Claude Code 2.1.268 targets gateway costs and closes agent security gaps

[{"h":"CONFIRMED: Claude Code 2.1.268 targets gateway costs","p":"Claude Code 2.1.268, published on September 10, does not ship a new model. Its important change is operational: Anthropic now synchronizes prices configured on a managed gateway with the customer’s spend meter and fixes a regression that made every request fail on some Anthropic compatible endpoints. For companies placing agents behind their own or a third party gateway, that means fewer billing surprises and less risk that an update breaks the entire workflow."},{"h":"Clearer cost, not a lower price","p":"The pricing feature works when an administrator sets pricing: in gateway.yaml. Signed in clients receive the same rates through managed settings, so /cost and telemetry track the spend meter. This is not an automatic price cut. It is an attempt to align the local panel with the rate actually applied by the gateway, a basic requirement for teams comparing cost per task, cost center, or customer."},{"h":"The fix that can unblock an entire workflow","p":"The most urgent fix affects third party endpoints configured through ANTHROPIC_BASE_URL. Since version 2.1.265, a regular expression in the Artifact tool schema was rejected by some compatible servers, causing HTTP 400 on every turn. Version 2.1.268 removes that break point. The lesson for a small or midsize company is simple: API compatibility does not end when the first request succeeds. Tools, schemas, and less common paths must be part of an upgrade test."},{"h":"Gateways become more explicit about exposure","p":"The package also hardens the security boundary around gateways. It warns at startup when access_control.allow_cidrs is empty and shows a one time warning when the first request arrives from a public address. The new gatewayInternalNetworks setting can allow login only from the organization’s own public IPv4 block. That is useful protection, but it does not replace a firewall, strong authentication, or an exposure review. The warning exists precisely because a permissive configuration remains possible."},{"h":"Sessions, plugins, and agents become less fragile","p":"Long running sessions get less visible improvements as well. The self hosted runner can now remove per session state when a run ends, claude auth status --json reports the configuration directory, and plugin commands gain JSON output with row level error details. The release also fixes a busy loop that kept idle sessions using high CPU, WebFetch hanging on servers that never finish a response, and respawned agents inheriting tools or prompts from files outside a trusted folder."},{"h":"Editorial reading","p":"My verdict: update development environments and move to production after a short validation, especially when gateways, MCP, plugins, or remote sessions are involved. For MaxIA and other small companies, the highest value is clearer cost tracking, the removed HTTP 400 failure, and more legible network policies. The marketing is in the new controls. The real value is making an agent less mysterious once it leaves a laptop and enters a shared operation."},{"h":"Sources","p":"Official GitHub release: https://github.com/anthropics/claude-code/releases/tag/v2.1.268 | Structured release record: https://api.github.com/repos/anthropics/claude-code/releases/tags/v2.1.268 | Independent distribution catalog, npm: https://registry.npmjs.org/@anthropic-ai%2fclaude-code | Changelog documentation: https://code.claude.com/docs/en/changelog"}]