Claude Opus 5 arrives, Gemini 3.6 Flash and Solar Open 2 follow

Anthropic had the week. The rest of it was Google keeping its Flash line moving and two open models from outside the usual US and China pair.

Claude Opus 5

Opus 5 landed on 24 July. Anthropic describes it as its flagship for complex agentic coding, enterprise work, long-horizon reasoning, computer use and professional analysis. The specs are generous: a 1M-token context window, up to 128K tokens of output in one response, text and image input, tool use, structured outputs and prompt caching. List price is $5 per million input tokens and $25 per million output.

The part that changes how I work is the effort control. Adaptive thinking is on by default, and a request can pick one of five effort levels from low to max. Those settings change how much the model thinks and how many tokens it spends, not how long the visible answer is. For anyone running many agents at once, that is a dial on cost per task. I'd rather route by effort than by switching models, because the prompts and tool setups stay the same.

Gemini 3.6 Flash

Google released Gemini 3.6 Flash on 21 July, together with Gemini 3.5 Flash-Lite and a gated Gemini 3.5 Flash Cyber. Flash takes text, image, video, audio and PDF input with a 1,048,576-token window and up to 65,536 output tokens. Pricing on the routes I looked at was $0.75 in and $3.75 out per million tokens. Google's guidance is simple: pick Flash when capability matters, Flash-Lite when cost and throughput matter more.

Two open models worth a look

Upstage released Solar Open 2, 250B total and 15B active parameters, with a 1M-token window and a design aimed at office documents and coding agents. The license is Apache 2.0 based with a branding requirement for derivatives.

The Swiss AI Initiative, working with ETH Zurich, EPFL and CSCS, released Apertus 1.5 70B on 24 July. What stands out is how open it is: weights, training data and recipes are public under Apache 2.0.

Pictures, worlds and policy

Alibaba's Qwen-Image-3.0-Pro went live on 20 July as a hosted API only. ABot-World-0 is a research paper on a video world model that runs long interactive rollouts on one desktop GPU. And Google announced $40 million in tokens and credits for the US Department of Energy's Genesis Mission, including access to DeepMind's science tools.

What I'd watch next

Whether anyone releases open weights that match the effort-dial idea, a model you can turn up or down per request, because that is the feature I'd actually pay for.