DEV Community

Cover image for AI Dev Weekly #35: GPT-6.2, Gemini 4 Neon and MAI Code 2
Joske Vermeulen
Joske Vermeulen

Posted on Originally published at aimadetools.com

AI Dev Weekly #35: GPT-6.2, Gemini 4 Neon and MAI Code 2

AI Dev Weekly is a Thursday series covering the AI developer news that changes what you can build, what it costs, and how you should operate it.

OpenAI, Google, and Microsoft all released coding-capable models this week. The choice looks like a row of names in an IDE, but the consequences show up in how an agent handles a correction, what it bills for a long run, and whether its patch survives review. GitHub's new Copilot Evals preview gives teams a way to test some of those changes against their own tools.

1. GPT-6.2 changes what survives between turns

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17. The API IDs are gpt-6.2-astra and gpt-6.2-sol. OpenAI lists standard rates per million input and output tokens of $8 and $40 for Astra and $2 and $10 for Sol, with cached input at $0.40 and $0.10 respectively. These are short-context rates. For Astra requests over 272,000 input tokens, the input and cache rates double and the output rate increases by half for the entire request. Release announcement · Astra model documentation

The Responses API changes matter if your application owns conversation history. GPT-6.2 can lose access to earlier reasoning items after you edit the conversation prefix, make certain model changes, or change the associated user or organization. Top-level instructions also come after input for GPT-6.2. OpenAI recommends putting reusable guidance in a developer message at the beginning of input and using instructions for guidance specific to the current request. GPT-6.2 migration guide

OpenAI also opened a beta for discounted background completion windows on selected models, including both GPT-6.2 releases. Eligible requests on Standard get 30% lower model token rates with a two-hour window and 40% lower rates with an eight-hour window. The discounted windows share capacity with Batch, and background mode temporarily stores response data. These are completion targets for individual responses. A workflow with a tool call and a follow-up still needs its own end-to-end deadline. Background mode documentation

My take. Replay a real multi-turn task where someone corrects the scope after the agent has gathered evidence. Check whether the agent carries forward the valid evidence, repeats work, or follows the old constraint. Compare an appended clarification with the application's usual history edit. Log cached input, repeated tool calls, and review time for both. This test can expose a regression that a single-turn model comparison would miss. Prompt caching guide · Previous issue

2. Gemini 4 Neon adds another option for sustained coding work

Google began rolling out Gemini 4 Neon on November 18 through the paid Gemini API, Gemini for business, and eligible Antigravity and Gemini app subscriptions. It lists a 1 million-token context window, with rates of $3 per million input tokens, $0.15 for cached input, and $15 for output. Google positions Neon for complex work that repeats often, Hydrogen for fast everyday tasks, and Carbon for the hardest long-running jobs. Google's Neon announcement

Carbon's temporary $2 input and $10 output rates make Neon look more expensive in a November price comparison. Once Carbon's promotion ends after November 30, the announced public rates are $4 input and $20 output. Even then, a model that reads more files or makes more repair attempts can cost more per completed task. Quote both prices at the date your team intends to deploy. Google's Carbon rollout and pricing · AI API pricing guide

My take. Try Neon on a bug that requires reading a screenshot, locating the component that renders it, and writing a regression test. Run the same repro against your current model with identical fixtures and tool permissions. Score whether the patch fixes the underlying behavior, then count the attempts and wall-clock time. If the cheaper run still leaves you fixing keyboard navigation by hand, include that time in the comparison.

3. MAI Code 2 and Flash reach GitHub Copilot

Microsoft introduced MAI Code 2 and MAI Code 2 Flash at Ignite on November 17. The larger model targets multi-file implementation, refactoring, and debugging. Microsoft positions Flash for routine edits and faster iterations and also lists it in a limited-access Microsoft Foundry preview. GitHub lists Copilot Pro, Pro+, Max, Business, and Enterprise for MAI Code 2, with Free and Student also listed for Flash. GitHub specifies annual-plan eligibility for subscribers and workspace model policies for Business and Enterprise. Microsoft's Ignite announcement · GitHub's Copilot release note

GitHub scheduled MAI-Code-1.1-Flash for retirement on December 18 across Copilot. If you have an agent profile or workflow that pins the old model, check what it selects after the change and whether the replacement is enabled for the people who use it. GitHub's retirement notice · Copilot usage-based billing guide

My take. Compare the two models inside the same Copilot client. Give Flash a dependency bump with a known failing test. Give both models a change that crosses a package boundary, keeping the tests and reviewer fixed. The second task can tell you when the smaller model needs escalation. Document that trigger with the model pin so a future default change has a test to run.

4. Copilot Evals puts regression tests next to agent configuration

GitHub launched Copilot Evals in public preview for VS Code, the Copilot CLI, and the Copilot app. It uses the local Vally evaluation engine. Developers can create tests for skills, custom agents, and MCP tools, approve individual runs, inspect failed checks, and compare results before and after a configuration change. GitHub's Copilot Evals release note

My take. Give a release-notes agent one fixture with a complete changelog and one with the changelog missing. Check that it links the source in the first case and reports the missing input in the second. An agent can produce polished copy from either fixture. The second result tells you whether it will disclose the obstacle before someone publishes the draft. Repeat the runs after changing the model or its instructions. LLM regression testing guide · Evaluation tools for multi-step tool use

Quick hit: Mistral ships Vibe Desktop

Vibe Desktop is generally available for macOS and Windows. It brings Work, local coding sessions, and coding in managed cloud sandboxes into one app. Local sessions share the CLI's core agent engine and session logs, with room for differences in prompts and tools. If you move a project from the CLI, check the files, commands, and credentials the desktop session can access. Mistral release notes · Vibe remote agents guide

FAQ

Do I need to change stored conversations for GPT-6.2?

Review any code that rewrites earlier messages or tool calls. Keeping the existing prefix stable and appending new guidance preserves the reasoning and caching behavior described in OpenAI's migration guide.

Is Gemini 4 Neon cheaper than Carbon?

Neon has lower listed token rates than Carbon's announced post-promotion rates. Carbon's introductory rates run through November 30. Actual task costs also depend on token use and retries.

Can I call MAI Code 2 Flash outside GitHub Copilot?

Microsoft lists a limited-access preview in Foundry. Check eligibility for that route separately from your Copilot plan.

What can I test in Copilot Evals?

The preview covers custom agents, skills, and MCP tools. A test can check whether the result meets your requirements and how the agent used its tools along the way.

Originally published at https://www.aimadetools.com

Top comments (0)