours later. Sam Altman framed it as “the first model that feels like talking to a thoughtful person” but a “giant, expensive model” — the last to use classic pre-training rather than chain-of-thought. SimpleQA hit 62.5% with a 37.1% hallucination rate (versus GPT-4o 61.8% / o3-mini 80.3%). API pricing landed at $75 per million input tokens and $150 per million output tokens — 15-30x more expensive than GPT-4o. OpenAI said it was “still deciding” whether to keep GPT-4.5 in the API long-term due to cost. DeepSeek opened Day 4 of Open Source Week with DualPipe, EPLB, and profile data. NVIDIA’s Q4 beat continued reverberating with the stock trading roughly flat as investors digested DeepSeek-driven capex anxieties.

Top stories

  • OpenAI launches GPT-4.5 “Orion” research preview. Largest model to date; last classic pre-training-only model. Pro ($200/mo) and paid API tiers first; Plus/Team rollout following week. SimpleQA 62.5% / hallucination 37.1%. Initial paper “GPT-4.5 is not a frontier model” line deleted hours later. Altman: “first model that feels like talking to a thoughtful person” but “giant, expensive.” via TechCrunch · via Simon Willison
  • GPT-4.5 API pricing draws sticker shock. $75 per million input tokens / $150 per million output tokens — 15-30x more expensive than GPT-4o. OpenAI said it was “still deciding” whether to keep GPT-4.5 in the API “long term” due to cost of running it. via Fortune
  • DeepSeek OSW Day 4 — DualPipe + EPLB + Profile Data. DualPipe: bidirectional pipeline parallelism that overlaps computation and communication to reduce pipeline bubbles. EPLB: expert-parallel load balancer for MoE that replicates high-load experts across GPUs. Profile data: prefilling analysis for training and inference. via DigiAlps
  • NVIDIA Q4 beat reverberates; stock flat post-earnings. Revenue $39.33B vs $38.05B expected; EPS $0.89 vs $0.84; Q1 FY26 guide $43B vs $41.78B expected. Despite the beat, shares traded roughly flat as investors digested DeepSeek-driven capex anxieties — the post-Jan-27 trauma still unresolved. via CNBC

Who shipped

OpenAI‘s GPT-4.5 was the day’s marquee release. DeepSeek shipped DualPipe + EPLB. Anthropic‘s Claude 3.7 Sonnet day-3 (62.3% SWE-bench Verified standard / 70.3% custom scaffold) continued landing as Cursor called Claude “best-in-class for real-world coding tasks.” Google DeepMind, Meta, xAI, and Mistral made no dated launches.

Open-source pulse

Fetch.ai (ASI Alliance) ASI-1 Mini continued its rollout week — first Web3-native LLM purpose-built for agentic workflows, running on just 2 GPUs (~80% hardware cost reduction). Extends MoE into Mixture-of-Models (MoM) and Mixture-of-Agents (MoA) with four dynamic reasoning modes. via Fetch.ai

Money, infra & hardware

GPT-4.5’s $75/$150 pricing landed in stark contrast to Claude 3.7 Sonnet’s $3/$15 (Feb 24). The cost gap (25x on input, 10x on output) became the dominant framing line — “if Anthropic can charge $3 for 70.3% SWE-bench Verified, what is OpenAI charging $75 for?” Microsoft Phi-4-multimodal coverage continued on MarkTechPost. via MarkTechPost

By the numbers

  • $75 / $150 per million tokens — GPT-4.5 API pricing
  • 15-30x — pricing premium over GPT-4o
  • 62.5% / 37.1% — GPT-4.5 SimpleQA accuracy / hallucination rate
  • $200/mo — ChatGPT Pro tier (first GPT-4.5 access)
  • Most-mentioned lab: OpenAI

Compiled by AI Feed’s editor from verified web sources for 27 February 2025.