Gemini 3.7 Flash

Who Should Read

A backend developer running Gemini Flash in coding or agent pipelines who is weighing a model switch and cost comparison.

Core Mechanics

• Follow-up to 3.6 Flash after three weeks
  • Google calls 3.7 Flash its most intelligent workhorse model for coding and agents.
  • It says the release comes from developer feedback and algorithmic improvements.
• Coding benchmarks jumped
  • FrontierCode 1.1 Main is 43.6% for 3.7 Flash versus 34.4% for 3.6 Flash.
  • DeepSWE v1.1 is 65.3% versus 49.0% for 3.6 Flash.
• Web dev and document tasks improved too
  • WebDev Arena Elo is 1588 versus 1538 for 3.6 Flash.
  • GDP.pdf rose from 22.0% to 34.0% and AutomationBench from 17.0% to 30.4%.
• Intro price is half the original 3.6 price
  • Through Dec 31, 2026: $0.75/1M input and $3.75/1M output tokens.
  • From Jan 1, 2027: $1.50/1M input and $7.50/1M output.
• Google says agent behavior changed
  • It claims the model adapts to roadblocks, asks for clarification, and follows instructions more faithfully.

Community Reactions

• Commenters say competitors are cheaper
  • Some say GPT-5.6 Luna at $0.2/1M input and $1.2/1M output undercuts Flash.
  • Another says DeepSeek V4 Flash/Pro is 13-26x cheaper with comparable text intelligence.
• Speed remains the strength, per commenters
  • One user finds Gemini much faster than waiting on Opus 5.
  • Another cites Artificial Analysis: score up from 52 to 56, but output tokens per task rose from 26k to 37k.
• One tester found 3.6 better in practice
  • A screenshot-plus-prompt chat UI build from 3.7 had broken category filtering and messages.
  • The same test with 3.6 mostly worked, they said.
• Hallucination concerns persist, commenters say
  • One user saw invented data despite giving tool and API access.
  • Another says 3.6 Flash claims to have completed tasks it did not.
• Commenters find the pricing odd

How to Apply

• Compare 3.6 and 3.7 on your own prompts today
  • Run 5-10 real coding or tool-calling prompts through both models under identical conditions.
  • Record correctness, tool call count, and output tokens.
• Calculate cost per task, not per token
  • More output tokens can erase the half-price benefit, so multiply total tokens per task by the rates.
  • Also compute with the January 2027 price ($1.50/$7.50).
• Benchmark against rivals like Luna on one eval set
  • Run candidate models on the same set and tabulate quality, latency, and cost.
  • Include latency criteria for agent steps where speed matters.
• Add a hallucination check to your agent
  • Verify tool-derived data against the source before the final answer.
  • Confirm any claimed completion against the real output.

Try It

# Paste into Claude or ChatGPT to plan a 3.6 vs 3.7 Flash comparison (simple version of the article's benchmark comparison)

Our service is {{service description}}, and we currently use Gemini {{current model}} for {{use: coding/agents/document processing}}.
Monthly input tokens: {{input tokens}}, output tokens: {{output tokens}}.

Please:
1. Calculate monthly cost at the Gemini 3.7 Flash intro price ($0.75/1M input, $3.75/1M output) and the post-Jan 2027 price ($1.50/1M input, $7.50/1M output).
2. Also calculate cost assuming output tokens grow by {{increase}}%.
3. Propose 10 test cases and scoring criteria to compare 3.6 and 3.7 on my service.

Next step: run the test cases on both models and paste the results table back for analysis.

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

Related Resources

• https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/
• https://artificialanalysis.ai/models/gemini-3-7-flash
• https://ai.google.dev/gemini-api/docs/latest-model
• https://deepswe.datacurve.ai
• https://developers.openai.com/api/docs/models/gpt-5.6-luna
• https://cognition.com/frontiercode