Gemini 4 Carbon goes to defenders first, GitHub ships Copilot Fusion and Agentic Workflows, OpenAI gives subscribers up to $500 a month in API credits
Today's thread is letting the system pick the model and managing the bill as a budget: GitHub moved model routing, agents in CI, memory and usage-based billing to general availability in one go, while Microsoft and AWS turn agent execution and approvals into managed services. At the frontier, Google again ships to defenders before developers.
Security benchmarks: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, 71.0% on CWE-bench for patching; up 89 Elo to #4 (1768) on WebDev Arena.
Access starts this week with a subset of cyber defenders, then phases to trusted partners through the Fairwind Program; no date yet for developers, businesses or consumers.
Introductory price matches Argon: $2 input and $10 output per million tokens, cached input 95% off; after the introductory period it is $4 and $20.
Google says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0% (0.4% attempts), makes about half as many unsupported claims as Argon, and had a 0.0% attack success rate on Gray Swan's indirect prompt injection benchmark.
Builder's takeCarbon follows Argon's playbook: defenders first, developers later. Two things are worth doing now: keep the eval set you used for Argon so you can compare on the day the API opens, and budget at the $4/$20 list price, not the introductory one. For long-running background jobs like PandaClaws, nothing needs to change until there is API access.
It merges Project HydraFusion's multi-model orchestration with auto model selection; pick Copilot Fusion in the model picker and then an effort level (High, Max). Workflows include Cascade (an efficient model drafts, a quality check decides whether to escalate) and Critique (one model drafts, another reviews, the first revises).
Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the average API cost per task, about half the tokens and about 40% less time.
Billing is per model used at standard rates, with a limited-time 30% discount on paid plans; admins control which models Fusion may use and can make it the default.
Available in VS Code, the Copilot app, Copilot CLI and JetBrains; an experimental preview routes between local models (such as MAI Code 1.1 Flash and Qwen3.8 Flash Next) and cloud models, though routing decisions still happen in the cloud.
Builder's takeCascade is the part to copy: a cheap model drafts, a quality check runs, and only then do you escalate. That is how PandaClaws' generation backend should be wired, rather than fixing one model. Before switching your team default, run a week of Fusion against a fixed model and compare per-PR cost with the by-client usage metrics from 10-25.
An open-source project from GitHub Next for tasks like issue triage, CI failure analysis, documentation updates and test writing; it compiles to standard Actions YAML, reusing your runner groups and policies.
Safeguards: read-only permissions by default, a sandboxed container behind the Agent Workflow Firewall, safe outputs validation, and a dedicated threat detection job that scans all proposed changes.
New at GA: self-hosted ARC runners with Docker-in-Docker, gh aw forecast out of experimental with monthly AI credit ranges, per-workflow model providers and a model router, and isolation in Cloud Hypervisor or NVX micro-VMs.
Available on GitHub Free, Team and Enterprise; the default Copilot CLI engine requires a Copilot license.
Builder's takeFor a small team, CI failure analysis and issue triage are the first wins: read-only by default plus validated outputs means it can only propose. I would start with a single workflow that comments an analysis on failed CI runs, and run gh aw forecast first to see the monthly credit range.
Memories can be created in the Copilot app, CLI, VS Code, cloud agent, code review, Slack, Teams and GitHub Mobile, and are reused across sessions and clients.
Among Copilot CLI sessions in the public preview where memory was used, 93% showed better adherence to developer preferences and 42% showed more efficient workflows.
Dreaming reflects on past work: it curates personal and repository memories, proposes changes to skills and custom instruction files as pull requests, and turns recurring accepted review feedback into code review guidance; runs consume AI credits.
Copilot can also search code across repositories in an organization through /research, including regex and semantic search.
Builder's takeRouting instruction-file changes through pull requests is the right line: memory that changes how a whole team works should be reviewable like code. I would turn on repository memory but hold dreaming PRs to the same review as any code, and skim personal memories now and then, because stale memories turn into silent wrong defaults.
Available now on GitHub Enterprise Cloud for customers working with a GitHub or Microsoft account team; self-serve opens in stages, current subscriptions are unaffected, and Copilot Business and Enterprise remain on sale.
Members without a GitHub Enterprise license can get Copilot Standalone at $10 per user per month, with the same features and models as Copilot Enterprise but no monthly AI credits per seat.
There is no included usage allowance; with no budget, code completion and next edit suggestions keep working, while Copilot Chat, code review and the cloud agent need budget; an exhausted budget never removes access.
