AI Dev Weekly #32: GPT-6.1 Terra, Gemini Carbon, Kimi K3.1 and Copilot Fusion

Week of Oct 23-29: GPT-6.1 Terra, restricted Gemini Carbon access, Kimi K3.1 Pro, Copilot Fusion and GitHub's agentic workflow launches.

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.

Three model announcements and a busy GitHub Universe make this a substantial release week. OpenAI added GPT-6.1 Terra to the middle of its lineup, Moonshot rolled out Kimi K3.1 Pro, and Google announced Gemini 4 Carbon. GitHub moved model orchestration into general availability and connected more of the work around the coding agent.

Availability is the first filter. Terra is available through the API. Kimi's hosted products are live, but its new weights are promised for next week. Carbon is beginning with selected security partners. Copilot Fusion is generally available, while Projects and intelligent local routing have different preview labels. A launch headline alone does not tell you which one you can deploy today.

This edition covers developments through October 29, including updates that arrived after last Thursday's issue. It is based on official documentation and announcements, not hands-on benchmarks of every new release.

1. GPT-6.1 Terra targets visual work at $1 input and $5 output

OpenAI released GPT-6.1 Terra on October 26, according to its API changelog. The model ID is gpt-6.1-terra; its positioning emphasizes visual reasoning, computer use, and writing rather than maximum reasoning capability.

The model reference lists a 1,050,000-token context window and 128,000 maximum output tokens, with text and image input and text output. Tool calling requires the Responses API. Reasoning ranges from none to max, with medium the default.

Standard rates per million tokens are:

For prompts above 272K input tokens, the full request uses doubled input and cache rates and 1.5 times the output rate. The context limit and the standard-price threshold are different numbers.

My take: Terra is worth evaluating for screen-heavy workflows, document interpretation, and bounded writing tasks that do not justify your most expensive model. I would test exact values, coordinates, and tool arguments, not just whether the description of a screenshot sounds convincing.

Keep the rest of the workflow fixed while comparing it with your current model. Record accepted results, elapsed time, retries, and review effort. A cheaper call can still make a more expensive task if it needs repeated correction. The AI model comparison helps structure the shortlist, and our agent reliability evaluation guide covers the acceptance criteria.

2. Gemini 4 Carbon is a frontier announcement, not a public API rollout

Google announced Gemini 4 Carbon on October 28. Access begins with a subset of cyber defenders and expands through the Fairwind Program. Google says broader developer, business, and consumer availability will follow; it does not give a public-access date in the announcement.

The emphasis is sustained engineering and cybersecurity work. Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench, with those external cyber evaluations run without guardrails. Those are vendor-published benchmark conditions, not a guarantee about the behavior of a future generally available endpoint.

This follows the broader Argon rollout covered in edition #30. Do not transfer Argon's availability to Carbon just because they share the Gemini 4 name.

My take: The interesting shift is models staying useful across a long engineering task, not simply producing a stronger first answer. Longer runs create more opportunities to act on a mistaken assumption, consume budget, or encounter untrusted instructions. Evaluation needs to cover those trajectories, including whether the model stops when its authority runs out.

Most readers cannot test Carbon directly yet. That makes this a planning signal rather than an immediate migration recommendation. Improve the evaluation set and containment you already have: representative repositories, controlled targets, explicit permitted actions, and external logs. Our agent sandboxing guide and agent audit logging guide cover controls that remain useful whichever frontier model arrives next.

3. Kimi K3.1 Pro ships in hosted products, with weights due November 4

Moonshot rolled out Kimi K3.1 Pro and Pro Highspeed on October 23. Its technical launch post describes a post-trained K3 successor with a million-token context window. Hosted access is available through Kimi, Kimi Work, Kimi Code, and the API. The full weights are promised by November 4, so this is not yet the downloadable release.

Moonshot reports 65.7% on Terminal-Bench 4.0 and says Highspeed delivers two to three times the speed. Both are vendor claims worth testing under your own harness and workload, not a cross-provider buying decision by themselves.

The Kimi Code release notes identify integration details that are easier to miss:

• Thinking must stay enabled. Supported levels are low, high, and max, with high the default; disabling thinking returns an error.
• Start a new session instead of swapping the model into an old K3-series conversation.
• Membership controls access: the full 1M context and Highspeed require higher plans than basic Pro-model access.

My take: Test the hosted model now if it is relevant, but keep the self-hosting decision separate until the actual checkpoint, license, and serving requirements are available. Promised weights are not the same artifact as a downloadable model you can inspect and operate.

The session guidance also matters for evaluation fairness. A freshly configured run and a long conversation carried across model families are not equivalent tests. Use the same task fixtures and tool permissions, but apply each provider's supported conversation contract. Our Kimi K3.1 Pro guide covers current access and integration details; the original K3 guide remains background on the earlier release. Use LLM regression testing to compare completed changes.

4. Copilot Fusion moves orchestration into GA, while Projects stays in preview

GitHub made Copilot Fusion generally available on October 28. It combines the earlier HydraFusion work with automatic model selection, offering Standard, High, and Max effort levels. A request can use a single model, a cascade with a quality check, or a draft-review-revise workflow. Existing Auto and HydraFusion selections transition to the corresponding Fusion levels.

