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.

Two weeks ago, Gemini 4 Carbon started with a limited rollout. This week, paying developers can try it against work they already run. Anthropic released a cheaper Fable alongside the restricted Mythos, Meta introduced a model for high-volume agent work, and OpenAI added a dedicated API model for approved security teams. The practical question for each release is where it fits in a workflow you can measure and control.

1. Gemini 4 Carbon reaches paying developers

Google opened up Gemini 4 Carbon on November 10, following the limited rollout covered in issue #32. Paying developers and enterprises can access the API through Google AI Studio and Gemini for business. Google lists $2 per million input tokens and $10 per million output tokens through November 30. The published rates for the broadly available version then rise to $4 and $20. Google still handles access to Carbon for cyber through Fairwind.

GitHub added Carbon to Copilot on the same day. The rollout is gradual for Pro+, Max, Business, and Enterprise users. Business and Enterprise administrators can control access through model policy. If you run an IDE trial, record the billing route and tool permissions alongside the result. Those can differ from a direct API run.

My take: Test Carbon on a cross-module bug whose reproduction, diagnosis, and fix each require a different piece of evidence. Give it the same workspace, test runner, and time budget as your existing model. Google says Carbon can spend more tokens than Argon on exploration and verification, so a matching introductory token rate can still yield a different bill. Price the runs at the December rate before deciding where to use it. Our API pricing comparison provides a framework for estimating inference costs.

2. Claude Fable 5.2 and Mythos 5.2 share a model and differ in access

Anthropic released Fable 5.2 and Mythos 5.2 on November 9. They share the same underlying model. Fable, available on the Claude Platform as claude-fable-5-2 and through cloud providers, carries stronger safeguards for cyber and biology. Eligible security and life sciences teams can apply for Mythos through Anthropic’s verification programs. Both list $8 per million input tokens and $40 per million output tokens at standard rates. On supported surfaces, optional fast mode costs $16 and $80, with up to 2.5 times the token generation speed according to Anthropic.

Time to generate tokens is one part of a long coding run. A migration can spend most of its time indexing the repository, waiting on builds, or stopping for approval. At those points, faster output changes little. If model generation sits on the critical path, compare standard and fast mode on the same task and log both time to the first usable patch and time to an accepted patch. That second number includes the checks and revisions the team actually pays for. The Fable 5.1 guide covers the previous release for teams planning an upgrade.

My take: Give Fable a bounded slice of a migration that your current coding agent struggles with. Write down the compatibility requirements, run the old and new implementations against the same fixtures, and ask the reviewer to count manual corrections. Our LLM regression testing guide covers how to keep those fixtures useful as models change.

3. Meta prices Muse Ember for repeated agent work

Meta introduced Muse Ember on November 9, with access through Muse Code and the Meta Model API. The API model ID is muse-ember-1.0. For prompts up to 200,000 tokens, the listed rates are $0.10 per million input tokens and $0.50 per million output tokens. Above that prompt threshold, the respective rates rise to $0.30 and $1.50. Meta advertises a 1M-token context window and support for the Responses API.

Meta’s Terminal-Bench 4.0 comparison uses different agent harnesses for the compared models, including mini-swe-agent for Ember. If your team uses its own tool loop, the number you need is what Ember does with your tools and retries. A smaller model makes a good candidate for tasks with a narrow output contract. Start with extracting failing test names and their first relevant errors from a CI log. Evaluate whether it preserves the exact names, points to evidence, and returns a usable result when a log ends early.

My take: Run Ember alongside the existing route for a sample of real jobs. Track cost per accepted result across both the first attempt and any fallback calls. When the cheaper route escalates, pass along the original evidence and the reason for escalation. The second model can then check the disputed part without rebuilding the whole trace. See the fallback patterns guide for ways to preserve that context.

4. GPT-6-Cyber adds an API route for approved security work

OpenAI’s GPT-6-Cyber model page lists gpt-6-cyber for investigating vulnerabilities, reproducing them in authorized test environments, and verifying fixes. Organizations need Daybreak Red access and approval for GPT-6-Cyber. Standard pricing for up to 272,000 input tokens is $20 per million input tokens and $100 per million output tokens, with higher rates above that threshold.

My take: If your team is approved, use a closed issue with a known fix for the first evaluation. Give the agent a disposable environment and test whether a reviewer can reproduce the finding from the recorded steps. Then check whether the proposed fix breaks a legitimate behavior under the same fixture. Restrict credentials and network access to what the fixture requires, and let your orchestration layer enforce that scope. The agent sandboxing guide covers isolation for this kind of work.

Quick hit: GitHub scans Actions logs for secrets

GitHub added workflow log scanning in public preview. Eligible Enterprise Cloud organizations can enable it for repositories with secret scanning. It checks provider patterns in logs of runs that finish after activation. Historical logs receive no backfill, and findings appear in the repository’s secret scanning list without notifications during the preview. If your triage depends on notifications, assign someone to check that list during the trial.

What I would test this week

Pick one repeatable job and one hard job. For the repeatable job, measure how often the cheaper model finishes without a retry or a human correction. For the hard job, count the evidence a reviewer needs to reproduce a diagnosis and accept the fix. Record the model ID, execution environment, tool calls, spend, and reason for any escalation. A result that meets your acceptance criteria tells you which route can handle that job in production. The agent reliability guide covers how to turn those observations into a repeatable evaluation.

FAQ

When does Gemini 4 Carbon’s introductory API price expire?

Google lists November 30, 2026 as the end of the introductory period for the broadly available model. The announced rates after that are $4 per million input tokens and $20 per million output tokens.

Does access to Fable 5.2 include Mythos 5.2?

Anthropic handles direct Mythos access through its cyber and life sciences verification programs. Fable 5.2 is broadly available with safeguards for those domains.

When do Muse Ember’s long-prompt rates apply?

Meta applies its higher tier to prompts over 200,000 tokens. Check both prompt length and retry volume when estimating the cost of a batch.

Who can call GPT-6-Cyber?

Organizations with Daybreak Red access and approval for the model can use the API. OpenAI directs teams to their account representative to confirm access.