GPT-5.5 nudges efficiency fight

OpenAI released GPT-5.5 on April 23 and expanded it to the API on April 24, putting a fresh flagship model into a market already crowded with near-peers. (openai.com 1) (openai.com 2) Large language models are prediction engines that guess the next token, or chunk of text, one step at a time. The expensive part is not just training them, but paying for every input token, output token and tool call once people start using them at scale. (openai.com) (api-docs.deepseek.com) OpenAI pitched GPT-5.5 as its “first fully retrained” base model in this cycle and said it is built for coding, research, spreadsheets and other multi-step computer work. Its standard API price is $5 per 1 million input tokens, $0.50 for cached input and $30 for output. (openai.com 1) (openai.com 2) That launch did not land in a vacuum. DeepSeek’s V4 line arrived days later with two Mixture-of-Experts models, including V4-Pro with 1.6 trillion total parameters, 49 billion activated parameters and a 1 million-token context window. (huggingface.co) (api-docs.deepseek.com) DeepSeek’s selling point was not simply a higher score. Its pricing page says V4-Pro is discounted 75% until May 5, 2026, and that cache-hit prices across its suite were cut to one-tenth of launch levels. (api-docs.deepseek.com) Google took a different route. On April 16, it said Gemini could use Personal Intelligence, Nano Banana 2 and a connected Google Photos library to generate images of users and their families without manual uploads or long prompts. (blog.google) (cnbc.com) Google said the feature is rolling out to U.S. subscribers on Google AI Plus, Pro and Ultra, and later said Personal Intelligence was expanding internationally with exclusions including the European Economic Area, Switzerland and the United Kingdom. Google also said Gemini does not train models on a user’s private photo library. (blog.google 1) (blog.google 2) The result is a three-way split in how major labs are trying to win buyers in late April 2026. OpenAI is selling a stronger work model, DeepSeek is cutting the bill for long-context use, and Google is tying its assistant more tightly to personal data people already store in its apps. (openai.com) (api-docs.deepseek.com) (blog.google) That is why the argument has moved beyond who tops a benchmark chart on a given week. Buyers choosing a model for coding agents, research tools or consumer assistants now have to weigh token cost, cache discounts, privacy settings and how much outside software the model can reach. (openai.com) (api-docs.deepseek.com) (cnbc.com) GPT-5.5 may have supplied the latest jolt, but the immediate response from rivals shows the contest is no longer just about raw model score. It is about how cheaply, safely and deeply these systems can fit into everyday work and personal software. (openai.com) (api-docs.deepseek.com) (blog.google)