VS Code 1.143 Adds Side-by-Side AI Model 'Bake-Off' and Local Coding Without Using Credits

Microsoft has shipped Visual Studio Code 1.143, introducing two features that directly solve the biggest headaches facing everyday software developers using AI assistants: burning through monthly GitHub AI credits on simple edits, and guessing which AI model will write the cleanest code for a tricky feature. With the new release, users can pit multiple AI models against each other on the exact same prompt in a side-by-side “bake-off” and run Microsoft’s MAI Code 1.1 Flash model locally on their own laptop or desktop at zero credit cost.

For anyone paying for a monthly GitHub Copilot subscription or managing a team’s AI seat budget, watching cloud credits evaporate on routine boilerplate, comment cleanup, or quick syntax fixes is a constant frustration. VS Code 1.143 adds an “On This Device” tab directly inside the model picker, allowing developers to download MAI Code 1.1 Flash to their machine. Any response generated locally is badged “On-device” and consumes zero GitHub AI credits, and users can turn on “Auto” model selection so VS Code handles simple prompts locally while escalating harder problems to the cloud.

When a programming task is complex enough to justify cloud credits—such as refactoring a multi-file payment checkout flow or debugging an asynchronous state bug—developers no longer have to run one model, revert the mess when it fails, and re-prompt a second model from scratch. By enabling the new comparison mode in the Agents window, a developer can type one prompt and watch Copilot, Claude, and Codex harnesses build competing solutions simultaneously in isolated sessions.

Once the competing models finish their drafts, VS Code 1.143 can invoke an optional Judge model to grade each attempt on execution time, token consumption, and code validity. Even better, a built-in “Synthesize” button lets the developer merge the best architectural ideas from the runner-up models into the winning code diff with a single click.

Fast Facts: What’s New in Visual Studio Code 1.143 for Practical AI Coding

• Release & Coverage: Shipped in Microsoft’s Visual Studio Code 1.143 update and detailed in Visual Studio Magazine’s release breakdown.
• Default Copilot Agent Harness: All new agent sessions now run on the Copilot SDK harness by default through the background Agent Host, accessible from both the main editor and the dedicated Agents window.
• Side-by-Side Model “Bake-Off”: Turning on sessions.chat.compareAgents.enabled lets developers select 2 or more models (across Copilot, Claude, and Codex harnesses at custom reasoning levels) to tackle the same prompt in parallel.
• Judge Scorecard & Synthesize Action: An optional Judge model scores competing attempts on token usage, elapsed time, and test evidence, while an on-demand Synthesize button blends the strongest parts of multiple runs.
• Zero-Credit Local AI Inference: Developers can download Microsoft’s MAI Code 1.1 Flash from the “On This Device” model picker tab; local runs are labeled On-device and use 0 GitHub AI credits.
• Granular Undo & “Keep” Controls: A new file-level working set lets users mark reviewed files with Keep (returning them to the review queue only if an agent edits them again) and undo changes from just the latest turn, the current session, or the whole branch without wiping out human-written code.

Feature Comparison: How VS Code 1.143 Changes Daily Coding Workflows

The table below shows how the new tools in VS Code 1.143 solve common day-to-day problems compared to earlier editor versions.

How to Use the AI Model Bake-Off and Free Local Inference in Practice

Two practical workflows make VS Code 1.143 an immediate upgrade for both solo builders and enterprise teams:

1. Running a Multi-Model “Bake-Off” on High-Stakes Code

Suppose you need to convert a legacy REST API endpoint into a paginated GraphQL resolver with unit tests. Different frontier models have very different habits—one might write cleaner TypeScript types while another writes more thorough edge-case tests.

In VS Code 1.143, you open Settings, enable sessions.chat.compareAgents.enabled, and open the Agents window. Before hitting Enter on your prompt, you pick two or three combinations—for example, a Claude harness at medium reasoning and a Codex harness at high reasoning. VS Code spins up separate background sessions through the Agent Host so each model works in its own sandboxed branch of the working set.

When both finish, the optional Judge model displays a side-by-side scorecard showing:

• How many tokens and seconds each model spent;
• Why one implementation is cleaner or safer;
• Validation evidence showing which attempt passed local build checks.

If Claude wrote the cleanest resolver but Codex wrote a better test suite, clicking Synthesize instructs the agent to fold Codex’s test cases into Claude’s implementation.

2. Stopping Credit Drain With Local MAI Code 1.1 Flash and “Keep” Locks

To stop wasting paid cloud credits on small tasks, open the VS Code model selector dropdown and click the On This Device tab. Downloading MAI Code 1.1 Flash installs a compact, Microsoft-tuned coding model directly onto your workstation. Once installed, you can either pick it manually whenever you are offline or switch the selector to Auto, which routes straightforward completions to your local hardware (badged On-device) and saves your GitHub AI credits for heavy multi-file tasks.

Meanwhile, the new file-level working set fixes the most annoying part of long agent sessions: losing track of which files you have already checked. As you review an agent’s multi-file pull request, clicking Keep on a file marks it as approved. It stays out of your review list unless the agent touches that file again in a later turn. If the agent makes a mistake on turn 4, the updated Changes view lets you discard only turn 4’s edits without touching turns 1–3 or the lines you typed yourself.

Practical Rollout Guide for Engineering Teams

To get the most out of VS Code 1.143 without running into local disk bloat or runaway token usage during parallel bake-offs, teams should configure four straightforward defaults.

The 4-Step Enterprise Implementation Playbook

1. Enable Auto Local Routing With MAI Code 1.1 Flash: Have developers open the VS Code 1.143 model picker, download MAI Code 1.1 Flash under the On This Device tab, and set model selection to Auto. This immediately offloads routine syntax fixes and single-function completions from your team’s monthly GitHub AI credit pool.
2. Reserve Multi-Model Bake-Offs for Architectural Tickets: Because running 3 models in parallel plus a Judge model uses roughly 3x to 4x the cloud tokens of a single prompt, enable sessions.chat.compareAgents.enabled specifically when starting complex refactors, framework migrations, or tricky bug investigations where a bad first attempt would waste an hour of human debugging.
3. Adopt the “Keep-Per-Turn” Review Habit: Train developers never to let an agent run 10 turns unchecked. After each turn in the Agents window, inspect the modified files, click Keep on validated files, and use the turn-level Discard button in the Changes view immediately if an agent drifts off course.
4. Standardize on the Default Copilot SDK Harness and CLI: Use the streamlined Copilot CLI installer included in 1.143 so terminal sessions and editor sessions share the same background Agent Host state, allowing developers to start a long task in the terminal and review the resulting file diffs visually in VS Code.