🎯 Top 3 Things to Know
1. Anthropic released Claude Sonnet 5, a cheaper model aimed squarely at running agents. The pitch is cost. Sonnet 5 closes most of the gap to Anthropic's top model, Opus 4.8, while charging a fraction of the price, which matters for agentic workloads that fan out into thousands of model calls per task. It beats the prior Sonnet 4.6 on every published benchmark: 63.2% on SWE-bench Pro (agentic coding), 81.2% on OSWorld-Verified, 57.4% on Humanity's Last Exam. Opus 4.8 still leads on the hardest coding tasks, but the margin is now small enough that teams running high-volume agents have a real reason to trade down. It is now the default model on free and Pro plans. One catch worth measuring before you migrate: Sonnet 5 ships a new tokenizer that produces roughly 1.0 to 1.35 times more tokens for the same input, which offsets some of the headline per-token discount. TechCrunch
2. Introductory pricing on Sonnet 5 runs only through August 31, then rises. Anthropic set launch pricing at $2 per million input tokens and $10 per million output, holding until August 31, after which it moves to $3 and $15. That is still cheaper than Opus 4.8, but the introductory window makes near-term cost math misleading for anyone modeling a production budget past the summer. Teams pricing an agent deployment should run the numbers at the post-August rate, not the launch rate, and factor in the tokenizer change above. Prompt caching (cache reads at 0.1x input) and the 50% Batch API discount still apply. VentureBeat
3. The EU published its final Code of Practice on labeling AI-generated content, weeks before the transparency rules bite. The European Commission released the Code of Practice on Transparency of AI-Generated Content on June 10, giving providers a concrete method for marking AI-made audio, image, video, and text, and deployers a standard for labeling deepfakes. The timing is the point: Article 50 transparency obligations become applicable on August 2. Adherence to the code is voluntary, but it is the clearest signal yet of what regulators will expect. Any team shipping generative output into the EU should check whether its watermarking and disclosure meet the code's machine-readable marking bar before the August date. European Commission
🚀 Frontier Models & Features
Claude Sonnet 5 ships as Anthropic's default. Covered above. The short version: a cheaper, more agentic Sonnet that narrows the distance to Opus 4.8 on coding and computer-use benchmarks, now the default for free and Pro users. The New Stack
OpenAI previewed GPT-5.6 "Sol" earlier in the week. OpenAI posted a preview and system card for a next model in the GPT-5.6 line ahead of a full release. Details remain thin pending the model card, so treat capability claims as unconfirmed. OpenAI
🔬 Research Worth Reading
- Masking Stale Observations Helps Search Agents — Until It Doesn't: A Regime Map and Its Mechanism (Zhang, Xu et al. / UC San Diego, McAuley lab). arXiv
- TL;DR: Long-horizon search agents pile up retrieved text across many tool calls. This paper maps exactly when dropping old ("stale") observations from context helps versus hurts, and explains the mechanism behind it.
- Stat: The accuracy gain from masking follows an asymmetric inverted-U: a plateau under weak retrievers, a peak when a strong retriever meets a mid-capacity model, and a sharp collapse once the model is already saturated.
- Apply it: Before adding observation-masking to your agent's context manager, check where your retriever-plus-model pairing sits on that curve. If you are already running a strong model with good retrieval, masking may cost you accuracy rather than save budget. Benchmark both ways on a held-out task set.
🏢 Enterprise in the Wild
General Mills reports more than $20 million in savings since fiscal 2024 from an AI-driven supply-chain optimization system, according to enterprise deployment roundups published this week. The figure is company-reported and spans roughly two years, so read it as cumulative, not annual. AI Monk
🛠️ Tooling & Ecosystem
Quiet day on this front. No major framework, SDK, or MCP server releases traced to a primary source in the last 24 hours.
⚖️ Policy & Regulation
EU AI Act transparency rules take effect August 2. The June 10 Code of Practice (above) is the compliance-facing companion to Article 50, which requires that AI-generated content be marked and that deepfakes be labeled. The code is voluntary, but the underlying obligation is not. artificialintelligenceact.eu
📌 Watch List
- Cost-per-quality math for agent workloads as cheaper "closes-the-gap" models arrive.
- Tokenizer changes that quietly shift real cost even when per-token prices drop.
- Context management for long-horizon agents: when to mask, summarize, or retain.
- EU AI Act Article 50 transparency: August 2 applicability date.
- Agent evaluation on continuous, multi-call tasks rather than single-shot benchmarks.