Gemini 3.7 Flash vs. 3.6: Price, Benchmarks, and Whether to Switch

What shipped on 13 August 2026

Gemini 3.7 Flash is Google’s generally available workhorse for coding and agents, launched on 13 August 2026 — about three weeks after 3.6 Flash.

📑Table of Contents

1. What Gemini 3.7 Flash is as a coding-agent workhorse
2. Benchmarks and intro pricing versus 3.6 Flash
3. How to reach it on Gemini API, Copilot, and Antigravity
4. Decision criteria before you switch a workhorse
5. FAQ
6. Summary

The original 3.6 Flash unit price cited by the launch blog is $1.50 input and $7.50 output per 1M tokens. Introductory Gemini API pricing for 3.7 Flash is half of that pair: $0.75 / $3.75 through 31 December 2026. From 1 January 2027 the 3.7 list returns to the same original $1.50 / $7.50.

---

What to decide before you switch

If a coding agent already runs on 3.6 Flash or another Flash-class model, the useful question is not “what did the launch post say?” It is the following three points.

• Which surface can actually serve the model
• Which published scores map to your workload
• Whether the 2027 list price still makes sense

Pick one surface among Gemini API, Antigravity, or Copilot, run the same 3.6 Flash tasks in isolation, and decide whether 3.7 Flash still earns its keep at the 2027-01-01 list price.

---

What you will use this comparison for

---

What Gemini 3.7 Flash is as a coding-agent workhorse

Official positioning

Google calls Gemini 3.7 Flash its most intelligent workhorse yet for coding and agents. The primary announcement is Introducing Gemini 3.7 Flash (13 August 2026).

The DeepMind Flash model page marks it generally available and lists these targets.

• Everyday tasks
• Agentic coding
• Advanced reasoning
• Multimodal input
• Knowledge work

---

Model id and token limits

The stable model id is gemini-3.7-flash. Gemini API model docs specify the following.

• Input: Text / Image / Video / Audio / PDF
• Output: text
• Input limit: 1,048,576 tokens
• Output limit: 65,536 tokens

DeepMind rounds that figure to 1M / 64k. Use the exact API integers as your configuration source of truth.

---

Supported and unsupported capabilities

This model supports the following capabilities.

• Caching
• Code execution
• Computer use (Preview)
• File search
• Function calling
• Maps grounding
• Search grounding
• Structured outputs
• Thinking (low / medium / high)
• URL context

The minimal thinking setting is unsupported and returns an error.

This model does not support the following capabilities.

• Audio generation
• Image generation
• Live API

If you require those surfaces, this model cannot serve as your only workhorse.

---

Availability surfaces

Official availability includes the following.

• Gemini API
• Google AI Studio
• Android Studio
• Google Antigravity
• Gemini Enterprise Agent Platform
• The Gemini Enterprise app
• Gemini Spark (Google AI Pro / Ultra, supported countries)

---

Consumer chat versus developer paths

The consumer chatbot path is limited to Spark. Ars Technica independently records that the regular Gemini chat UI stayed on 3.6 Flash.

Developer traffic should assume the API, Studio, Antigravity, or Enterprise tier — not the default chat box. For product-level Antigravity pricing versus Cursor, see the existing ⟦DEVGENT_IMMUTABLE_0⟧.

---

Benchmarks and intro pricing versus 3.6 Flash

How to read intro pricing

Pricing authority sits in the Keyword blog footnote and the Gemini API changelog for 13 August 2026.

• Intro rates: $0.75 / $3.75 through 31 December 2026
• Standard rates: $1.50 / $7.50 from 1 January 2027

The blog describes this as half the original 3.6 Flash cost. DeepMind’s footnote applies the same intro pair to both 3.6 and 3.7 Flash.

A matching current list is not the same claim as “half of what 3.6 cost at launch.”

---

Independent list-price records

Independent operators often repeat the introductory list.

• Artificial Analysis: 13 August 2026 release, Intelligence Index 56 (17 of 188 in class), $0.75 / $3.75 list, a 90% cache discount, and $0.40 per Intelligence Index task
• BenchLM (data as of 15 August 2026): 61.4/100 overall (50 of 218), coding rank 12 of 135, and agentic rank 25 of 130
• OpenRouter: Google AI Studio at $0.75 / $3.75 and a Vertex 50% intro at $0.375 / $1.875

Cache hits can reduce the effective rate below the list price.

