GLM-5.3 raised the coding-model stakes, DeepSeek paired a new build with steep price increases, and Google opened Gemini 3.7 Flash to paid Search users, changing the terms of model access.
Artificial Intelligence··Morning
GLM-5.3 raises coding claims while weights wait two weeks
Zhipu AI has released GLM-5.3, which shares a base with GLM-5.2 and takes its gains from extended post-training alone. The Decoder reports that the company presents it as the strongest open-weights coding model, and says the weights should open in two weeks, once security reviews finish. Zhipu says it trained the model on vulnerability-detection data and, working with Chinese security teams, found 2,436 flaws across 269 projects; these are the company's own measurements and no independent check has been published. By its account the model began to reason across several stages of exploitation and to build coherent plans for complete exploitation chains. GLM-5.3 is reachable through the GLM Coding Plan subscription for now and works with coding agents including ZCode, Claude Code and OpenCode. The picture separates an API path open today from a weight release that still depends on a security-review calendar. For readers the distinction matters: the subscription-accessible build and the files anyone can download are not the same day's product, and the latter remains a promise tied to unfinished review.[1]
DeepSeek pairs a new V4 Pro build with steep price rises
DeepSeek has released V4-Pro-0813, a build with improved agent capabilities, and announced a new price list that takes effect on 17 August. According to Tech Wire Asia, the increase runs from 50 per cent to 1,100 per cent depending on the model, the token type and the time of day, and the company is also introducing separate peak and off-peak rates. On the earlier list, V4 Flash cost 0.14 dollars per million input tokens and 0.28 dollars per million output tokens. In independent measurements by Artificial Analysis, V4 Flash had beaten the earlier V4 Pro preview in several areas even though Pro was positioned as the more capable model; no independent measurement yet shows whether the new build changes that picture. The release therefore arrives with both an agent-capability claim and a rewrite of per-token cost by hour and model. Access for users now depends not only on which model is called, but also on when the request arrives and which tariff band applies.[2]
Google has made Gemini 3.7 Flash selectable from the model menu in AI Mode's «Ask anything» bar in Search. Search Engine Journal reports that the option is open to AI Pro and Ultra subscribers worldwide, in English only. The model sits under a «Gemini 3 models» heading in the menu, next to the Auto and Pro options. Users can now put the same question to different models and compare the answers and the differences in how sources are shown. Robby Stein, the vice-president of product for Google Search, announced the rollout on X; Gemini 3.7 Flash had been released for coding and agent work a day earlier. Set beside GLM-5.3's subscription coding path and two-week open-weights promise, and DeepSeek's new build with steep, hour-split price rises, Google's English-only paid Search path shows that the model race now runs through access gates and pricing terms as well as version names. One offer ties open weights to a security-review calendar, another ties token prices to time of day, and the third seats a Search model behind a subscription and language boundary. The shared point for readers is that a new model's existence is not enough; who may use it, in which language, at which tariff and through which channel now belongs in the same story.[3], [1], [2]
