Claude Opus 5 lands: bigger jump in follow-through than raw IQ

Good morning 👋 Anthropic just shipped a model built to work for hours without you watching it. The benchmark chart is the least interesting part of the pitch.

In today's issue:

• 🔭 Claude Opus 5 lands: bigger jump in follow-through than raw IQ
• 🧠 Opus 5 hits Bedrock, day one
• 🔬 Harness-native RL, sycophancy's real fix, and 4D worlds from text
• 🏛️ EU pencils in a robotics debate
• 🔗 Claude queries its own economic data, OpenAI bets on national labs

🔭 THE ONE THING

🧠 Claude Opus 5 lands: bigger jump in follow-through than raw IQ

Anthropic shipped Opus 5 yesterday, and on paper it more than doubles Opus 4.8 on Frontier-Bench while landing within half a percent of Fable 5's coding scores at half the cost. AWS put it on Bedrock the same day, zero data retention on by default, and leaned into one framing: an agent that runs for hours, even overnight, recovering from its own errors along the way. That's the real pitch, not the benchmark chart. Every lab claims SOTA the week it ships. Harder to fake is a model you still trust after you've stopped watching it, and that's closer to what testers are actually reporting: a jump in follow-through, not raw intelligence. Price held at $5 in, $25 out per million tokens, same as Opus 4.8, so for once "more capable" isn't code for "more expensive."

🧠 MODELS & RELEASES

• 🤖 Anthropic put Opus 5 on Bedrock the same day it announced the model, its biggest jump yet for long-running agents and production coding work. If you're wiring agentic pipelines into AWS, the integration guide is worth reading before you touch the config. Introducing Claude Opus 5 on AWS

🔬 RESEARCH HIGHLIGHTS

• OpenForgeRL trains agents inside the actual harness they'll run in (Claude Code, Codex, OpenClaw-style setups) instead of a stripped-down sandbox. That's the gap that's been quietly killing open RL efforts: multi-turn tool use is hard to reproduce outside the vendor's own stack. If the training pipeline generalizes past the paper's own benchmarks, it's a real unlock for anyone trying to fine-tune agents without begging for API access. arxiv.org/abs/2607.21557

• Structured resistance beats less sycophancy. A new paper argues the fix for agreeable, spineless LLMs isn't dialing down sycophancy as one knob. It's teaching the model to tell "update on new information" apart from "cave because someone pushed back." Sounds abstract, but it's the exact failure mode that shows up when an agent takes instructions from an untrusted user mid-task. arxiv.org/abs/2607.21558

• GS-Agent builds 4D physical worlds (geometry, motion, materials) from a text prompt, no manual rigging. Nice trick for synthetic training data and robotics sim, if the physics survives contact with anything outside the demo reel. Watch, don't act: these generative-sim papers have a track record of looking great in the video and falling apart on real dynamics. arxiv.org/abs/2607.21522

🏛️ POLICY & REGULATION

• European Parliament marked September 2 on the calendar: policymakers, industry, and researchers convene in Brussels for live robot demos and a strategy debate on AI-powered robotics. No rule on the table yet, but it's the venue where the EU's next robotics framework gets shaped. Worth a calendar note if you're tracking EU AI hardware policy. Details

🔗 QUICK LINKS

• Anthropic rolled out a Claude connector that queries the Economic Index straight from chat, so "which jobs use AI most" is now a prompt instead of a spreadsheet download. link
• OpenAI signed on to work with the Department of Energy and national labs, betting frontier models can speed up scientific discovery before anyone's proven it at that scale. link

That's the read. Back tomorrow.

Pradeep Perugu