Gemini 3.8 Flash and 3.8 Flash Cyber
TL;DR Highlight
Gemini 3.8 Flash boosts coding and reasoning at the same price as 3.7
Who Should Read
A backend developer using cheap Flash-class models for coding agents or long-running automation
Core Mechanics
- Same price as 3.7 Flash
- It costs $0.75 per 1M input tokens and $3.75 per 1M output tokens.
- This is an introductory price that expires on December 31, 2026.
- From January 1, 2027, it becomes $1.50 input and $7.50 output.
- The model works harder
- On complex tasks it takes extra reasoning steps and calls tools iteratively.
- So it can use more tokens, especially at higher effort levels.
- Beats larger models on long coding tasks
- On DeepSWE v1.1 (long-horizon software engineering), it outperforms most larger frontier models, per Google.
- Google says it does so at a fraction of the cost.
- Also ahead of 3.7 Flash in professional domains
- It beats 3.7 Flash and other frontier models on Vals Finance Agent V2 and Harvey's Legal Agent Benchmark, per Google.
- It scores 54.9% on HLE-Verified.
Community Reactions
- Commenter got an HTML result in 13 seconds for 1.8 cents
- They used the prompt 'make me a cool thing in html' and were impressed by the speed and HTML/JS skill.
- Opinion that low thinking level regressed
- A commenter's pelican SVG test suggests 3.8 at low effort is worse than 3.7.
- High-effort cost was 8.9742 cents for 3.8 vs 8.4387 cents for 3.7.
- Commenter found 3.7 slower than 3.5
- For large generations, 3.5 Flash took 8-10 seconds and 3.7 Flash took 15-25 seconds.
- Both ran at the lowest reasoning setting, and the commenter noted this release does not mention speed.
- Opinion that low reasoning scores are more telling
- A commenter cited Artificial Analysis index scores: 3.7 at 51/53/57 and 3.8 at 52/57/59.
- They use the lowest reasoning level in production with good results.
- Commenter finds Flash models reckless for coding
- Even with a generous Ultimate quota, they still use Opus 4.6 because it is more reliable when the context window is managed carefully.
How to Apply
- Compare 3.7 and 3.8 on five of your own prompts
- Run five real prompts on both models at the same effort level today.
- Record quality, latency, and output tokens in a table.
- Measure cost per effort level
- Run the same task at low, medium, and high effort.
- Multiply tokens by price to get cost per task.
- Rebudget for the January 2027 price increase
- Compute monthly cost at $1.50 input and $7.50 output.
- Check whether simple tasks can be routed to 3.7 Flash to cut cost.
- Put 3.8 Flash in your agent retry loop
- Attach it to coding tasks that tests can verify.
- Run it with retry on failure and log success rate and cost.
Try It
# Prompt for Claude or ChatGPT: design a 3.7 vs 3.8 comparison experiment (simplified version of the article's idea)
I am deciding whether to switch my LLM API model from 3.7 Flash to 3.8 Flash.
3.8 Flash has the same price as 3.7 ($0.75 input, $3.75 output per 1M tokens) but may use more tokens through extra reasoning steps and tool calls.
My service description: {{service description}}
Five representative tasks: {{task list}}
Using this, produce:
1. An experiment table comparing both models at low/medium/high effort (columns: quality, latency, output tokens, cost per task)
2. Automatic scoring criteria per task
3. Decision criteria for switching to 3.8
Next step: ask it to turn the table into an actual API call script.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