Anthropic AI Updates: September 29, 2026

1. Anthropic Released Claude Sonnet 5.5 at Sonnet 5 Prices

Anthropic. Claude Sonnet 5.5 (claude-sonnet-5-5) launched on September 28 at $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20 and cache writes at $2.50, the same list prices as Sonnet 5. Anthropic says it generates output more than 30% faster and costs up to 30% less per task because it needs fewer tokens and tool calls. Reported scores against Sonnet 5 include 70.6% vs. 10.3% on Terminal-Bench 4.0, 55.5% vs. 34.1% on CursorBench 4.0, and 80.1% vs. 57.0% on OSWorld 2.1. It is the first Sonnet model to ship with the cyber safeguards used for Opus 5.5, with high-risk cyber requests falling back to Sonnet 5, and Anthropic says Claude Haiku 5.5 will follow “in the coming weeks.” Source

2. Sonnet 5 Code Can Break on Sonnet 5.5 in Five Ways

Anthropic. The Claude Platform release notes list five breaking changes for Sonnet 5.5. To turn off up-front thinking, requests must send thinking: {"type": "between_tools"} instead of "disabled" (at high effort or below). Forced tool use (tool_choice types any and tool) returns a 400 error, and the older computer_20251124 computer use tool is rejected on the Claude API and Google Cloud. The advisor tool also rejects Opus 4.8, Opus 4.7, and Sonnet 5 as advisors. Thinking blocks produced by Sonnet 5.5 are now bound to the originating account (or a linked one), and the API silently drops them if another account sends them. Anthropic published a migration guide with before-and-after requests. Source

3. Claude Managed Agents Pairs With NVIDIA OpenShell for Layered Agent Control

Anthropic. Anthropic described a layered setup that runs Claude Managed Agents alongside NVIDIA OpenShell, part of NVIDIA’s Open Agent Safety Platform. Managed Agents keeps credentials in a vault the agent never reads directly, supports long-running sessions and multi-agent orchestration, and can run on customer-controlled sandbox infrastructure. OpenShell is Apache 2.0 licensed and enforces default-deny policies on every agent action, logs allowed and blocked decisions, and includes a policy prover that checks access boundaries. The idea is that model safeguards, credential isolation, and external runtime enforcement each fail independently. Notion, Rakuten, and Asana are cited as Managed Agents customers. Source

4. Claude Code 2.1.284 Adds Sonnet 5.5, One-Time Auto Mode Approvals, and Gateway Spend in Dollars

Claude Code 2.1.284 added Sonnet 5.5 as the default Sonnet model on the Anthropic API. Auto mode’s prompt before reading outside the working directories gained a “Yes, but ask again next time” option, and /mcp reconnect all retries every MCP server that failed to connect or needs authentication. For Claude apps gateway deployments, /usage and the status line now show spend limits in dollars (with used_usd, limit_usd, and period in rate_limits.spend_limit), telemetry can be forwarded directly to Google Cloud’s OTLP endpoint with auth: { google: {} }, and the gateway supports private_key_jwt certificate authentication with identity providers. Source

5. Trump Met Amodei and Framed the Talks Around Beating China

Anthropic. President Trump hosted CEO Dario Amodei for a late-night White House dinner on September 27, their first one-on-one meeting since the Pentagon labeled Anthropic a supply-chain risk. Asked about Amodei’s calls to slow AI development, Trump said, “I’m for, ‘Let’s go and let’s win,’” and put the US lead over China at about a year to 18 months. On September 28 Amodei declined to tell reporters at the Capitol what the two discussed or whether he would attend a White House meeting with tech leaders on Tuesday. Source