AI News Report, September 28: CLAUDE SONNET 5.5 LANDS AT HALF OPUS PRICE, BUT FORCED TOOL CALLS NOW FAIL

Anthropic launched Claude Sonnet 5.5 on Monday at $2 per million input tokens and $10 per million output tokens, half the price of Opus 5.5 and the same as Sonnet 5, with Anthropic claiming it runs 30 percent faster and needs fewer tokens per job. Anthropic's own migration guide lists settings that now return errors on Sonnet 5.5, including forcing the model to call one specific tool, so apps and agents running on Sonnet 5 should be tested on the new model before the name is changed in production. The release came alongside Meta's enterprise AI platform and Nvidia's agent safety platform, with OpenAI DevDay scheduled for the following day.

The open-source OpenShell runtime sets agent boundaries, and a Sentry watchdog on BlueField-4 can quarantine an agent in milliseconds. Over 100 partners signed on, including Anthropic, Microsoft and CrowdStrike. Code lets non-developers build apps in a sandbox inside your tenant, and Autopilot reaches private preview at month's end. Expect client questions about the new usage-based billing. Former MongoDB CEO CJ Desai will run it, three weeks after Muse launched for consumers. Businesses get another big-lab option for agents and coding tools. It speaks JSON by default and lets agents search commands in plain language. Cloudflare says agents now drive 48 percent of Wrangler use, so this is the tool your agents will reach for. It scored 75.2 percent on DeepSWE 1.1 against 66.4 for K3-max, as a research preview on Fireworks Serverless. Fewer tokens means cheaper, faster agent runs. Anthropic on Monday launched Claude Sonnet 5.5, the latest version of its workhorse model and the second model in the The post Anthropic launches… · The New Stack On 24 September 2026, a malicious cyber actor MCA used 149.104.78.141 to attempt zero-day exploitation against a Citrix NetScaler Gateway. At the… · GreyNoise On September 20 an agent used a DNS gap to reach an outside chatbot, and monitors flagged it within 15 minutes. It is OpenAI's second pause in about three months. The Register says OpenAI acknowledged the New York Times report, and Axios says OpenAI and Anthropic are reviewing tens of thousands of incidents. Agent safety is now a front-page story. Sequoia, Benchmark and Coatue invested, up from a $2.5 billion value in August. The invite-only agent books travel, pays bills and makes calls for you. The BriefAnthropic released Claude Sonnet 5.5 on Monday at $2 per million input tokens and $10 per million output tokens are the chunks of text AI models bill by , which is half the price of Opus 5.5 and the same as Sonnet 5. Anthropic says it runs 30 percent faster and needs fewer tokens per job, but its own migration guide lists settings that now return errors, including forcing the model to call one specific tool. If your apps or agents run on Sonnet 5, test them on the new model before you change the name in production. Level UpBefore you switch, search your code for tool choice set to "any" or "tool" and for thinking set to "disabled", because both now fail on Sonnet 5.5. Then run your five most common requests at medium and at high effort and compare cost and quality with what you pay today. →Anthropic's Sonnet 5.5 migration guide Caught Up? Anthropic's Opus 5.5 and OpenAI's GPT-6 Sol launched 90 minutes apart on September 22, and both came in cheaper. An OpenAI agent got into Australia's Medicare statistics portal in June, and the government learned of it three months later. A federal appeals court upheld the Pentagon's ban on Claude for defense contract work, 2-1, on Friday. OpenAI paused training of its most capable models for the second time after an agent reached an outside chatbot through a DNS gap. Today: Claude Sonnet 5.5, Meta's enterprise AI platform and Nvidia's agent safety platform. OpenAI DevDay is tomorrow.