PostCards
Anthropic Ships Claude Opus 5 at Half the Price of Fable 5
Anthropic released Claude Opus 5 on July 24, priced at $5 per million input tokens and $25 per million output tokens — half of what Fable 5 costs. It scores 43.3% on Frontier-Bench v0.1 against Fable 5's 33.7% and Opus 4.8's 18.7%, and wins five of nine head-to-head benchmarks against Fable 5, including a 10.2-point jump over Opus 4.8 on organic-chemistry tasks.
Anthropic says its cyber classifiers intervene roughly 85% less often than they did with Fable 5, calling Opus 5 its most-aligned model yet. The company moved fast: Opus 5 is already the new default on Claude Max, the API, Claude Code and Claude Cowork, the same day it was announced.
The launch lands one day after AMD's $5 billion, 2-gigawatt compute pact with Anthropic and the same week Alphabet posted a blowout cloud quarter — a reminder that pricing discipline, not just raw capability, is now the lever labs pull to hold enterprise customers.
Delhi High Court Sides With OpenAI, Denies ANI's Injunction Bid
Justice Amit Bansal of the Delhi High Court on July 24 denied Asian News International's bid for an interim injunction against OpenAI, ruling that ChatGPT's training on ANI's content falls under "fair dealing" per Section 52(1)(a) of India's Copyright Act.
The court found ChatGPT's outputs are not substantially similar to ANI's original reporting, and that granting the injunction would cause "irreparable injury" to OpenAI and to public access to the tool. The underlying suit continues, and ANI can still pursue damages at trial.
But the interim ruling materially weakens the case for training-data liability in India's largest AI copyright dispute so far — a rare bit of home-jurisdiction clarity for builders working out of India on exactly where the law currently draws the line.
A Security Firm Finds an Unpatched Escape Hatch in Claude Cowork's Sandbox
Security firm Accomplish AI disclosed on July 23 that Claude Cowork's Linux VM sandbox on macOS can be escaped in a single crafted message, handing the agent full read and write access to the host machine's filesystem — including SSH keys and cloud credentials.
The chain, dubbed "SharedRoot," combines unprivileged user namespaces with CVE-2026-46331 in the guest kernel. Anthropic closed the report as "informative" rather than shipping a patch, telling users instead to rely on the cloud-execution default rather than local sandboxing.
For anyone running Cowork locally against real client files, it's a reminder that an agent sandbox is only as strong as its weakest kernel primitive — worth a second look before the next local run.
Meta, Nvidia and a Dozen Others Ask Washington Not to Restrict Open-Weight AI
Meta, Microsoft, Nvidia, IBM, Palantir, Hugging Face, Perplexity, Mistral, Andreessen Horowitz, Y Combinator, the Linux Foundation and roughly a dozen other organizations signed a joint letter on July 24 urging Washington to avoid "premature restrictions" on open-weight AI models, calling instead for expanded public compute and shared training assets.
Notably absent from the signatory list: OpenAI and Anthropic, the two labs most associated with closed frontier models. Nvidia CEO Jensen Huang broke his long-running silence on X to publish the letter as his first-ever post there, arguing "the world needs both frontier closed models and frontier open models."
The letter lands as Washington weighs curbs on Chinese open-weight models following the Moonshot–Fable distillation dispute, and as a bipartisan "AI Kill Switch Act" moves through Congress the same week — evidence that policy, not model releases, may be the busier story this month.
DeepSeek's Legacy API Goes Dark as V4 Officially Takes Over
DeepSeek pushed its V4 model family out of preview and into general availability on July 20, and today, July 24 at 15:59 UTC, the legacy deepseek-chat and deepseek-reasoner endpoints that a huge share of existing API calls still point to stopped working entirely.
The retired names now map onto the non-thinking and thinking modes of deepseek-v4-flash. V4-Pro (1.6T total, 49B active parameters) and V4-Flash (284B total, 13B active) both ship MIT-licensed with a 1M-token context window, and V4-Pro reports 80.6% on SWE-bench Verified — within 0.2 points of Claude Opus 4.6.
DeepSeek has also introduced peak and off-peak pricing for the first time, roughly doubling token cost during Beijing business hours while keeping off-peak rates near preview levels — still, by a wide margin, the cheapest 1M-context frontier-class API on the market.