Claude Opus 5.5
TL;DR Highlight
Opus 5.5 cuts costs 40% and runs 30% faster than Opus 5 while leading benchmarks
Who Should Read
Backend developers running agent and coding workloads on Opus 5 who worry about API cost and speed
Core Mechanics
- 40% cheaper than Opus 5
- Input and output tokens are $4 and $20 per million, 20% less than Opus 5.
- Cache reads are $0.20 per million tokens, 60% less than Opus 5.
- Anthropic says typical workloads at default settings cost 40% less than Opus 5.
- 30% faster output, plus Fast mode
- Output generation is more than 30% faster than Opus 5.
- Fast mode in Claude Code and Claude Platform is up to 2.5x speed at $8 input and $40 output per million tokens.
- Leads coding benchmarks
- Terminal-Bench 4.0 is 66.4%, versus 55.8% for Fable 5.1 and 52.3% for Opus 5.
- FrontierCode v1.1 is 54.4% and CursorBench 4.0 is 57.8%, the highest among the models compared.
- Anthropic notes that at this level benchmark margins are a less reliable guide to real differences.
- Strong cost-performance at default effort
- On FrontierCode, Opus 5.5 at medium effort beats GPT-6 Astra's top score (53.3%) at about a fifth of the cost per task.
Community Reactions
- Price cut read as a market signal
- One commenter noted Opus 5 tops OpenRouter spend, so the cut says something about the market and Anthropic's profitability.
- Hopes for a better writing style
- One commenter moved to Astra because of Opus 5/Fable's writing style and wants to test this claim.
- Another still reacted negatively to the new style.
- Real-world cost and quota experience
- One user on $20 plans said GPT-6 Astra burns the 5-hour quota in about an hour while Opus 5.5 lasts nearly all day.
- Another built two games in about 45 minutes at $11 to $14 with xhigh effort, with quality similar to Fable 5.1 and Astra.
- Concern about expanded safeguards
- One commenter worried that flagged conversations may be stored longer and reviewed by humans, as with Fable.
- Cheaper alternatives and model fatigue
- One commenter said DeepSeek v4.1 finished a long UI task for about $0.07.
How to Apply
- Run one existing workload on Opus 5.5 and compare cost
- Pick your costliest agent task and run the same input on Opus 5 and Opus 5.5 today.
- Record token usage, time and output quality side by side.
- Drop effort from max to medium
- Set your coding agent to the default effort (medium) and check quality on your internal eval set.
- Keep it as the default if quality holds.
- Recompute your budget from cache read share
- Pull the share of cache-read tokens from your usage logs.
- Apply the $0.50 to $0.20 cache read price and recalculate monthly cost.
- Log safeguard fallbacks
- If you do cybersecurity or biology work, check responses for requests handled by a fallback model.
- Flag requests whose results differ.
Try It
# Compare effort levels with Simon Willison's llm CLI (run in terminal)
uv tool install llm
llm install llm-anthropic --upgrade
llm keys set anthropic
for e in low medium high; do
llm -m claude-opus-5.5 -o thinking_effort $e "{{your task prompt}}" > out_$e.md
done
# Save usage logs
llm logs -cu > logs-with-usage.md
# Next: compare quality of out_low.md and out_high.mdTerminology
Related Papers
Mercury 3: Diffusion at scale
Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality
Timestral: Mistral's time-series foundation models
A 209M time-series model ranks first on three benchmarks and runs on CPU.
DeepSeek V4.1 Flash 1017 (Max)
DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.
If you don't run your own LLM proxy, how do you know whats going on?
Running your own LLM proxy shows which keys, models, and cache misses drive your usage.
DeepSeek: DeepSeek v4.1 Flash 1017
DeepSeek V4.1 Flash 1017: $0.05/M input, 1M context, built for agents
DeepSeek-v4.1-Flash-1017
DeepSeek V4.1 Flash minor update lifts coding and agent benchmarks while cutting run time and output length
Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware