Co-Design Bench: top coding agent solves 26% of real tasks

GPT-6 Astra leads yet stalls on real reworks; the harness matters as much as the model. Plus a sales MCP, Claude reads your bank, Gemini 4 Argon doubles Dec 1.

Nowline NOV 9 10:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • The best agent still leaves 3 in 4 real tasks unfinished

    On Co-Design Bench — real rework requests for existing apps — top scorer GPT-6 Astra fully solved just 26.4% (41% on multimodal tasks, under 18% repo-wide), with Opus 5 and Qwen3.8-Max next. The harness mattered as much as the model — Qwen3.8-27B swung from 56% to 75% on the same tasks by changing environment — and more tokens didn't help: GLM-5.2 burned ~6x the tokens and still scored worse. Budget for manual cleanup, and report your model and environment together.

  • 1,200 sales and marketing actions, now one MCP away

    AXL launched an AI-native sales and marketing OS that exposes 1,200+ actions — lead records, email, outreach — to ChatGPT, Claude, or Cursor over MCP, with a free tier. If you're a solo founder, you can run outbound from the same agent you code in instead of wiring CRMs together by hand.

  • Claude can read your bank accounts now

    Personal Finance in Claude (beta, US Pro/Max/Ultra) links accounts through Plaid and reads balances plus up to 24 months of transactions — read-only, no transfers — then turns them into reports or dashboards. Watch the data trail: disconnecting stops new reads but leaves details in existing chats and any shared artifact links until you delete them.

  • Gemini 4 Argon's API price doubles December 1

    Argon's introductory rate ends Nov 30: from Dec 1, input goes $2 to $4 and output $10 to $20 per 1M tokens, and cached input keeps its 95% discount so it doubles in dollars too. If Argon is in your pipeline, lock budgets or move volume before month-end.