DeepSeek V4 Flash 0731
TL;DR Highlight
DeepSeek V4 Flash 0731 scores 89.0% on ARC-AGI-1 at $0.02 per task at max effort
Who Should Read
Backend developers looking for a cheap model for agents and coding automation to cut LLM API costs
Core Mechanics
- 89.0% on ARC-AGI-1 at max effort
- This is on ARC-AGI-1 Semi-Private at $0.02 per task.
- ARC-AGI-2 Semi-Private is 61.4% at $0.04 per task.
- Score varies a lot by reasoning level
- ARC-AGI-1: Max 89.0%, High 87.0%, Low 84.0%, None 11.8%.
- ARC-AGI-2: Max 61.4%, High 56.0%, Low 46.0%, None 2.1%.
- Turning reasoning off collapses performance
- The None variant gets 11.8% on ARC-AGI-1 and 2.1% on ARC-AGI-2.
- Even Low reaches 84.0% on ARC-AGI-1.
- Token pricing is very low
- Input is $0.13 and output is $0.26 per 1M tokens.
- There are 4 reasoning variants, released Jul 31, 2026.
- Per-task pass/fail tables are published
- They cover ARC-AGI-2 Public Eval (120 tasks) and ARC-AGI-1 Public Eval (400 tasks) per variant.
Community Reactions
- Cost stopped mattering in daily use
- One commenter running 5-6 sessions in Oh My Pi struggles to spend $5 a day.
- They suggest CI auto-fix proposals, auto-generated tests, and log monitoring become feasible.
- Local inference experience
- One commenter saw about 250 tok/s single stream on 2x RTX Pro 6000 Blackwell and about 1000 tok/s with 64 streams on vLLM.
- They attribute the speed to a ~300B model with 13B active.
- Pushback on agentic quality
- One user reported more infinite loops and tool-less self-talk than the prior Flash in Pi agent.
- Another's own benchmark found more errors than Gemini 3.5 Flash Lite and GPT 5.6 Luna.
- Price hike warning and caching tips
- A commenter noted DeepSeek announced a significant API price increase.
- On Fireworks AI cache hits cost 20%, and one user found GitHub Copilot poor at using the cache compared to Cline.
How to Apply
- Feed a CI failure log to the cheap model
- Paste one failing CI log into DeepSeek V4 Flash 0731 today and ask for a fix.
- Start with use cases where a wrong answer can be discarded.
- Compare reasoning levels starting at Low
- Run 20 of your own tasks at None, Low and Max, recording accuracy and cost.
- Skip None and choose between Low and Max.
- Fix your prompt prefix to hit the cache
- Keep the system prompt and shared context at the front.
- Check your provider's cache pricing and whether your harness supports it.
- Make the model swappable before the price hike
- Move the model name to an env var so you can change it in one line.
- Recalculate unit costs after the pricing notice.
Try It
# Paste into Claude or ChatGPT (simple practical prompt for comparing before a model swap)
You are an LLM model-swap evaluator. For the failure log {{CI failure log}}, propose a fix in 3 lines or fewer.
After answering, summarize in a table: fix confidence (high/medium/low), extra info needed, and whether it is safe to auto-apply.
Next step: run the same log through DeepSeek V4 Flash 0731 at Low and Max and compare results and cost.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