Claude Opus 5
TL;DR Highlight
Opus 5 delivers major gains over Opus 4.8 at the same price, nearing Fable 5 at half the cost.
Who Should Read
A backend developer running coding agents or business automation on the Claude API who is weighing a model switch and cost optimization
Core Mechanics
- Big jump at Opus 4.8 pricing
- Priced at $5 per million input tokens and $25 per million output tokens, same as Opus 4.8.
- The API model name is claude-opus-5.
- Near Fable 5 at half the cost
- On CursorBench 3.2 at max effort, it scores within 0.5% of Fable 5's peak at half the cost per task.
- On OSWorld 2.0 it beats Fable 5's best result at just over a third of the cost.
- More than double Opus 4.8 on Frontier-Bench
- On Frontier-Bench v0.1 it beat all other models and more than doubled Opus 4.8's performance.
- Cost per task is lower than Opus 4.8.
- Stronger self-verification and iteration
- On a Frontier-Bench task with no way to view the drawing, it wrote its own computer vision pipeline to rebuild the 3D part.
- No competing model solved it in five attempts with the same setup.
- Effort setting trades intelligence for tokens
Community Reactions
- Fable-class model without data retention requirement
- One commenter noted Fable has a 30-day data retention requirement, while Opus 5 has none for general access.
- They argued this matters most for organizations.
- Disappointing on C/C++ code review
- A commenter gave Opus 5 a large C/C++ review that GPT-5.6-sol had already cleaned up and got 4 errors.
- GPT judged all 4 false, and the commenter felt Opus misses wider context.
- OSWorld 2.0 numbers differ from the paper
- A commenter pointed out Anthropic reports 55.7% for Opus 4.8 while the benchmark paper reports about 21%.
- An edit explained the gap as full completion versus partial credit.
- Better image-to-HTML than Fable
- A commenter testing image-to-HTML found Opus 5 followed the design source more closely, down to button shapes.
- They said Fable was previously best, with Gemini 3.1 pro second.
How to Apply
- Run your existing prompts on claude-opus-5 and compare
- Today, swap the model name and run 5-10 representative prompts side by side with Opus 4.8.
- Record quality, output tokens and response length.
- Measure the cost curve by lowering effort
- Run the same task set at effort levels from low to max in your agent or pipeline.
- Set the default at the best pass rate per cost.
- Try mid-conversation tool changes for agents with shifting tools
- Enable the beta in agents that change their tool list mid-conversation.
- Verify cache hit tokens to confirm the prompt cache survives.
- Decide your fallback setup for security requests
- If you serve security code analysis, first measure how many requests the classifiers block.
- Then decide whether to enable automatic fallbacks or join the CVP.
Try It
# Paste into Claude Code or Claude (a quick prompt to compare Opus 5 behavior)
You are a senior developer. Read the bug report below and find the root cause, not just the surface symptom.
{{bug report or code}}
Steps:
1. Form a root-cause hypothesis and state how you will verify it.
2. Write or run a test to confirm the hypothesis.
3. After the fix, re-verify edge cases and report the results.
Next: run the same prompt on Opus 4.8, vary the effort, and record token usage and results.Terminology
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