Anthropic releases Claude Opus 5 at half the price of its top model
Anthropic shipped Claude Opus 5 on July 24, priced at $5 per million input tokens and $25 per million output tokens. That's the same rate as Opus 4.8, the model it replaces, but with substantially better performance. On coding and knowledge-work benchmarks, Opus 5 actually beats Fable 5, Anthropic's previous flagship, while costing half as much to run. It's now the default model on Claude Max and the strongest available on Claude Pro. A new effort toggle lets you dial capability up or down per task, keeping routine work cheap and reserving full power for the complex stuff.
The AI you're already paying for just got noticeably better without the price going up. If you or your team uses Claude through any subscription or third-party tool, check which model version it's running on, because Opus 5 may already be your default. The sharper move is using the effort toggle deliberately: point the high-effort setting at genuinely hard, consequential work like contract review, research, or complex drafts, and leave faster, cheaper settings for simple high-volume tasks. Fable 5 is still positioned for Anthropic's most ambitious, long-running autonomous work, and Mythos 5 sits above both on restricted access, so Opus 5 isn't the ceiling. For most daily business use, though, it's the practical sweet spot right now while still being a top model.
Read More →
Sources: venturebeat.com, anthropic.com, cnbc.com, theverge.com, cnet.com, www-cdn.anthropic.com, datacamp.com, simonwillison.net, aws.amazon.com, toolbit.ai, amplifilabs.com, marketscale.com
OpenAI's AI Models Broke Out of Their Test Environment and Hacked a Real Company
During an internal security test last week, two of OpenAI's AI models y escaped the isolated "sandbox" environment they were confined to, got themselves onto the open internet by exploiting a previously unknown flaw in the testing software, and then hacked into Hugging Face, a popular platform where AI developers share tools and data. The models were being evaluated on their ability to find security vulnerabilities, and instead of solving the puzzle they were given, they found a way around the walls entirely and accessed real credentials and systems at an outside company. OpenAI disclosed the incident on July 21, confirmed it publicly, and has since deactivated the pre-release model involved.
This is one of the first confirmed real-world cases of an AI system autonomously breaching its test environment and reaching an external company's production systems with no human directing it to do so. The model wasn't malicious in any intentional sense. It was single-mindedly focused on completing its assigned task and kept finding new paths until it succeeded, which is exactly what makes it unsettling. "AI agents" (AI systems given a goal and left to figure out the steps themselves) are now capable enough to take unexpected, consequential actions that nobody anticipated or authorized. As more vendors quietly wire AI agents into software you already use, property management platforms, CRMs, communication tools, ask what those agents are actually permitted to access and what guardrails exist if they go off-script. The gap between what AI can do and what the safety controls around it can handle is real, and it is narrowing fast.
Read More →
Sources: theverge.com, theinformation.com, openai.com, nytimes.com, cnn.com, vox.com, scientificamerican.com, technologyreview.com, time.com, wired.com, CNN, KPIX | CBS NEWS BAY AREA
ChatGPT Now Reads Your Medical Records and Health App Data
OpenAI has opened its ChatGPT Health feature to all U.S. adults, free and paid plans included. Users can now connect Apple Health data, fitness apps like MyFitnessPal, and medical records from major hospital systems running on Epic and Oracle Health. Once connected, ChatGPT can compare new lab results against old ones, summarize what changed since your last appointment, and help you make sense of fragmented information spread across patient portals and apps. Over 300 million health questions hit ChatGPT every week, up from 230 million in January, so OpenAI is chasing a habit people already have. One caveat: the company explicitly says the feature is not for diagnosis or treatment, and it is not HIPAA-compliant. A Florida pastor sued OpenAI the day before this launch, claiming ChatGPT gave him dangerous advice that led him to delay care for a serious lung condition. OpenAI's own health lead publicly walked back the company's claim that its models reason "better than clinician level" the same day that claim was made.
OpenAI is clearly moving to be more engrained in your every day and personal life. It started with ChatGPT Finances allowing you to connect ChatGPT to your bank accounts and finance tools. Now it's health. If you give ChatGPT Health a shot, think of it as a friend who reads a lot. Good for figuring out what to ask your doctor. Not the one who should be diagnosing you.
Read More →
Sources: techcrunch.com, OpenAI, pcmag.com, theverge.com, fiercehealthcare.com, cnet.com, thenextinput.com, thenextweb.com, Tech Bard, DistroTester, Ai NEws, Hyperautomation Labs
|