CLAUDE OPUS 4.8 • MESSAGES API

Opus 4.8
Utilities

Small, self-contained Python scripts that demonstrate real Claude Opus 4.8 capabilities on the Anthropic Messages API.

Explore the three demos View source
Read the companion post on Substack
Claude Opus 4.8 — Frontier Intelligence
3 self-contained scripts
Zero dependencies beyond the SDK
Each prints exactly what it demonstrates
Get started

Requirements

All scripts target the model claude-opus-4-8.
Runnable demos

The Three Utilities

Each script is a single file. Run it and watch the behavior (and the token accounting) in real time.

Effort
effort_demo.py

Same model, same prompt — turn the effort dial from low to max and watch output tokens rise.

See the effort dial in action →
Caching
cache_safe_system_injection.py

Inject mid-conversation instructions as system messages without destroying your prompt cache prefix.

Preserve cache across turns →
Independence
disagreement_probe.py

Test whether the model pushes back on confident-but-wrong claims or caves under pressure.

Calibrated pushback vs sycophancy →
01

effort_demo.py — Effort as a dial

Run the exact same prompt on the exact same model at every effort level. The only thing that changes is output_config={"effort": "..."}.

Key insight

Output tokens (and quality) scale with the effort setting. You don’t switch models — you turn a dial. high is the default. xhigh and max are Opus-tier only.

Run it
python3 effort_demo.py
What the script does

It loops over ["low", "medium", "high", "xhigh", "max"], calls the model with thinking={"type": "adaptive"} + the chosen effort, and prints the answer plus input/output token counts.

You’ll see output length (and usually depth) increase as effort goes up — all on the identical base model.

Source: effort_demo.py
02

cache_safe_system_injection.py — Mid-conversation system messages

How do you give new instructions to a long-running agent without blowing away your expensive prompt cache?

Key insight

Put the new instruction as a {"role": "system"} entry inside the messages array. The top-level system (the cached prefix) stays byte-identical.

Run it
python3 cache_safe_system_injection.py
Three calls it prints
  • turn 1 — cold start, writes the big system block into cache (cache_write > 0)
  • turn 2a (good) — injects a one-off system message in the messages list → cache prefix survives (cache_read > 0)
  • turn 2b (bad) — mutates the top-level system → cache is invalidated, full rewrite (cache_write > 0, cache_read = 0)

The gap between 2a and 2b is the real cost of cache invalidation in production agents. Requires a Claude 4+ model.

Source: cache_safe_system_injection.py
03

disagreement_probe.py — Does it push back?

Frontier models can be sycophantic. This script probes native behavior with no system prompt coaching.

Two tests
A) Correction test — four claims presented as fact (three false, one true). Does it agree with everything or correct the falsehoods?
B) Pressure test — after an initial response, push back hard with authority. Does it hold its ground or fold?
Run it
python3 disagreement_probe.py

The script prints both the model’s answers and the token usage. Look for whether it confidently agrees with provably false statements or stands its ground under pressure.

Source: disagreement_probe.py
These scripts make live API calls.
Keep max_tokens and effort modest while experimenting. Token costs add up quickly on the higher effort tiers.
Clone the repo on GitHub Read the full Claude Opus 4.8 post