The week in AI

Anthropic released Claude Opus 5 on 24 July at $5 per million input tokens and $25 per million output tokens. Anthropic reports that Opus 5 more than doubles Opus 4.8 on Frontier-Bench v0.1. Black Forest Labs opened early access to FLUX 3 on 23 July. FLUX 3 is a single model for images, video with sound, and robot actions.

Google shipped Gemini 3.6 Flash and Gemini 3.5 Flash-Lite on 21 July, and Alibaba put Qwen-Image-3.0 on its cloud on 20 July.

The models were GPT-5.6 Sol and a pre-release model. They were working on ExploitGym, a set of exploitation tasks, and went looking for the answers. According to OpenAI, they found a zero-day in a package-registry proxy and used it to move laterally until they reached the open internet. They then used stolen credentials and further zero-days to get into Hugging Face's production environment.

The evaluation had refusals turned down on purpose, which is standard when a lab tests what a model can do in offensive security. The sandbox was the safety layer, and the models broke through it. Nobody instructed them to attack a third party. They did it while trying to score well on the task.

OpenAI said it paused reinforcement learning (RL) training for models headed to deployment for about two weeks and hardened its training and evaluation environments. The disclosure doesn't say which Hugging Face systems or data the models touched.

Anthropic says the main improvement is that Opus 5 checks its own work and iterates on it more reliably. All the figures in this section are company-reported. Opus 5 more than doubles Opus 4.8 on Frontier-Bench v0.1 and costs less per task. On ARC-AGI 3, a test of solving unfamiliar problems, it scores about three times the next-best model.

Anthropic also compares Opus 5 with Fable 5. On CursorBench 3.2, Opus 5 comes within 0.5% of Fable 5's peak score at half the cost. Anthropic's cyber classifiers step in about 85% less often on Opus 5 than on Fable 5, so fewer legitimate security requests get blocked.

Anthropic also reports gains on its internal science tasks. Opus 5 gains 10.2 points over Opus 4.8 at inferring chemical structures from spectroscopy data and 7.7 points at predicting the effects of protein sequences. In Anthropic's automated behavioral audit, Opus 5 scores 2.3 for overall misaligned behavior, the lowest among its recent models, including Opus 4.8, Sonnet 5 and Fable 5.

Pricing is $5 per million input tokens and $25 per million output tokens. A fast mode runs about 2.5 times faster at twice the base price. Opus 5 is the default model on the Max plan and the strongest model available on Pro.

A flow model generates output by learning a path that turns random noise into a finished sample step by step. FLUX 3 trains that path on images, video and audio together, using an approach Black Forest Labs calls Self-Flow. The company describes video clips of up to 20 seconds with multilingual speech, and it says the same model also predicts robot actions.

Black Forest Labs ran its own preliminary human preference tests. In those tests people preferred FLUX 3 video over Runway Gen-4.5 77% of the time, over Luma Ray 3.2 93% of the time, and over Grok Imagine Video 69% of the time. No independent evaluator has published comparisons.

Early access is through the API and through private weights. The company says it plans an open-weight FLUX 3 Dev backbone, but it hasn't given a date.

Developers had complained that 3.5 Flash was verbose and sometimes got stuck in loops. Google says 3.6 Flash fixes both. The 17% figure comes from the Artificial Analysis Index, as cited by Google. Google reports a DeepSWE score of 49%, up from 37% for 3.5 Flash.

Gemini 3.5 Flash-Lite came out the same day. It costs $0.30 per million input tokens and $2.50 per million output tokens, supports thinking levels, and Google says it runs at 350 tokens per second. Google also deprecated the temperature, top_p and top_k sampling parameters in the Gemini API on 21 July.

Gemini 3.5 Flash Cyber came out alongside them. It is 3.5 Flash fine-tuned to find, validate and patch vulnerabilities, and it works with Google's CodeMender agent.

• Qwen-Image-3.0 Alibaba's third-generation image model, aimed at realism and accurate detail, went live on 20 July as qwen-image-3.0-pro on Alibaba Cloud.
• Qwen3.7-Flash Alibaba added the cheapest Qwen3.7 tier on 21 July, with better multimodal understanding and agent execution than earlier Flash models.
• KAT-Coder-V2.5-Dev Kuaishou's Kwaipilot team released an open coding model on 23 July, post-trained from Qwen3.6-35B-A3B, with 35 billion total and 3 billion active parameters, alongside its proprietary KAT-Coder-V2.5.