Z.ai released GLM-5.3 on August 14, an update to its GLM-5.2 coding model that reused the same underlying base and got every improvement from further post-training. The company reports a jump from 4.6 to 28.3 on Terminal-Bench 3.0 and says the model already flagged a potentially serious vulnerability in Cursor, the coding tool recently acquired by SpaceX.
The bigger story is what happened alongside the coding gains. Working with security teams, Z.ai says GLM-5.3 turned up 2,436 vulnerability findings across 269 real software projects, including the Linux kernel and WebKit, with 1,097 rated critical or high severity after expert review. Z.ai says cyber capability developed faster than expected as training scaled, particularly the model's ability to chain a discovered flaw into a working exploit rather than just spotting it.
That is why API access and downloadable weights are being held back roughly two weeks for what the company calls safety hardening, even though the model is live now through Z.ai's GLM Coding Plan and ZCode tool. It is the first release in the GLM line to gate open weights behind extra review, a sign that the same skill making a model a better engineer is also making it a sharper security researcher, for better and for worse.
OpenAI dropped the price of GPT-5.6 Sol, the top tier of its three-model GPT-5.6 family, by more than a fifth on August 21. Sol now runs $4 per million input tokens and $20 per million output tokens, down from $5 and $30, with the promotional pricing holding through at least November 21.
The cut follows July's steeper reductions to the mid-tier Terra and budget Luna models, showing a pattern of frequent repricing rather than a one-time launch discount. It also lands the same week Google cut Gemini 3.7 Flash and DeepSeek raised its own API prices, underscoring that the cost of frontier-grade inference is now moving in opposite directions depending on which lab needs volume and which needs margin.
Google released Gemini 3.7 Flash on August 13, calling it its most capable workhorse model yet for coding and everyday agent work. It costs $0.75 per million input tokens and $3.75 per million output tokens through the end of 2026, half of what Gemini 3.6 Flash charged at launch just three weeks earlier.
On Google's own benchmarks the model scored 43.6% on FrontierCode 1.1 Main, up from 34.4%, and now powers Gemini Spark, the company's personal AI agent, in more than 160 countries. The cheaper rate roughly doubles on January 1, 2027, and it arrives while Google's larger Gemini 3.5 Pro flagship remains delayed with no new timeline, leaving February's Gemini 3.1 Pro as the company's newest large reasoning model.
Alibaba published open weights for Qwen3.8-Max on Hugging Face in mid-August, following through on a promise made when it first released the model earlier in the month. The system carries 2.4 trillion total parameters but activates only 95 billion per request, and the hosted version supports a 1-million-token context window across text, images, and video.
Alibaba had kept its most recent flagship Qwen models closed, so this marks its first public release at Max scale. A smaller companion, Qwen3.8-27B, shipped under an Apache 2.0 license and runs on a single ordinary GPU. The openly downloadable checkpoint is text-only, so developers wanting the full multimodal package still need Alibaba's paid API at $2 per million input tokens and $6 per million output tokens. Alibaba reports the model ranks fifth in Text Arena and second in Vision Arena, trailing mainly Anthropic's Claude line.
Anthropic began switching on Auto Mode by default for Claude Code starting August 14, letting the agent carry out multi-step coding tasks with less back-and-forth approval. A classifier reviews each tool call before it runs, clearing safe actions automatically and blocking or escalating anything that looks risky, such as mass deletions.
The change sits between Claude Code's old fully-interactive default, which stops for permission at nearly every step, and the bypass flag some developers were already using to skip permissions altogether. It follows months of steady additions to Claude Code, including automated code review and sandboxing rules, as coding agents from Anthropic, OpenAI, and Meta all move toward longer, less supervised task runs this year.