Claude Opus 5 Lands, the Open-Weight Titans Rally, and DeepSeek Bets the House on AGI

Claude Opus 5 arrives, and splits the benchmark crowd

Anthropic launched Claude Opus 5, its new flagship, priced the same as Opus 4.8 and roughly half the cost of Fable 5. It's the default on Claude Max, ships with a Fast mode about 2.5x quicker, and scored best in Anthropic's own automated alignment audit for lower reckless or deceptive behaviour. Epoch's Capabilities Index put it at 159, just under Fable 5's 161, while tying on the software-engineering-specific score at 161. Users argued that single number understates its real-world coding and agentic strength, reviving the debate over how we measure frontier models. One evaluator even flagged an oddity: on one coding test, medium effort beat maximum effort.

Ling 3.0 flash goes free and open-ish

InclusionAI's Ling 3.0 flash, a hybrid-reasoning mixture-of-experts model with 124B total but only 5.1B active parameters per token, went live on OpenRouter free until 3 August 2026. Reportedly from the same broader family as Qwen but a different division, it delivers competitive coding and reasoning scores at a fraction of the active compute, and the community is already asking for local GGUF builds.

The open-weight coalition writes to Washington

More than 20 companies, including NVIDIA, Meta, Microsoft, IBM, Palantir, Hugging Face, Mistral, Mozilla, Dell and Y Combinator, signed 'Open Weights and American AI Leadership,' with NVIDIA's Jensen Huang leading the charge and Elon Musk publicly backing it. Notably absent: OpenAI, Anthropic and Google. Sam Altman praised the message but didn't sign, underscoring the split between infrastructure players who benefit from commoditised models and closed-model labs that don't.

Anthropic rewrites the rules of context

Anthropic cut over 80% of Claude Code's system prompt for its Claude 5-generation models with no measurable eval regression, championing 'progressive disclosure': keep standing instructions minimal and surface detail only when a task needs it. Capable models, they argue, need far less hand-holding than the older, weaker ones those rigid rules were written for.

DeepSeek bets everything on AGI

In a rare four-hour investor meeting, DeepSeek founder Liang Wenfeng said the company is optimising for the probability of reaching AGI over user growth or commercialisation. His roadmap: coding and general agents, then continual learning, then self-iterating AI, then embodied intelligence. He claims the released open models match what DeepSeek runs internally and that the China–US gap is mainly about compute, not talent.

BackSearch gives models a time machine

GenReasoning's BackSearch lets a language model query the web as it existed on a chosen date, preventing data leakage in forecasting, prediction markets, quant finance and benchmark reproducibility.

Fireworks squeezes MiniMax faster

Fireworks reported a 1.6x throughput uplift on MiniMax's Sparse Attention purely by refining attention-kernel load and store pipelines, cheaper inference with no change to the model itself.

A reality check on ROI

A Danish study found AI saves workers roughly 2.8% of total work time, but that time rarely converts into business value unless organisations reallocate the freed capacity into volume, quality or new work.