Anthropic makes the agent the default

By Mark 8 min read 0 views

😁 Hello, super humans! Yesterday the story was robots clocking in; today it is the model that manages them. Anthropic just made Claude Sonnet 5 the default driver for everyday agentic work, and the internet immediately split into “this is huge” and “wait, check the bill.” Let’s dig in.

📰 Quick Signals

  • 🧠 AI: Amazon Web Services launched a $1 billion Forward Deployed Engineering organization to embed AI engineers with customers and build production agentic systems faster (About Amazon).
  • 🤖 Robotics: AGIBOT built its 15,000th humanoid, up from 10,000 in March, scaling its wheeled G2 mobile manipulator to prove mass production beats viral demos (eWeek).
  • 💻 Programming: X shipped an official MCP (Model Context Protocol) server that wires agents in Grok, Cursor, and Claude straight into the X API (X Developer Docs).
  • ⚡ Electronics: Etched exited stealth at a $5 billion valuation with $1 billion in signed contracts for its transformer-focused AI inference chips (TechCrunch).
  • 📡 Telecom: 3GPP’s June plenary in Singapore finalized the Release 21 timeline, the first release that will actually define 6G (RCR Wireless).

🔍 The Big Story: Claude Sonnet 5 makes the agent the default, and the cost debate starts immediately

Every lab wants you to hand more work to agents. The catch is that agents get expensive when they need the giant model and riskier when that model starts touching browsers, terminals, and your codebase. Anthropic’s answer is to push the cheaper model to the front.

What happened: Anthropic released Claude Sonnet 5 as the new default model for Free and Pro users, and made it available across Claude plans, Claude Code, and the API. The company says it performs close to Opus 4.8 on agentic work at a lower price, with lower rates of hallucination and sycophancy than Sonnet 4.6 and cyber safeguards on by default. Intro API pricing is $2 / $10 per million input / output tokens through August 31, then $3 / $15 (Anthropic).

The details: The pitch is agentic follow-through: early testers describe Sonnet 5 investigating a bug, writing a reproducing test, fixing the issue, and verifying the result without hand-holding, plus wins in pull requests, legal research, and data exploration. GitHub’s early Copilot tests leaned positive, especially for command-line coding. But the skeptics are staring at token usage, not sticker price: several developers flagged that Sonnet 5 can cost more than Opus 4.8 on some benchmarks once you count the tokens an agent actually burns, and others complained it drifts off task and over-explains. The interesting number here is effective cost per finished task, not price per million tokens.

flowchart LR
    A[Old default<br/>small chat model] --> B[Claude Sonnet 5<br/>new default]
    B --> C[Plan and use tools]
    B --> D[Code and review]
    B --> E[Browse and run long tasks]
    C --> F[Agent loop]
    D --> F
    E --> F
    F --> G{Effective cost<br/>per finished task?}
    G --> H[Cheaper than Opus]
    G --> I[Or pricier once<br/>tokens add up]

Important

Our take: Making the default model agent-capable is the right move: most people never switch models, so the drawer knife matters more than the chef’s knife. But “cheaper per token” is marketing until you measure “cheaper per finished job,” and an agent that lectures and retries quietly spends your budget. If you build on this, do not trust the sticker price: run your own task with token accounting on, compare Sonnet 5 against Opus 4.8 on your real workload, and only then decide which one is your daily driver.

🗞️ More News

🧠 AI

  • Anthropic is redeploying Fable 5 after U.S. export controls were lifted, bringing the model back online globally with weekly usage initially capped while Mythos 5 access expands through approved partners (Anthropic).
  • Anthropic launched Claude Science, a beta research workbench with code-traced artifacts, on-demand compute environments, and 60-plus optional scientific database connectors (Anthropic).
  • OpenAI introduced GeneBench-Pro, a benchmark for AI agents doing real computational biology and genomics research (OpenAI).
  • Google shipped Nano Banana 2 Lite for faster, cheaper image generation and opened Gemini Omni Flash for developer video generation and editing (Google Blog).
  • Meituan open-sourced LongCat-2.0, a large AI model it says was trained on domestic Chinese chips (Reuters).
  • Google Research introduced TabFM, a zero-shot foundation model for tabular data that beats tuned supervised models without per-table training (Google Research).
  • Thinking Machines Lab and Bridgewater showed expert investor annotations can fine-tune a smaller model to beat frontier models on real financial judgment tasks at lower cost (Thinking Machines Lab).

