Vercel AI Gateway gives any model web search and fetch on one key

A new ultrafast DeepSeek Flash lands on the same gateway, a calibrated-judge trick cuts eval costs ~4x, and Copilot CLI switches on shared repo indexing.

Nowline NOV 8 1:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Any model gets a web browser, one key

    Vercel's AI Gateway now exposes DeepSeek Search and Fetch as tools any tool-calling model can invoke — live web search plus full page retrieval — without a separate search API or provider account. One key, with spend tracking and request traces. It's the quickest way to bolt current-web grounding onto whatever model you already route through the gateway.

  • DeepSeek V4.1 Flash lands on the gateway

    DeepSeek V4.1 Flash 1017 (Ultrafast) is now callable through Vercel AI Gateway with that same single key — no DeepSeek account required — with spend tracking and traces included. A fast, cheap model to slot into latency-sensitive routes.

  • Cheap judge, Opus on call: ~4x cheaper evals

    Hightouch ran a calibrated small model (Jev) that scores every case 0-1 and escalates only the uncertain 30-90% band to Claude Opus 5.5. On their identity-resolution test that hit 98.8% accuracy versus Opus's 99.0%, at $0.88 vs $3.86 per 1,000 cases — roughly 4.4x cheaper and ~3x faster. The takeaway for your own eval pipelines: calibration beats raw accuracy — GPT-6 Luna scored higher solo but, being overconfident, flagged none of its own errors for escalation.

  • Copilot CLI switches on shared repo indexing

    Copilot CLI 1.0.108 turns on managed shared tgrep indexing by default for eligible repos and adds /index commands, so repo context is prebuilt instead of re-scanned each session. Autopilot under Assisted Permissions also stops halting on permission prompts — it tries another approach and reports what it couldn't do — alongside a new Windows enumerate-only directory grant.