OPUS 4.8!!! (also maybe GPT5.6??)

TL;DR

• 1

  Opus 4.8 lands fast and keeps the same base price — Anthropic says the new model improves judgment, honesty about progress, and long-running autonomy while holding standard pricing at $5 per million input tokens and $25 per million output tokens.

• 2

  Fast mode got materially more attractive — Berman highlights that Opus 4.8 fast mode is now 3x cheaper than before, making it effectively around 2x the standard price while delivering roughly 2.5x the speed.

• 3

  Benchmarks look strong, but the vibe check still matters more — Opus 4.8 posts 69.2% on SWE-Bench Pro and beats GPT 5.5 on Humanity’s Last Exam, yet Berman argues benchmarks like DeepSuite better reflect real coding feel, where many users still prefer GPT 5.5.

• 4

  Dynamic workflows is Anthropic’s big product swing — The new Claude Code feature can orchestrate tens to hundreds of parallel sub-agents for bug hunts, migrations, and adversarial verification, though Berman immediately heard “my API bill skyrocket” when reading the description.

• 5

  Live testing showed both frontier capability and jagged intelligence — Opus 4.8 built a playable 3D soccer game, a mostly working Rubik’s Cube solver, and a rough video editor, but still stumbled on the car-wash logic question and produced a weak homepage redesign after 29 minutes and 18 agents.

• 6

  Anthropic looks newly unconstrained on compute and cash — During the stream Berman spots news of a $65 billion Series H at a $965 billion post-money valuation, ties it to Anthropic’s xAI/Amazon/cloud deals, and notes his Max plan usage was shockingly low despite heavy parallel-agent testing.