Anthropic AI Updates: September 25, 2026

1. Project Swap Found That Agents Negotiate Fine and Understand You Badly

Anthropic ran a book-trading market with 201 employees across six offices, where each person had a short conversation with Claude about their reading taste before a Claude agent traded on their behalf. Claude’s inferred rankings matched participants’ real rankings on 61 percent of book pairs, against 50 percent for random and 53 percent for popularity-based ordering, and doubling the intake conversation length bought roughly 4 percentage points. Participants ended up with books at 0.55 on their own preference scale against an optimal 0.89, and the research attributes 85 percent of that shortfall to imperfect preference representation and only 15 percent to bargaining inefficiency. Source

2. Model Choice Mattered More Than Telling the Agent to Be Ruthless

In the same 200-plus reruns of the market, Anthropic varied both the model and the agent instructions. Haiku reached 0.75 market efficiency against Opus at 0.88, while ruthless agents beat prosocial ones by only 0.02. Between 78 and 96 percent of agents disclosed their top preference outright and fewer than 1 percent lied about it, and participants said they would hand an agent about 30 percent of a yearly book budget, falling to 23 percent when they felt their preferences had been captured poorly. Source

3. Anthropic Published the Usage Shift That Explains the Opus 5.5 Pricing

Anthropic put numbers behind why it rebuilt Opus around long sessions: between March and September 2026, context per request rose 2.6x, the input-to-output token ratio moved from 189:1 to 324:1, the model now runs 3.3x longer per prompt with 40 percent more function calls, and interruptions fell 68 percent. Opus 5.5 is billed as roughly 40 percent cheaper to run than Opus 5 on typical token-billed workloads, with cached tokens down 60 percent and input and output down 20 percent, and generates output over 30 percent faster. Claude Code changes cut cache misses by more than half, and forked subagents now inherit the parent cache instead of reprocessing context. Source

4. Claude Tag Can Now Use Your Own Connectors Inside a Shared Channel

Anthropic extended Claude Tag so a user’s personal connectors work in Slack channels, not just the admin-attached channel connectors. Output visibility is controlled per invocation through review mode, where the user approves each response before it posts, or auto mode, where Claude screens for sensitive content first. Connector activity lands in that tool’s own log under the user’s account, channel connectors stay the route for unattended scheduled routines and org-wide actions, and enterprise admins can force review for everyone. It is rolling out on Team plans first. Source

5. Claude Code 2.1.282 Adds a Readiness Grace Period for Postgres Failovers

Claude Code 2.1.282 added store.readiness_grace_seconds to the apps gateway so /readyz stays ready through a short Postgres outage such as a database failover. The release also added a maxProseWidth setting that caps prose width in wide terminals while tables and code blocks keep the full width, a startup notice plus /status and claude doctor entries naming telemetry variables that were ignored or that disabled telemetry, and an allowClaudeInChromeWithManagedMcp managed setting that lets claude --chrome run alongside an exclusive managed-mcp.json. Source