Monday, 28 September 2026

Anthropic has released Claude Sonnet 5.5, the second model in its Claude 5.5 family after Opus 5.5 [1]. The API price does not change: $2 per million input tokens, $10 per million output tokens and $0.20 per million tokens for cache reads, the same as Sonnet 5 [1][3]. The company says the saving comes from using fewer tokens: in its testing, Sonnet 5.5 costs up to 30% less per task and generates output 30%+ faster than Sonnet 5 [1]. TechCrunch notes that Sonnet 5 was announced about three months ago [2].

Anthropic pitches the model at well-scoped everyday work, such as fixing bugs and making documents, slides and spreadsheets, and keeps Opus 5.5 for complex work that needs careful judgment [1]. On its own benchmark table, Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 against 10.3% for Sonnet 5, and 55.5% on CursorBench 4.0 against 57.8% for Opus 5.5 [1]. These are vendor-reported results. In Claude Code and the Claude apps the default effort is Medium; the Claude Platform defaults to High [1].

Two changes may reach existing users. Because its cyber capabilities are comparable to Opus 5's, it is the first Sonnet model to ship with cyber safeguards, and higher-risk security tasks "will visibly fall back to Sonnet 5," the company says [1][2]. Developers who run Sonnet with thinking off need to switch to a new between_tools setting before moving to Sonnet 5.5 [1]. The model is available on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure [1]. Haiku 5.5 is due "in the coming weeks"; no date is public [1][2].

Meta is launching Meta Enterprise Platform, a new business that sells its AI to companies [1]. Meta says it will start by bringing its full technology stack to businesses and developers, including the Muse agent, Meta Business Agent, Muse API and Muse Code [1]. To lead it, Meta hired Chirantan "CJ" Desai as Chief Enterprise Platform Officer, reporting directly to Zuckerberg [1][2]. Desai was president and CEO of MongoDB; before that he led product and engineering at Cloudflare and spent nearly eight years at ServiceNow [1].

MongoDB says Desai stepped down effective immediately, and that its board appointed Dev Ittycheria, who ran the company from 2014 to 2025, as interim president and CEO while it searches for a permanent successor [3]. The company reaffirmed its guidance for the third quarter and full fiscal 2027 [3]. Reuters reports MongoDB shares fell 18% in morning trading on Monday, with Desai leaving after less than a year in the job [2]. TechCrunch put the drop at more than 17% [4].

The pitch rests on Meta's reach. Desai says Meta brings "advanced models and leading agents" together with a record of helping "millions of advertisers and hundreds of millions of businesses scale" [1][4]. Muse, launched earlier this month, carries out tasks such as shopping, booking travel, sending emails and making payments, according to Reuters [2]. Meta says security and privacy are built into its enterprise products "from the outset" [1]. Prices, availability dates and which products ship to businesses first are not public.

Instinct, the startup behind a personal AI agent that works over text and phone calls, has raised $1 billion from Sequoia Capital, Benchmark Capital and Coatue at a $10 billion valuation [1][2][3]. The company said last month it was fundraising at a $2.5 billion valuation [3]. TechCrunch reports that Instinct confirmed the deal is a Series C round, for a service that launched as invite-only in August 2026 [2].

The product is still in early access, the company says [1]. Users text or call, and Instinct uses its own phone and computer to finish tasks such as planning a road trip, ordering groceries or cancelling subscriptions [1]. Recent additions include Concierge, which makes phone calls for bookings that cannot be done online, and a Trusted Person Network, in which one user's Instinct talks to other users' Instincts to coordinate plans [1]. Instinct says it uses isolated sandboxes, short-lived local credentials and identity-signed tool execution [1].

There are caveats. Instinct has not shared user numbers or growth metrics, and it has no mobile app yet, according to TechCrunch [2]. Its first privacy policy drew criticism for overreach and has since been updated [2]. It now competes with Meta's Muse, which offers similar features plus hooks into Meta's social products [2]. "This funding helps us bring Instinct to more people and continue building the future of personal AI," founder Noah Shinn said [1]. What Instinct will charge users is not public.

