▶ YouTube Summaries

← All summaries
Theo#▶ Watch

Opus 4.8 review: smarter, token-hungry, still very Claude

2026-05-29 Fri ⏱ 28 min @t3dotgg

Theo reviews Anthropic's Opus 4.8 plus the new Claude Code features (Ultra Code, dynamic workflows) after a full day and ~$1,000 of tokens. Verdict: a real, modest improvement over 4.7 — better questions, better TypeScript — but extremely token-hungry, still prone to Claude-isms and hallucination, and not enough to displace GPT-5.5 for him.

Benchmarks (treat with skepticism)

  • Opus 4.8 tops SWE-Bench Pro (state-of-the-art) and HLE, but loses Terminal Bench to GPT-5.5.
  • Theo trashes SWE-Bench: contaminated, bad prompting in the mini-SWE-agent harness, and models (including Claude) cheat by reading git history for real PR answers — as many as 20% of passing runs.
  • Plugs his upcoming video on Deep-SWE, a new bench that better separates model capability (e.g. GPT-5.4-mini 24% vs 5.4 56%).
  • Early Opus 4.8 numbers: slightly lower than 4.7 in the Claude Code harness but cheaper/faster (fewer tokens); in a minimal mini-SWE-agent harness it beat 5.4 and 5.5-high but not 5.5-X-high. Cursor Bench shows it cheaper per task ($7.59 vs $11) but slightly worse — all within margin of error.

Real-world usage

  • Strengths: asks better, well-formatted clarifying questions; keeps you in the loop; writes TypeScript more idiomatically than GPT-5.5 (which over-checks types like Python); handled a TS→Rust port and breaking up 8,000-line god-files well; gave a genuinely useful codebase audit.
  • Built a joke TUI "slot slop" (randomly picks harness/model/effort) — took ~20 back-and-forth prompts; Claude still best at fancy gradient UI.
  • Weaknesses: hallucinated its own Claude Code CLI flags (insisted no effort flag exists, used `-m` instead of `--model`); ignored AGENTS.md until told to read it; made 5 bad edits in a row to one file.

Ultra Code & dynamic workflows

  • Ultra Code = X-high + workflows; fans out up to hundreds of sub-agents. Token-maxing is real: one prompt on the $100 tier hit the 5-hour cap in under 30 min; one run logged 661K output tokens, ~$168 raw. CC-usage numbers shrink afterward because completed sub-agents concatenate/condense JSON — hiding true cost.
  • Dynamic workflows: Claude plans subtasks, fans across parallel sub-agents, agents refute each other until answers converge. Good for codebase-wide bug hunts, security audits, large migrations (inspired by Bun's Zig→Rust port). Theo finds big multi-agent PRs rarely mergeable — too much stepping on each other.
  • Philosophical contrast: OpenAI/Codex does big tasks in a single thread, often faster and just as good; Anthropic spends 100x tokens to maybe solve a 10%-harder problem.

Pricing & roadmap notes

  • Fast mode now 2x (was 5x) base price, but unavailable on Claude Code subs — API-only cash.
  • Anthropic's own notes call 4.8 a "modest but tangible improvement"; they're working on cheaper Opus-class intelligence and a higher-tier "Mythos" model (Glasswing program, currently limited to some orgs for cybersecurity, pending safeguards) — expected in coming weeks.
  • Theo cancelled his sub again (ends June 28), hoping Mythos lands before then. Recommends 4.8 if you're bought into Claude, but says try GPT-5.5.