Cline v4.1.22 Update Breakdown: New Providers and Improved Error Recovery

AI coding assistant Cline has released version v4.1.22. As an open-source tool with nearly 70,000 stars on GitHub, every Cline release deserves attention from developers. This update focuses on expanding the provider ecosystem, refreshing the model catalog, and refining a series of error recovery workflows. Let's break down what actually changed.

New Providers: Expanding the Ecosystem

This update adds two new model providers: Bee, offered by HEOSSI, and Pareto Inference, which requires configuring a PARETO_API_KEY.

For a coding assistant like Cline — positioned as a universal gateway — expanding the provider roster means users have greater freedom in model selection and more leverage on pricing. Developers can flexibly switch between multiple upstream providers based on the cost-performance requirements of different tasks, without being locked into a single service. This open, multi-provider architecture is one of the core advantages open-source AI tools hold over closed products.

Reliability Improvements: Failover for Anthropic Requests

Reliability is the headline feature of this release. After the update, requests sent directly to Anthropic now use Anthropic's server-side rejection fallback mechanism. On OpenRouter and Cline's own channel, Anthropic models can failover to another upstream provider on failure instead of dropping the request entirely.

This change delivers a fairly direct improvement to the actual user experience — during peak hours or when an upstream node becomes unstable, tasks won't be interrupted by a single point of failure. For long-running coding tasks, this resilience design can significantly reduce the frustration of mid-task failures.

Additionally, when a response is blocked by a content filter, Cline will now explicitly inform the user that "content was blocked" and suggest rephrasing, rather than vaguely reporting "the model returned an empty response." Clearer error messages like this may seem like a minor detail, but they can dramatically reduce the time users spend troubleshooting.

---

What is Failover? Failover is a fault-tolerance mechanism in distributed systems: when a primary link or service node experiences a fault, timeout, or rate limit, the system automatically redirects the request to a backup node — transparently to the user. In AI inference scenarios, large model APIs are prone to 429 (rate limit) or 5xx server errors during peak hours. Without failover, clients can only wait and retry or receive an outright error. OpenRouter, as a multi-model aggregation gateway, maintains connections to multiple Anthropic upstream providers, giving it the ability to route failed requests to other available nodes. Cline's integration of this capability into its own request pipeline means that even if one Anthropic node experiences instability, the system automatically tries alternative paths — from the user's perspective, the task is slightly delayed rather than outright failed.

---

Model Catalog Refresh: Major Default Model Changes

This release includes a large-scale refresh of Cline's model catalog, with default models changing across multiple providers. GPT-6.1 Sol is now the new default model for a range of providers including OpenAI, OpenRouter, GitHub Copilot, Cortecs, Eden AI, Kilo Gateway, NanoGPT, OpenCode Zen, and Requesty.

Other changes include: Vercel AI Gateway switching to Ling 3.1 Flash, Tempr Gateway moving to MiMo V2.6 Flash, CrossModel and Ofox shifting to Claude Sonnet 5.5, Pioneer adopting GLiNER 2.5 Decide, and Scaleway using Qwen 3.8 27B. Cline's recommended model list has also been updated to include Claude Sonnet 5.5 and Claude Opus 5.5.

One important note: if you're using any of the providers above without having manually pinned a specific model, you will encounter a default model change after upgrading. Users with strong model preferences are advised to review and lock their configurations to avoid unexpected behavioral differences.

Multiple Fixes to Error Recovery Workflows

Despite the minor version bump, Cline has packed in a dense set of fixes around interaction details, centering on error recovery and task lifecycle management:

• Input restoration after API errors: After an API error, users can now type and send a new message directly instead of being forced to click Retry. Commands entered during recovery — such as /compact, /smol, or /newtask — are preserved in the input field and won't accidentally be sent to the model as plain text.
• Task resumption after login failure: If a new task is interrupted because Cline fails to log in, submitting a revised prompt after logging in will now launch the task with that prompt — previously, this situation would produce no response at all.
• Checkpoint-related fixes: Reset Code is now properly disabled when a message has no checkpoint, and no longer remains disabled for messages that do have a checkpoint on slow disks. Changes to Checkpoint settings now take immediate effect on the currently active task, and subsequent messages entered during the process are preserved in order.
• Cancellation now works: Canceling immediately after sending a message now actually stops the task. Previously, if the cancel action landed during the turn initialization phase, it would be ignored and the task would continue running.
• Reasoning token count fix: Reasoning tokens are no longer double-counted, though cost calculations themselves are unaffected.

These fixes collectively reflect the Cline team's attention to edge cases — most of them occur under non-ideal conditions like network anomalies, slow disks, and login interruptions: precisely the moments most likely to cause real frustration during actual use.

Targeted Optimizations for Amazon Bedrock

For Amazon Bedrock, commonly used by enterprise users, this release also includes several targeted fixes. GPT-6 and GPT-5.6 are now routed through inference profiles, and the India region (ap-south-1, ap-south-2) will correctly resolve profiles with the in. prefix.

OpenAI models behind inference profiles now receive supported reasoning effort settings. Nova 2 Lite's high-reasoning mode, application inference profile ARNs, and Nova Micro no longer send requests that Bedrock would reject. Additionally, a legacy AWS configuration saved without the "use profile" flag is now correctly identified as profile-based authentication instead of being silently discarded.

Furthermore, gateway models on hybrid endpoint providers now retain their respective API protocols rather than falling back to the provider-level default protocol. These changes are especially important for teams operating in multi-cloud or multi-endpoint deployments.

---

About Amazon Bedrock Inference Profiles: Amazon Bedrock's Inference Profile is an abstraction layer for cross-region model invocation provided by AWS. Users reference a model through a logical ARN (Amazon Resource Name), and AWS automatically selects the actual physical endpoint in the background based on regional load and availability. Profile ARNs typically carry a regional prefix (e.g., us., eu., in.) to identify the geographic partition. Previously, Cline had a bug in parsing the in. prefix for the India region (ap-south-1/ap-south-2), causing requests from those regions to fail to route correctly. With this fix, enterprise users deployed in the India region can now use profile-based routing normally — which is especially important for organizations with data localization compliance requirements that must keep inference traffic within a specific geographic region.

---

Summary: A Solid Maintenance Release

Overall, v4.1.22 is a classic, solid maintenance release: no groundbreaking new features, but meaningful refinements across provider ecosystem, reliability, error recovery, and enterprise deployment compatibility. For developers who rely on Cline daily for AI-assisted coding, these improvements deliver a smoother and more predictable experience. Upgrading promptly is recommended — and be sure to check for any configuration adjustments needed due to the default model changes.