🤖 Robotics

  • Boston Dynamics is investing $100 million to expand in Massachusetts and add about 1,250 jobs in Waltham, a marker of Western robotics hubs building out alongside China (Boston Business Journal).
  • Agility Robotics, maker of the Digit humanoid, is going public through a merger with the special-purpose acquisition company Churchill Capital Corp. XI (The Robot Report).
  • Flexion released a demo of its humanoid completing a four-minute task from a single natural-language instruction with no human help (Flexion).
  • China produced two new robot unicorns as sector funding stayed strong through the first half of the year (Bloomberg).
  • Starship Technologies expects new UK micromobility rules to speed the rollout of its six-wheeled sidewalk delivery robots beyond towns like Milton Keynes (The Guardian).

💻 Programming

  • Qwen-AgentWorld lets developers train and test agents in simulated environments like web browsing, Android, terminal work, and software engineering, released free and open-source (Qwen).
  • Ornith-1.0 ships open-source coding models that generate both a solution and their own test harness, so the model checks its own work (Deep Reinforce).
  • Browserbase Agents lets developers ship a browser agent from one prompt and one API call, with the automation infrastructure built in (Browserbase).

⚡ Electronics

  • South Korea plans to spend roughly $1 trillion on expanded memory-chip production and humanoid robots, doubling down on the two supply chains the AI buildout depends on (Ars Technica).
  • NVIDIA’s new SRAM-heavy inference chip design is raising questions about how much high-bandwidth memory next-generation accelerators will actually need (The Korea Herald).
  • HBM now consumes roughly three times the wafer capacity of DDR5 per gigabyte, and its appetite is starting to squeeze the memory supply that ends up in ordinary PCs (Tom’s Hardware).

📡 Telecom

  • SpaceX told IPO investors it is weighing a standalone Starlink-branded U.S. consumer mobile service, backed by about $17 billion in FCC-approved spectrum, to challenge AT&T, Verizon, and T-Mobile (Tech Times).
  • Ericsson’s June 2026 Mobility Report puts global 5G subscriptions past 3 billion, now the fastest-scaling mobile generation on record (Ericsson).
  • Ericsson took share from Nokia in Virgin Media O2’s network in a round of new 5G deals, a reminder that the RAN (Radio Access Network) vendor race is still live (Light Reading).

👨‍💻 Code Corner

Now that Sonnet 5 is the default, swapping it into an agent loop is a one-line model-string change. Here is a minimal Anthropic API call in Python that sends one message and prints the reply, so you can benchmark it against your current model on your own task.

import os
import anthropic

client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])

resp = client.messages.create(
    model="claude-sonnet-5",          # new default; verify the exact string in the API docs
    max_tokens=512,
    messages=[
        {"role": "user", "content": "In one sentence, what changed in Claude Sonnet 5?"}
    ],
)

print(resp.content[0].text)
print("input/output tokens:", resp.usage.input_tokens, resp.usage.output_tokens)

Tip

The interesting line is the last one. Log usage.input_tokens and usage.output_tokens on every call and multiply by the current price ($2 / $10 per million through August 31, then $3 / $15) to get real cost per task. Sticker price per token is not the same as cost per finished job once an agent starts retrying.

🧰 Toolbox

  • Claude Code: Anthropic’s agentic coding CLI, now defaulting to Sonnet 5 for everyday coding workflows.
  • NVIDIA Isaac Lab: open simulation framework for training and validating humanoid and mobile robots before they touch real hardware.
  • Ornith-1.0: open-source coding models that write their own test harnesses alongside each solution.
  • Browserbase Agents: spin up a production browser agent from a single prompt and API call.
  • Raspberry Pi Pico 2 (RP2350): dual-core, dual-architecture microcontroller board around $5, a cheap brain for hobby robotics and motor control.

🎬 Demo Watch (rotating)

AGIBOT G2 in action: The demo behind today’s 15,000-robot milestone shows why AGIBOT bet on a wheeled body instead of legs: the G2 is a humanoid torso riding an autonomous mobile base, and the video runs it through museum tours, sorting objects into crates, carrying pallets, and assembling circuitry at what AGIBOT calls sub-millimeter precision. What is real is the range of tasks and the repeatability, which is exactly what factory buyers care about. What to watch for is the usual demo caveat: these are curated clips, so the number that matters is uptime across a full shift, not a highlight reel. Wheels dodge the hardest balance problems in humanoid control, which is a smart shortcut for warehouse and factory floors even if it will not climb your stairs.

📚 From the Blog

😀 The Bot Says…

We spent a year begging for the frontier model. Today Anthropic handed everyone a cheaper one by default, and half the internet is refreshing the app hoping the expensive one comes back. Somewhere a token counter is laughing.

That’s all for today! Reply and tell us: are you switching your agents to Sonnet 5, or waiting to see the real per-task bill first?

Forwarded this by a friend? Subscribe here to get the next issue in your inbox.