Builder's takeWhat you manage changes from seat counts to budget: set per-team budgets and alerts before rollout, then open up Chat and the cloud agent. Completions are now effectively free, so leave them out of the rollout cost.
The allowance depends on the Pro plan or the Business workspace's paid seats, up to $500 a month; Business credits are pooled and capped at $500 per workspace per month, or $6,000 per annual billing cycle.
The subscription must be paid, active and at least seven days old, with free trials and complimentary subscriptions excluded; you must own both the subscription and the API organization, using the same email.
Credits expire at the end of the billing cycle and don't roll over, apply only to the linked API organization, are used before purchased credits, and also cover built-in tools such as web search, file search and Code Interpreter.
Builder's takeFor a small team this is a fixed monthly prototyping budget. I would link it to a separate API organization used only for experiments, not production, or the production bill jumps the moment credits expire. Since credits reset every cycle, they are best spent on eval runs, such as comparing Terra and Sol.
Weights are on Hugging Face and ModelScope: it transcribes 40 languages, translates 31 of them to or from English and attributes speakers in 13; recordings up to an hour and up to eight speakers; output is steered with natural-language prompts, with word-level timestamps.
5.33% average WER on the Hugging Face Open ASR English leaderboard, first among open-weight models; 4.62% mean WER across seven languages; 2.8% AA-WER on Artificial Analysis, tied for first among open-weight models with about 4B active parameters (30B total) versus 24B for the model it ties.
NVIDIA also released Granary v2 multilingual data and its curation pipeline; in a medical case study, fine-tuning cut WER on held-out speakers by 45.8% relative to the best prompt. Language support differs by task, audio Q&A is trained in English, and quantized versions are coming.
Builder's takeAI Interview currently stitches ASR to a separate diarization step; one model that returns speaker labels and word-level timestamps would remove a whole layer of alignment code. But the leaderboard results are English: check the model card for which tasks support Chinese, run it on our own recorded Chinese interviews, and confirm the license before deciding to switch.
Agents are defined declaratively; reusable skills package instructions, scripts and resources; Foundry and MCP tools connect with explicitly configured authentication and approvals; agents work with files and code and delegate bounded tasks to built-in subagents.
Agent definitions can live in source control, and you can choose a supported Foundry model without rewriting the orchestration.
Long-running resilience for hosted agents (bring your own framework and loop) is generally available: Foundry detects interruptions and restarts the work with its original input, while your code is responsible for saving progress and preventing duplicate actions.
Builder's takeSkills plus MCP plus subagents is now the common shape across Anthropic, GitHub and Microsoft: write agent capabilities as skills and MCP tools rather than tying them to one platform. The part to remember is resilience: a restart replays the original input, so any write action such as sending email or creating a report must check whether it already happened. PandaClaws' publish step needs that idempotency check whatever platform it runs on.
It only reviews actions where the user's setting is "ask" each time; denied actions stay denied, it never grants new access, each automatic approval covers only the exact action reviewed, and a failed or timed-out review goes to the user.
Higher-impact actions go to the user: sending content outside the organization, changing permissions, deleting content, making payments, and actions beyond what the user asked for.
On by default for all users; admins control access with custom permissions, users can turn it off for conversations and scheduled runs, and AWS CloudTrail records each review and each settings change.
Builder's takeThis is the same idea as Copilot assisted permissions from 10-28: a model screens, people see the exceptions. Its five always-escalate categories (outbound content, permission changes, deletion, payments, out-of-scope actions) make a ready default checklist for your own agents, and every automatic approval should be logged so it can be audited later, as CloudTrail does.
Stateless and zero-dependency, built on the Interactions API: generate text and media, analyze files, and run Gemini managed agents.
Subcommands include image, tts, generate, analyze (PDFs, images, audio, video, CSV, text and public YouTube URLs), transcribe (md, text, json or srt with speaker labels and timestamps), and agent run, status and cancel.
The docs include a snippet that tells a coding agent how to use it, and the install script can pin a release with GEMINI_VERSION.
Builder's takeGiving an agent a capability through a CLI instead of an SDK is the cheapest route, and it fits neatly into a skill. I would start with one-off CI jobs such as generating images; but pin the version with GEMINI_VERSION rather than piping the latest install script into bash, with the LiteLLM poisoning from 10-24 still fresh.
Researched and drafted with AI assistance; editorial standards and views set by Darius.
Every Monday, the week in AI in one email
Every Monday morning: the 6 AI stories that mattered most in the past seven days, with my take, in your inbox.
Your email is only used to send the weekly and is never shared. Every email has a one-click unsubscribe link. See the latest issue →