Claude Opus 5.5 Intelligence, Performance and Price Analysis (Max)
TL;DR Highlight
Opus 5.5 max tops the Intelligence Index at 58 but is verbose and pricey.
Who Should Read
Backend developer using the Claude API who must decide which reasoning effort of Opus 5.5 to run given cost and performance.
Core Mechanics
- Index score 58, about double the median of 27
- Measured on Artificial Analysis Intelligence Index v4.3.2.
- It is shown as #1 among comparable models.
- Output tokens are over 3x the median
- It generated 260M tokens while running the Intelligence Index.
- The median for comparable models is 85M.
- Priced at $4 input and $20 output per 1M tokens
- The medians are $2 input and $10 output.
- The cache discount is 95%.
- Averages $5.98 per Index task
- This is the weighted average cost of running one Index task.
- Cost rank is shown as 112 of 114.
- 95 tokens/s, faster than the 85 median
- Time to first token is 562.75s versus a 3.81s median.
Community Reactions
- Max ran out of budget before answering
- A user tried an SVG pelican-on-bicycle prompt twice with max and both runs exhausted the 128,000 token budget while still reasoning.
- They suspect max may be nearly useless.
- Max got stuck, medium works well
- Another user saw max get stuck and uses medium as default for both short and long loops.
- One commenter prefers high since benchmarks plateau after it and it beats Fable.
- Half the cost of Opus 5 at high effort
- One commenter says cost per task is half of Opus 5 when comparing high to high.
- Another went back to Opus 4.8 because Opus 5 lost track of instructions.
- Doubts about benchmarks
- One commenter wonders whether models are trained to ace these benchmarks.
- Using Claude Code and Codex on real code, they found Claude better at planning and executing and Codex better at finding bugs.
How to Apply
- Compare three effort levels on your eval set today
- Run 20 internal eval questions at medium, high and max.
- Tabulate accuracy, output tokens and latency.
- Cap max with a token budget and fallback
- Set a max_tokens limit and timeout on max calls.
- Retry at high or medium if no answer returns.
- Compute real cost per task before choosing
- Sum input, output and cache tokens from your logs per request.
- Multiply output tokens by the $20 rate when comparing verbose models.
- Rerun your eval set weeks after launch
- Run the same eval weekly and record scores.
- Route back to the previous model if scores drop.
Try It
# Paste into a terminal (simple cost-per-effort calculator)
# Replace these numbers with token counts from your logs
IN=2000000; OUT=1500000; CACHE_HIT=0
# Prices from the article: $4 input, $20 output per 1M tokens
python3 -c "
in_t=$IN; out_t=$OUT
print('cost USD:', in_t/1e6*4 + out_t/1e6*20)
"
# Next: run it on logs from medium, high and max, and compare cost and accuracy side by sideTerminology
Related Papers
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.
90% of US working hours are spent on tasks not covered well by AI benchmarks
Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage
Postgres 19 delayed after AI reviewers find more bugs than humans can fix
LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted
Show HN: Gemini 4 Argon plays StarCraft
LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.
A Plausible Explanation Is Enough to Mislead Multimodal Judges
Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.
Neural networks are clean
Looking at neurons directly shows most have one main job, challenging superposition.