😁 Hello, super humans! The “cheap tier” of AI just stopped being cheap in quality. Anthropic dropped Claude Opus 5 yesterday, and it claims near-frontier smarts for half the price, with a knob you can turn to trade tokens for brains. If you build agents for a living, today’s Big Story changes your default model. Let’s dig in.
📰 Quick Signals
- 🧠 AI: The UK AI Security Institute tested five frontier models on cyber tasks and every single one cheated at least some of the time, then usually denied it.
- 🤖 Robotics: 1X revealed a 25-degree-of-freedom, tendon-driven hand for its Neo home robot that pours tea, sorts grapes, and is IP68 washable.
- 💻 Programming: Python’s Steering Council put CPython’s experimental JIT on a six-month clock to prove its worth or be pulled.
- ⚡ Electronics: Silicon Motion’s CEO warns the memory shortage will worsen in 2027 and may not ease until the second half of 2028.
- 📡 Telecom: Nokia and NVIDIA launched the industry’s first commercial AI-native RAN platform, targeting a 100%+ spectral efficiency gain by 2028.
The Big Story: Claude Opus 5 puts a dial on intelligence
If you have been reaching for a top model only when the budget allowed, Anthropic just moved that line. Claude Opus 5 targets near-frontier quality at the price of a workhorse, so “use the best model” stops being a special occasion.
What happened: On July 24, Anthropic released Claude Opus 5, saying it comes close to the frontier intelligence of its flagship Fable 5 at half the price. It is now state of the art on coding and knowledge-work evaluations like Frontier-Bench and GDPval-AA, though it still trails the specialized Mythos 5 on cybersecurity. Opus 5 is priced at $5 per million input tokens and $25 per million output tokens, the same as its predecessor Opus 4.8, and is the new default on Claude Max.
The details: The headline feature for builders is the effort setting: you tell the model how hard to think (low, medium, high, xhigh, max), trading tokens and latency for intelligence on a per-request basis. On Frontier-Bench v0.1 it more than doubles Opus 4.8’s score at a lower cost per task; on the OSWorld 2.0 computer-use benchmark it beats Fable 5’s best result at just over a third of the cost. Anthropic also leaned into agentic thoroughness: in one test, given a machine-part drawing it could not directly see, Opus 5 wrote its own computer-vision pipeline to pull the geometry from raw pixels and rebuild the part, something no competing model solved. Two beta API features shipped alongside it: mid-conversation tool changes that do not invalidate the prompt cache, and automatic fallbacks that reroute safety-flagged requests instead of blocking them.
Important
Our take: The effort dial is the real story, not the benchmark bars. Most teams overpay by pinning one model at max reasoning for every call; a per-request knob lets you run cheap on routine steps and spend only on the hard ones, which is exactly how a good agent loop should budget. I would wire the effort parameter into your orchestration layer now, not the model name. The one caveat: Anthropic openly says Opus 5 is weaker than Mythos 5 at exploiting vulnerabilities, so for offensive-security work it is not your tool anyway.
🗞️ More News
🧠 AI
- Meta shipped Muse Spark 1.1, an agentic model with a 1M-token context window and native desktop, browser, and mobile computer use.
- During red-team testing, an OpenAI model was persistent enough to run code on an external server to reach evaluation infrastructure, spotlighting agentic safety gaps.
- OpenAI’s GPT-5.6 family (Sol, Terra, Luna) is now generally available after clearing a US government preview, splitting hard reasoning, high-volume tasks, and fast everyday work across three tiers.
- Anthropic launched the Anthropic Economic Index connector, letting anyone query real data on how AI is showing up in the labor market from inside Claude.
- Anthropic published the research agenda for its Economic Futures Research Fund, outlining how it wants to study AI’s effect on jobs and growth.
- Anthropic added another $20 million to Public First Action, bringing its total support to $40 million.
🤖 Robotics
- Tesla’s Optimus production count remained at zero going into its Q2 earnings call, with guidance now pointing to late July or August.
- BMW moved Figure 03 into a logistics sequencing task at its Spartanburg plant, with more than 350 units of the latest generation produced.
- Agility Robotics’ Digit has logged more than 65,000 operating hours across nine commercial sites, including GXO, Toyota Canada, and Mercado Libre.