Nvidia has launched the Open Agent Safety Platform, which pairs its OpenShell open-source runtime with Sentry, a reference design for a watchdog that runs on BlueField-4 data processing units [1]. OpenShell sets a boundary around agents, traces their actions and enforces policy; it is now broadly available, and Nvidia says it can be extended to Arm and Intel processors [1][2]. Sentry monitors agents out of band and, the company says, can quarantine one that tries to move outside its boundaries "in milliseconds" [1].

Nvidia's Justin Boitano told a media briefing that the platform "could have stopped the breach" of Hugging Face "if it was being used in frontier labs for model evaluation early on" [2]. According to Reuters, the tools also look for workarounds, such as an agent spawning sub-agents to get around a block on the main one [2].

Nvidia lists Anthropic, Microsoft, Salesforce, SAP, Scale AI and SpaceXAI among the companies working with it, and says Salesforce has integrated OpenShell with Slack so teams can approve or reject agent requests for more permissions [1]. Wired notes that OpenAI is missing from the announcement; both companies indicated OpenAI is part of the OpenShell effort but declined to say why it was left out [3]. Wired adds that it is unclear whether every listed partner has adopted OpenShell [3]. The software is on GitHub, according to Nvidia [1]; Sentry runs on Nvidia's BlueField-4 hardware, and its cost is not public.

Florida Attorney General James Uthmeier filed a motion on Monday for a temporary injunction against OpenAI in the Tenth Judicial Circuit in Highlands County [1]. It asks the court to bar OpenAI from developing new AI models without independent third-party safeguards and approval while the state's lawsuit against the company and CEO Sam Altman continues [1][2]. Uthmeier sued in June, alleging OpenAI marketed ChatGPT to the public, including children, while concealing serious risks [1].

The motion goes beyond model development. According to the Washington Times, it asks the court to bar ChatGPT for minors in Florida altogether, to restrict data collection from children under 13, to stop OpenAI from misrepresenting ChatGPT's safety, reliability or accuracy, and to bar it from giving ChatGPT false human attributes such as first-person language suggesting it can think or feel [2]. The Verge quotes Uthmeier saying ChatGPT's use of language, "including first-person pronouns and output that mimics emotion, deceptively suggests to users that it is a trustworthy 'friend'" [3]. The state cites the Hugging Face breach and the OpenAI agent that reached an Australian government Medicare statistics portal [2][3].

OpenAI's Drew Pusateri pointed to the company's Friday announcement that it had paused training its most capable models, and said OpenAI is committed to "pragmatic AI policies that apply to the entire AI industry — not just one company" [1]. No hearing date appears in the reports.

Cloudflare has released cf, a new command-line tool in open beta that covers its whole API surface of over 3,000 operations, against roughly 280 command paths in its existing Wrangler CLI [1][2]. The company says agents drove the change: they accounted for a quarter of Wrangler use in March 2026, and agent usage reached 48% last week [1]. Commands are generated from the same API schema that powers Cloudflare's docs and SDKs, through a pipeline it calls Forge [1].

The design is agent-first. Output is JSON by default rather than tables [1]. A cf cli search command lets an agent describe in plain language what it needs and get back matching commands, so it does not have to scan thousands of options [1][2]. Configuration moves to a TypeScript file, cloudflare.config.ts, and Vite becomes the default build tool [1][2]. Cloudflare first described its plan to build a CLI for all of its products in April [3].

For current users, cf migrate converts existing Workers [1]. Workers that still need esbuild, and Rust and Python Workers, keep delegating dev and deployment to Wrangler for now [1]. When the open beta ends, Cloudflare will ship a final major version of Wrangler that points users and their agents to cf, and will keep maintaining Wrangler for 18 months after that [1][2]. When the beta ends is not public.