Anthropic releases Claude Opus 5 at unchanged pricing
Anthropic launched Claude Opus 5 on July 24, 2026, pricing it the same as its predecessor Opus 4.8 at $5 per million input tokens and $25 per million output tokens. The company reports state-of-the-art results on coding and professional-work evaluations such as Frontier-Bench and GDPval-AA, and roughly double Opus 4.8's performance on several of them. It becomes the default model on the Claude Max plan and the strongest on Pro. Artificial Analysis listed it first on its Intelligence Leaderboard soon after release.
Sources & depth
UK and US government labs rate Kimi K3 well below frontier on cyber tasks
The UK AI Security Institute and the US CAISI published a joint preliminary assessment of Moonshot AI's open-weight Kimi K3. They found it performs significantly below the most recent frontier cyber-capable models, but ahead of GLM-5.2, previously the strongest open-weight option on the same tests. In a simulated 32-step corporate network attack it completed the full chain in one of ten attempts and averaged 17 steps, against 28.5 for the leading US models. The evaluators also noted that the model's safeguards did not stop it from assisting offensive cyber work.
Sources & depth
Washington's open-weight fight: industry letters and startups push back on restrictions
A wave of lobbying surfaced in the week of July 22 over whether the US should restrict open-weight AI, particularly Chinese models. Nvidia, Microsoft and Meta issued a joint letter, "Open Weights and American AI Leadership", cautioning against overregulating open-weight models, while a group of startup founders urged the administration not to cut off access to Chinese open-weight models. Axios framed OpenAI and Anthropic as aligned on the other side, citing risks to their business. Commentary such as Tobi Knaup's essay argues open-weight models are becoming a Kubernetes-style shared platform, and that the US should compete by releasing frontier-grade open models, using procurement strategically and setting independent standards rather than banning rivals. Details of the letters themselves: unfetchable — link only.
Sources & depth
Echo claims frontier-level results at a third of the cost by routing across open-weight models
Echo, from Tracer ML, is an experiment in building one system from a pool of open-weight models, including GLM-5.2 and Kimi K2.7, instead of using a single model for every task. It decides per request how much compute to spend, which models take part and how their outputs combine. The creator reports that its evaluation dashboard shows roughly Fable-level performance at about one-third of the inference cost, while acknowledging remaining routing failures and that coding and agentic tasks are still being tested. These are self-reported results.
Sources & depth
Palmier Pro: a macOS video editor with built-in generation and an agent interface
Palmier Pro is a Swift-native video editor for Apple Silicon Macs running macOS 26 or later. It includes built-in generative video models such as Seedance and Kling, and a local MCP server that lets AI agents work directly on the editing timeline. Releases up to v0.7.6 are open source under GPLv3, while newer binaries are proprietary.
Sources & depth