Opus 4.8 Just Dropped. Here's How To Actually Use It.

Product reviewed

Claude Opus 4.8 (via Claude Code / Cloud Code, available “everywhere today”), presented as an upgrade to Opus 4.7.

Key improvements / main features

Built on Opus 4.7 with:

• Sharper judgment
• More honesty about its own progress (better alignment between claimed vs. actual work)
• Ability to work independently longer on long-running tasks

Effort control in Claude / Claude Code (Cloud.ai)

A new “effort control” lets users decide how much effort Claude puts into tasks.

• Default = high effort
• Effort levels include low / medium / high, plus higher modes that can increase output quality and/or autonomy
  • The video mentions XH, high, max, ultra, and “X high + workflows”
• Higher effort:
  • Costs more tokens
  • Can be slower

Slider effect (described):

• Move left → faster outputs
• Move right → more thorough

Dynamic workflows feature in Claude Code

• Intended to handle very large scale problems
• Accessible by typing “workflows”
• Also referenced via CLI effort modes (e.g., “Ultra code” / workflows combo)

Token / rate-limit adjustments (API via Cloud Code)

• Same pricing as Opus 4.7 for input/output tokens
• Increased API rate limits to support higher token usage from higher effort levels
• Other limits (e.g., weekly session limits / 5-hour rolling window) are stated as unchanged

Benchmarks & comparisons (as claimed in the video)

• The presenter reports benchmarks “look amazing” and claims:
  • Opus 4.8 is better than 4.7 across major categories
  • Opus 4.8 is also better than “GBT 5.5” (likely GPT-5.5, per transcription)
• The video emphasizes caveats:
  • Benchmarks can be marketing-oriented
  • The “best model” depends on your use case

Example comparison mentioned

• The presenter claims Codeex with GPT-5.5 can be better at “agentic computer use” than Opus 4.7/4.8, even if Opus looks better “on paper” for that task.

Specific rating / numerical score mentioned

In an honesty / misaligned behavior evaluation:

• Lower score is better
• Opus 4.8 scores at “almost half” of Opus 4.7 and Sonnet 4.6 (No exact number was provided; only the relative improvement.)

Issues from Opus 4.7 that 4.8 aims to fix (and how)

The video summarizes community complaints about Opus 4.7, then ties Opus 4.8 improvements to them.

Reported 4.7 problems

• Felt lazy / gave up early on tasks
  • Mentions “/goal” as a band-aid style improvement in 4.7 (goal-directed behavior), implying earlier attempts to force persistence
• Overly rigid safety overreach
• Attitude / sassy / pushback / stubborn tone
• Important nuance: sometimes it’s not purely “the model”—it can be user prompting / strategy / effort selection

4.8 improvements intended to address them

• More honesty & self-correction (especially the progress/claim mismatch issue)
• More sustained autonomy on long tasks (less “giving up”)
• Warmer, more collaborative vibe
• Better efficiency and quality of life, including:
  • Better tool calling
  • Better reasoning
  • Better question asking
  • Better token efficiency (stated as “based on documentation,” not fully verified in the video)

User experience guidance (how to use it effectively)

The presenter’s guidance focuses on changing how you prompt/configure compared to 4.7.

1. Effort is the #1 lever
   • Don’t leave it at defaults if you see poor behavior
   • Low effort may cause “lazy” completion on hard tasks
   • High / extra-high effort on simple tasks may cause overthinking/overengineering and slower outputs
   • Switching effort can feel like switching to a “different version” of the model
2. Tell it what to do, not what not to do
   • Example prompting docs were said to emphasize positive instructions over prohibitions
3. Give the “why” behind instructions
   • Add context so it follows style/constraints more reliably
4. Default behavior: reasoning before tool use
   • The model may reason internally before calling tools/sub-agents
   • If you need more external context first, adjust prompting and/or effort
5. Response length is calibrated
   • Shorter answers for simple lookups
   • Longer answers for open-ended analyses
6. When switching 4.7 → 4.8, don’t blindly reuse everything
   • Re-check workflow behavior and adjust effort/workflow settings

Pros (based on the video)

• Tangible improvement over 4.7, especially in honesty and task persistence
• Better self-correction
• Longer independent work on extended tasks
• More collaborative / warm tone (as described)
• More control via:
  • Effort slider
  • Better CLI / Cloud Code controls
• Dynamic workflows for large-scale problems
• API rate limits adjusted to support higher-effort modes

Cons / cautions (based on the video)

• Higher effort costs more tokens (tradeoff: quality vs cost/speed)
• Benchmarks may not match your actual workload (explicit warning)
• Early rollout issues: some early reports of bugs were mentioned
• Some “negative” 4.7 behaviors might be influenced by user configuration/prompting, so expectations should be managed

Overall “verdict” / recommendation (from the video)

• Recommended for most Claude Code users who had issues with Opus 4.7, especially:
  • early stopping / laziness
  • incorrect or overconfident progress claims
  • overly rigid or overly attitude-driven interactions
• Still, users should tune effort and prompt positively with context (“why”) rather than assuming 4.8 will automatically fix every workflow frustration.

Benchmarks and specific behaviors may vary with your setup; don’t expect identical performance across all “agentic computer use” scenarios.

Unique points mentioned about the product (consolidated list)

• Released May 28, 2026
• Built on Opus 4.7; adds sharper judgment
• More honesty about progress; better self-correction
• Works independently for longer
• Same pricing as Opus 4.7 (input/output tokens)
• API rate limits increased for higher token usage from effort levels
• Default effort = high; can switch effort in Cloud Code
• Effort modes include low/medium/high plus higher settings (max/ultra) and a workflows combo
• Moving the effort slider affects speed vs thoroughness
• New dynamic workflows feature for very large scale tasks
• Still has a 1 million context window (Opus)
• Positioned as fixing major 4.7 community issues: laziness, safety overreach, attitude
• Guidance emphasizes:
  • effort control
  • positive prompting
  • include rationale (“why”)
  • expect reasoning before tool calls
  • calibrated verbosity
• Caution:
  • Benchmarks ≠ your use case
  • Some tasks (like agentic computer use) may favor other models in practice
• Mentions a future model class “Myt(h)os” with stronger cyber safeguards (not the product focus, but part of the discussion)

Different speaker/view contributions

• Single main speaker (presenter): makes all claims, comparisons, and usage guidance.
• Mentions “community reports” and “people saying”, but these are referenced rather than delivered by separately identifiable speakers.