話題 / Topic
Vals AI compared GPT-6 Sol and Claude Opus 5.5 as single agents versus teams. Even paying 1.8–5.1× the cost of a single agent, the teams improved scores significantly in only one of four comparisons—putting hard numbers behind the practitioner's instinct that reasoning strength beats head count.
Vals AIがGPT-6 SolとClaude Opus 5.5を1体とチームで比較。チームは1体の1.8〜5.1倍のコストを払っても、成績が有意に伸びたのは4回中1回だけ。「頭数より推論の強さ」という実務の勘どころを数字で示した。
OpenAI expands Codex's default context from 272K to 500K—up to 1M, with no long-context surcharge on subscription usage. It effectively rolls back a number it had quietly cut itself weeks earlier, though concerns about quota consumption and retrieval accuracy remain.
OpenAIがCodexの既定コンテキストを272K→500Kへ拡張し、最大1M、サブスク利用分は長文追加料金なしに。数週間前に自ら削った数字を巻き戻した格好だが、クォータ消費や検索精度への懸念も。
Released by Sakana AI with UC Berkeley, MASS has a single model play executor, orchestrator, and grader, using agent collaboration as training material to self-improve. It posted up to 1.6x on research benchmarks, yet on coding tasks like SWE-bench it fell below the base model.
Sakana AIがUC Berkeleyと公開したMASSは、一つのモデルに実行・指揮・採点の三役を兼ねさせ、エージェントの協働を教材に自己改善させる手法。研究系ベンチで最大1.6倍を記録した一方、SWE-bench等のコーディングでは土台モデルを下回った。
Microsoft has released Decision-1, a model built purely for judging. It writes no prose—it returns only calibrated per-option probabilities as JSON. Input costs $0.042 per 1M tokens; output is free. It lets you cheaply carve out the “checkpoints” of agent proceed/stop decisions, model routing, and A
Microsoftが判定専用モデル Decision-1 を公開。文章は書かず、選択肢ごとの較正済み確率だけをJSONで返す。入力100万トークン$0.042・出力無料。エージェントの続行/停止判断やモデルルーティング、AI採点の"関所"を安価に切り出す。
Harvard analyzed roughly 300 million development events across 700 companies. After adopting AI agents, code grew 30% but task resolution stayed flat. The bottleneck had shifted from "writing" to "reviewing and verifying."
ハーバードが700社・約3億件の開発イベントを分析。AIエージェント導入でコードは3割増えたのに、課題解決率は横ばい。ボトルネックは「書く」から「レビューして確かめる」へ移っていた。
At "Gemini at Work 2026," Google unveiled "Gemini agent," an all-purpose agent that takes a goal and handles planning, coding, and execution. At the same time, it is ending standalone sales of Gemini Code Assist and consolidating it into Antigravity. We break down the impact on development teams and
Googleが「Gemini at Work 2026」で、目的を渡すと計画・コーディング・実行まで任せる万能エージェント「Gemini agent」を発表。同時にGemini Code Assistの単体販売を終了しAntigravityへ集約します。開発チームへの影響と囲い込みの懸念まで整理します。