Billing is still based on the individual models used. GitHub offers paid users a temporary 30% model-cost discount, not a permanent flat price for an entire agent task. The Fusion announcement also explains organization model-policy controls. This is the GA successor to the research-preview work covered in edition #30, not an unrelated new foundation model.

Two accompanying releases change where that work lives:

• Copilot Projects is in public preview. It groups context, repositories, agents, and outputs around a goal, including work spanning multiple repositories. The app release notes distinguish that from a focused single-repository session.
• The connected cloud experience is GA on paid plans, with rollout beginning October 28. Sessions can continue across clients, using cloud sandboxes. GitHub's cloud announcement says sandbox billing begins in November and can draw from the shared Actions compute entitlement.

Local execution has a separate lifetime. The VS Code release notes describe Agent Host work continuing beyond an editor window, but the machine and its host process still have to run. Closing a window, disconnecting from a remote machine, and shutting down the execution machine are different operations.

My take: Fusion is worth comparing with a pinned-model baseline, but the larger operational change is persistent work across clients and projects. Teams need a clear owner for a run, its spend, and its approval state when someone moves from phone to desktop or hands a task to another person.

Measure the accepted result and the complete bill. Check the model policies before relying on the orchestrator's defaults, and do not assume a review pass is independent merely because it uses another model. Our Copilot app guide covers the product context; agent governance covers the responsibility that comes with ongoing work.

5. GitHub Agentic Workflows becomes generally available

GitHub's Agentic Workflows project also reached GA on October 28. You define a reasoning-based automation in Markdown, then compile it into standard Actions YAML. Example workloads include triage, CI failure analysis, documentation, and test maintenance.

The GA announcement describes read-only defaults, sandboxed execution behind an Agent Workflow Firewall, validated safe outputs, and a threat-detection job for proposed changes. GA adds more runner options, model routing, and budget forecasting. The default Copilot CLI engine requires a Copilot license, even though the project itself is available across GitHub Free, Team, and Enterprise.

My take: Start with an automation that creates a reviewable report or proposed pull request. Define the behavior when the evidence is insufficient, a test fails, or a dependency cannot be installed. A scheduled agent should not convert every obstacle into permission to broaden the task.

Keep deterministic checks outside the agent: test execution, required approvals, branch protection, and deployment authorization. Also budget for analysis that produces no code. A workflow that correctly decides not to change anything can still consume inference and runner time. Our agent versus workflow explainer helps decide whether the job needs reasoning at all; the production deployment checklist covers the surrounding gates.

Quick hits

• Fusion can route to local models, in experimental preview. The local-routing announcement distinguishes managed Windows downloads from manually configured CLI providers. Routing decisions still happen in the cloud. Local inference does not make the session offline.
• GitHub AI Scan now covers whole repositories in public preview. The release notes distinguish it from pull-request scanning. A merged fix does not automatically close an alert: run another scan to verify it. Account and repository eligibility apply, and scans consume Code Security AI credits.
• Eligible ChatGPT Pro and Business subscribers can claim API credits. The benefit guide lists plan-dependent amounts, ownership requirements, and a $500 monthly workspace cap for Business. This is not $500 for every subscriber. It does not increase signed-in Codex limits or cover Codex usage through API keys.
• Decisions API is now GA. OpenAI's October 22 update, after our previous issue, adds input caching, input_json, images in question instructions, and up to 400 questions per request. Recheck the beta-to-GA contract before changing an existing integration.
• Liquid AI released LFM3 for specialization. The launch post brings several modalities onto a shared small-model backbone. The 1.7B model card lists 256K context but explicitly does not recommend it for agentic coding or complex multi-step reasoning. Local task specialists and frontier coding agents are different categories.
• NVIDIA has a smaller local-assistant option. The Nemotron 3.5 Bolt E4B 4GB card describes an 8B-total, 4B-effective, text-only model in Q4_K_M, aimed at device assistants and multi-turn tools. Check the runtime footprint rather than treating the checkpoint name as a complete memory budget.
• Claude Platform has an Anthropic-managed Google Cloud path. The October 22 launch is GA with Google Cloud IAM and billing, currently with US and Global endpoints. Anthropic-managed and Google-managed offerings have different data processors and regional controls; shared billing does not make them identical services.

What I would test this week

Run one visual task on Terra, one repository task on K3.1 Pro, and the same bounded coding task through Fusion and a pinned model. Keep Carbon on the watch list rather than inventing public access. For automation, start with read-only analysis and inspect the generated workflow before enabling actions. The interesting metric is accepted work under a known budget and permission boundary, not how many releases you switched on.

FAQ

Can I use Gemini 4 Carbon through the public API today?

Google's announcement starts access with selected cyber defenders and Fairwind partners. It does not announce a general public API rollout date.

Are Kimi K3.1 Pro's weights already downloadable?

The launch post promises full weights by November 4. Hosted access is live, but that is a separate availability milestone.

Is Copilot Fusion the same as Projects?

No. Fusion routes and orchestrates models for a task. Projects organizes a broader goal, context, agents, and results. Their release statuses differ too.

Does local Copilot routing remove the cloud dependency?

No. The experimental routing feature still makes routing decisions in the cloud. Review the complete data path, not just where generation runs.

Are the new API credits unlimited?

No. Eligibility, plan amounts, linked organization ownership, and expiry rules apply. Check the benefit documentation and your actual grant before budgeting a workload around it.

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