Skip to content
Daily BriefNewsDaily Brief

5 AI stories from September 12, 2026: Sakana Fugu Ultra v2 outscores Opus 5 via open-weight orchestration, GitHub Copilot auto-resolves code review, Claude Code v2.1.269 plugin evals, Jira in Copilot, and OpenAI GPT-Live-1 at $0.05 per minute

· by Pondero Newsdesk · 5 stories

AI news daily brief: 2026-09-12

Five stories today: a new multi-agent orchestration architecture from Sakana AI that outperforms frontier models on visual reasoning without calling them, two GitHub Copilot changelog drops (code review and weekly releases), a Claude Code version update with a plugin eval framework, and OpenAI's full-duplex voice model arriving in the API.

Sakana AI Ships Fugu Max and Fugu Ultra v2: A Multi-Agent Orchestration Model That Beats Frontier Models Without Using Them

Sakana AI released Fugu Max and Fugu Ultra v2 on September 11, 2026. Rather than a single monolithic model, Fugu is a trained orchestration system that routes tasks across a pool of open and specialized models and recursively calls itself on complex problems. Fugu Ultra v2 scored 48.3 on the Chartography visual-reasoning benchmark versus 27.3 for Opus 5, without drawing on Fable 5, Fable 5.1, or GPT-6 Astra in its model pool, per Sakana AI. Fugu Max is the cost-first tier at $2/$6 per million input/output tokens, adding NVIDIA Nemotron models and cutting costs 40-60% versus competing frontier models.

Third-party reproducibility of those benchmark scores is the first signal to watch. Context pricing above 272K tokens is the second for teams running agentic workloads.

Full story: Sakana Fugu Max and Fugu Ultra v2

GitHub Copilot Sept 11: Code Review Now Auto-Resolves Its Own Comments, and VS Code Agent Activity Now Appears in Usage Metrics

Two September 11 GitHub Changelog entries brought substantive code review changes. Copilot auto-resolves its own comments when a developer addresses them in a new commit, writes contextual commit message suggestions when applying Copilot code suggestions, and uses a broader shell-tool set to detect high-severity issues. The Lite effort level now runs an ensemble of agents instead of a single reviewer, raising high-severity finding detection by 47% and medium-severity by 31% at approximately 8% lower cost. A separate entry adds VS Code Agents activity to Copilot usage reports, giving admins per-agent adoption data.

Watch whether the Lite ensemble approach rolls out to Standard and Max effort levels.

Full story: GitHub Copilot code review Sept 11

Claude Code v2.1.269 Ships Plugin Evals, Output-Style Switching, Agent Map in VS Code, and Workflow Fan-Out Controls

Anthropic released Claude Code v2.1.269 on September 11, 2026. The headline addition is claude plugin eval: a command that runs a plugin's eval suite against Claude Code, scores results with six grader types (four free computational graders plus two billable model-based graders), and produces JSON and HTML reports with differential scoring, per the Claude Code changelog. Default is three runs per eval case. A new /output-style [name] command switches output styles in any session, including Remote Control and headless runs. VS Code gains an agent map that shows all active sub-agents during Workflow fan-outs. CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS (1-256) caps fan-out concurrency, giving teams a direct cost-control lever.

Whether the plugin eval framework becomes a requirement for the Skill Marketplace is the development to watch.

Full story: Claude Code v2.1.269

GitHub Copilot Weekly Releases Sept 7: Jira Integration in the Copilot App and AI Scan REST APIs for PR Code Scanning

GitHub's September 7 weekly releases post, published September 10, 2026, added Jira integration in the Copilot app, letting developers view, create, and update Jira issues without leaving the Copilot workflow, per the GitHub changelog. The same update brought VS Code 1.137 agent automations in public preview (scheduled recurring tasks), VS Code voice mode and PR/issue review in the Agents window (both experimental), and JetBrains enterprise sandbox controls in public preview.

Whether Jira is the first of a broader project-management integration push (Linear, Asana) is the commercial signal to watch.

Full story: GitHub Copilot Jira integration and AI Scan

OpenAI Releases GPT-Live-1 in the API: Full-Duplex Voice at $0.05 per Minute That Can Delegate Reasoning to Other Models

Full-duplex voice reached the OpenAI API on September 10, 2026. GPT-Live-1, which has powered ChatGPT Voice since July, is now available to developers at $0.05 per minute for the front-end voice layer, per OpenAI. The model listens and speaks simultaneously and can delegate reasoning and tool calls to paired backend models, including GPT-6 Astra. OpenAI cited a Tau3 score of 86.2% versus 45.7% for prior STT-LLM-TTS chains, and end-to-end latency of 0.798 seconds versus 1.41 seconds for the pipeline approach.

At $0.05 per minute, the pricing changes the cost equation for voice-first AI products. How latency holds when delegating mid-conversation to a heavier backend model is the practical question for production deployments.

Full story: OpenAI GPT-Live-1 API

Sources