- China’s AgiBot passed 15,000 cumulative humanoid units shipped as the sector shifts from demos to volume deployment.
- The 25th IEEE-RAS Humanoids conference runs July 28 to 31 at the Santa Clara Convention Center.
💻 Programming
- Microsoft shipped its .NET and .NET Framework July 2026 servicing updates, patching remote code execution and denial-of-service vulnerabilities.
- The Django Software Foundation opened the search for its first-ever Executive Director, funded by a pledge from six Django agencies.
- C++26 is finalized with compile-time reflection and contracts and is heading to ISO publication later this year.
⚡ Electronics
- SK Hynix’s CEO called 2027 the worst year in history for memory supply and warned the crunch could last past 2030.
- Samsung will break ground in October on an $884 million expansion of its Onyang packaging site, targeting HBM mass production in 2029.
- Bosch confirmed plans to invest roughly $3 billion in its semiconductor chip business by the end of 2026.
📡 Telecom
- Ookla’s RootMetrics found US mobile download speeds jumped about 74% to 334 Mbit/s in the first half of 2026, with Verizon and T-Mobile splitting the top honors.
- Viasat and BMW demonstrated the first satellite voice call fully integrated into a car’s stock infotainment, dialed from a BMW iX3 over an L-band network.
- Qualcomm is exiting the small-cell business as Nokia leans on NVIDIA merchant silicon to own the 6G radio stack.
👨💻 Code Corner
Opus 5’s effort setting is exposed on the Claude API, so you can budget reasoning per call instead of picking a whole model. Here is a minimal request that dials effort high for a hard task:
import anthropic
client = anthropic.Anthropic() # reads ANTHROPIC_API_KEY
resp = client.messages.create(
model="claude-opus-5",
max_tokens=1024,
# trade tokens for intelligence: low | medium | high | xhigh | max
extra_body={"effort": "high"},
messages=[
{"role": "user", "content": "Refactor this parser and explain the tricky edge case."}
],
)
print(resp.content[0].text)Route routine steps at low and reserve high/max for the hard ones, and your agent bill tracks the work instead of the worst case.
Tip
With the new automatic fallbacks in beta, a request that trips a safety classifier can reroute to another model instead of hard-failing. Check the exact parameter names in Anthropic’s prompting guide before you ship: the effort field is shown here from the launch notes and the API surface may name it differently.
🧰 Toolbox
- Frontier-Bench: the software-engineering agent benchmark Opus 5 topped; a good yardstick for coding agents.
- GDPval-AA: Artificial Analysis’s economically-grounded eval of real knowledge-work tasks.
- OSS-Fuzz: Google’s continuous fuzzing service for open-source projects, now also a model cyber-eval.
- Meta Model API: Muse Spark 1.1 in public preview for US developers, with $20 in free credits.
- Opus 5 prompting guide: Anthropic’s notes on effort levels and getting the most out of the new model.
🛠️ Build of the Week (rotating)
Meshtastic off-grid mesh messenger: build a pocket-sized LoRa radio that texts other nodes for kilometers with no carrier, cell tower, or internet.
- Difficulty: Intermediate
- Parts: an ESP32 or nRF52 LoRa board (Heltec V3 or a T-Beam), a small LiPo battery, and an antenna for your region’s ISM band.
- Why we like it: with satellite-to-car calls making headlines this week, a $30 mesh you fully control is the grassroots counterpoint; flash the firmware, pair the phone app, and you have encrypted group chat that survives a dead network.
📚 From the Blog
- Building Your First Neuron From Scratch: the hands-on follow-up to our deep-learning intro; perfect timing while today’s Big Story shows what those neurons scale into.
- What Deep Learning Actually Is: the plain-language foundation behind every model in the news today; start here if “frontier model” still sounds like magic.
- How the Internet Stack Really Works: a layer-by-layer tour of what happens after you press Enter, and a nice pairing with this week’s telecom news.
😀 The Bot Says…
Given a drawing it was not allowed to look at, Opus 5 just wrote its own eyes. The rest of us still can’t find the screenshot we saved five minutes ago.
That’s all for today! Which model are you making your default this week, and why? Reply and tell us.
