Cline v4.1.23 fixes Claude 400 errors on custom Anthropic URLs, changes defaults for 19 providers

Cline v4.1.23 fixes a regression from v4.1.22 that caused Claude requests to fail with a 400 error when routed through a custom Anthropic base URL, such as Azure AI Foundry or a corporate gateway. According to the release notes, 4.1.22 began sending an option that only Anthropic's own API accepts.

The release was published by github-actions on 07 Oct (commit afaf4f2). Pricing and context-window figures are not part of this release, and the notes do not disclose them for any model named below.

Fixes

• Custom Anthropic endpoints: Claude no longer returns a 400 error through custom base URLs.
• Reasoning levels: Models that accept only certain reasoning levels, including Kimi K3, no longer reject requests. Cline now picks the closest level the model supports.
• Large MCP output: When an MCP tool returns more output than fits in context, Cline receives a preview plus a link it can page through with read_files. Previously, output past the cutoff was lost.
• Generate Commit Message: The feature now follows .clinerules, so conventions for language, format and ticket references apply to commit messages.
• Tool approvals: Approve and Reject work immediately, even while the panel is still updating. A late update failure no longer rejects a later, unrelated approval.
• Focus Chain: Reopened old tasks no longer show leftover Focus Chain checklist rows.

Behavior change

If a model's response ends without a recognized finish reason, Cline now asks it to continue once. Before, it treated the response as complete.

Model catalog refresh

• Cline recommended list: GPT-6.1 Sol added.
• Free list: Solar Mini 4 added. DeepSeek V4.1 Flash and space-bunny-alpha removed.
• Default models changed for 19 providers: 302.AI, AKI.IO, Blue Claw, CoralBricks, CrossModel, DevPass, DigitalOcean, GMI Cloud, LLM Gateway, Mistral, NanoGPT, Neon, Nvidia, Ofox, Requesty, Subconscious, The Grid AI, Vivgrid and Vultr.

The notes do not say which new default each provider received. Users on these providers who have not pinned a model should expect a different default. Vultr's model IDs were renamed upstream, so a pinned Vultr model may need to be re-selected.

What this means

The Claude fix matters most for enterprise users. Azure AI Foundry and internal gateways are common in organizations that route Anthropic models through controlled infrastructure. A regression that breaks them outright makes 4.1.23 a necessary upgrade for anyone who moved to 4.1.22.

The silent default changes are the larger operational risk. Across 19 providers, unpinned users may see different model behavior, latency and cost after updating, with no change to their own configuration. Teams that depend on reproducible agent behavior should pin models explicitly.

The reasoning-level and finish-reason changes point to the same problem: provider APIs differ in what they accept and how they signal completion. Mapping to the nearest supported reasoning level and retrying once on an unrecognized finish reason are practical workarounds. They also make Cline's requests less predictable than a strict pass-through would.

The MCP paging change addresses a real failure mode. Silent truncation of tool output can lead an agent to act on incomplete data, and a pageable preview lets it retrieve the remainder.

Source: Cline v4.1.23 release notes.