Claude, the essentials — edition of July 29, 2026
Claude Opus 5 Arrives as Anthropic Urges a Slower AI Race
Anthropic launched Claude Opus 5 the same day its chief executive joined rivals in asking Washington to temper the pace of model releases, while fresh security research, a patched flaw and a new defensive product underscored what is at stake when AI moves fast.
- Anthropic officially launched Claude Opus 5, its new flagship model, which scored 24% on SlopCodeBench, a benchmark built to expose coding models that look competent but fail under scrutiny.
- Anthropic's CEO joined counterparts from OpenAI, Meta and Google in asking U.S. authorities to slow the release cadence of AI models, even as other LLM providers publicly pushed back on his more cautious stance toward open-weight models.
- Research conducted with Claude uncovered significant cryptographic weaknesses in internet security mechanisms, while Anthropic separately patched a virtual-machine escape flaw in Claude Cowork on Mac.
- A new defensive cybersecurity model, tested by SonicWall through Anthropic's Glasswing program, was pitched as safer and cheaper to run than prior offerings.
- Microsoft is closing a difficult fiscal year amid stock pressure attributed in part to intensifying competition from Anthropic.
Opus 5 lands, with mixed signals on real-world coding
Anthropic officially rolled out Claude Opus 5 as the successor to its flagship line. Its 24% score on SlopCodeBench — a benchmark specifically designed to catch models that produce code which looks plausible but fails under scrutiny — is a useful data point rather than a headline number: it says as much about how far coding benchmarks still have to go as it does about the model itself.
A parallel item from the developer community offers a grounded counterpoint: an independent developer used Claude to diagnose and fix memory issues in production during a sudden spike to 4,300 simultaneous users. Read together, the two items suggest a model whose everyday utility in applied, high-pressure engineering work is not fully captured by adversarial benchmarks alone.
Sources: Anthropic dévoile Claude Opus 5 · Claude Opus 5 obtient 24 % sur le benchmark SlopCodeBench · Used Claude to handle a 4,000+ user spike
A rare industry consensus on pace — and a public split on openness
In an unusual show of alignment, Anthropic's chief executive joined leaders at OpenAI, Meta and Google in asking U.S. regulators to slow the rate at which new AI models are released. The joint request signals that companies competing hard against each other now see uncontrolled release velocity as a shared risk rather than purely a competitive lever.
That alignment does not extend to open-weight models. Anthropic's CEO restated that the company does not want open-weight AI banned, but voiced clear reservations about it — a position other major LLM providers are now publicly contesting. The contrast between a joint slowdown request and a solo stance on openness suggests Anthropic is trying to shape the safety debate on its own terms even while seeking common ground with rivals on pacing.
Sources: Les dirigeants d'Anthropic, OpenAI, Meta et Google demandent un ralentissement de l'IA · Le patron d'Anthropic précise sa position sur les modèles ouverts face aux critiques · D'autres fournisseurs de LLM défendent l'IA open weight face à Anthropic
Security cuts both ways: new findings, a patched flaw, a defensive product
Anthropic disclosed research, conducted with Claude, that uncovered significant cryptographic weaknesses in mechanisms underpinning internet security — a reminder that AI-assisted research is increasingly capable of surfacing infrastructure-level vulnerabilities. On the same day, Anthropic confirmed it had fixed a virtual-machine escape vulnerability in Claude Cowork on Mac, an issue that would have let the tool break out of its intended sandbox.
Anthropic also introduced a new model built specifically for defensive cybersecurity work, described as both safer and less costly to run than previous options; SonicWall is now testing it — Claude Mythos 5 — through the company's Glasswing program. Taken together, these three items describe a company simultaneously discovering new vulnerabilities, closing its own, and productizing defense — a posture that only makes sense if security research is treated as core infrastructure rather than a side project.
Sources: Claude découvre des failles cryptographiques critiques dans des systèmes de sécurité · Anthropic corrige une faille d'évasion de machine virtuelle dans Claude Cowork sur Mac · SonicWall rejoint le projet Glasswing d'Anthropic pour tester Claude Mythos 5 · Anthropic présente une IA de cybersécurité plus sûre et moins coûteuse
Competitive pressure shows up in the numbers
Microsoft is wrapping up a difficult fiscal year marked by a stock decline that observers link in part to intensifying competition from Anthropic — a sign that Anthropic's product momentum is now visible in the financial results of the largest incumbents, not just in usage charts or benchmark tables.
On a smaller but practical note, Anthropic shipped version 0.120.2 of its Python SDK, fixing a compatibility issue with MCP SDK v2 — routine maintenance that keeps the developer tooling underpinning Opus 5 and other releases usable in production.
Sources: Microsoft termine un exercice fiscal difficile, sous la pression concurrentielle d'Anthropic · Nouvelle version du SDK Python d'Anthropic (v0.120.2)
This edition is an original synthesis written by Claude from aggregated news (press, Hacker News, Reddit, GitHub), under the editorial supervision of Héra SASU. Every fact links to its source. See the live feed →
Claude News is published by Héra SASU. Independent media, not affiliated with Anthropic.