---

Score table versus 3.6 Flash

Item	Gemini 3.7 Flash	Gemini 3.6 Flash	Notes
Intro input/output $/1M	0.75 / 3.75	Blog calls 3.7 half of original 3.6	Ends 2026-12-31
Post-intro $/1M	1.50 / 7.50	Same footnote	From 2027-01-01
FrontierCode 1.1 Main	43.6%	34.4%	Production-style code quality
DeepSWE v1.1	65.3%	49.0% / 48.6%	Blog vs DeepMind table
WebDev Arena Elo	1588	1538	UI / web
AutomationBench	30.4%	17.0%	Business workflows
GDP.pdf	34.0%	22.0%	Complex PDFs
AA Intelligence Index	56	52	DeepMind table

---

Sources for the score table

• Google Keyword blog
• DeepMind Flash page
• Artificial Analysis (as of August 2026)

---

Peer scores on the official table

DeepMind’s Performance table also ranks competing models.

• FrontierCode: Gemini 3.7 Flash at 43.6%, Claude Sonnet 5 at 42.7%, and GPT-5.6 Terra at 41.3%
• DeepSWE: GPT-5.6 Terra at 69.6% versus 65.3%
• Terminal-bench 2.1: 85.8%, up from 78.0% on 3.6 Flash
• Terminal-bench 3.0: 14.9%

Short-horizon gains do not automatically transfer to long-horizon agents.

---

Knowledge-work eval (Box)

Outside Google’s own suite, Box’s Complex Work Eval records the following.

• Overall: Gemini 3.7 Flash at 67% versus 62% for 3.6 Flash
• Financial-services tasks: 83% versus 65%
• Mean completion: around 73 seconds versus 115 seconds
• Availability in Box: soon

Treat vendor rubrics as a reason to run an isolated bake-off, not as a substitute for your own repository. A broader model-family frame appears in the existing ⟦DEVGENT_IMMUTABLE_0⟧.

---

How to reach it on Gemini API, Copilot, and Antigravity

Gemini API / AI Studio

The shortest developer path is to select gemini-3.7-flash in Google AI Studio or pass that model id to the Gemini API. The model docs do not carry the price table, so pair them with the changelog deadline and the blog footnote.

1. Select gemini-3.7-flash in AI Studio or the Gemini API
2. Use thinking low / medium / high only; do not send minimal
3. Enable computer use only if you accept Preview status; do not expect Live API or image generation
4. Model both the intro rate through 31 December 2026 and the standard rate from 1 January 2027

---

Antigravity

Antigravity appears on the official developer list and on DeepMind’s availability line. Managed Agents and the earlier 3.6 Flash default are covered in Gemini API Managed Agents. The 3.7 GA is a default-model swap on that path, not a redesign of the product surface.

---

GitHub Copilot

GitHub’s 13 August 2026 changelog says Gemini 3.7 Flash is rolling out in GitHub Copilot for Pro, Pro+, Max, Business, and Enterprise.

Available surfaces include the following.

• Visual Studio Code
• Visual Studio
• Copilot CLI
• The cloud agent
• The Copilot app
• JetBrains
• Xcode
• Eclipse

Business and Enterprise users cannot pick it until an admin enables the Gemini 3.7 Flash Preview policy. Billing follows provider list pricing under usage-based billing.

GitHub’s early-test claims about web, app, and agentic coding are Copilot-side statements, not a replacement for your own evaluation set.

---

Constraints to keep separate

Keep the constraints clearly separate.

• Regular Gemini chat remaining on 3.6 Flash is an independent report, not a developer blocker
• Spark is limited to AI Pro / Ultra in supported countries
• Official note: the EEA, Nigeria, Switzerland, and the UK are outside Spark
• On OpenRouter, google/gemini-3.7-flash shows both the Studio list and the Vertex 50% intro; pin the billing backend before you compare runs

---

Decision criteria before you switch a workhorse

Comparison axes

The goal is not a ranking. It is whether Gemini 3.7 Flash should become the default coding-agent workhorse or stay on a discounted evaluation lane.

Axis	Reason to try 3.7 Flash now	Reason to wait
Unit cost	$0.75 / $3.75 through year-end; Vertex / OpenRouter may discount further	Price doubles on 2027-01-01; long contracts should be modeled at standard rates
Coding	FrontierCode 43.6% and DeepSWE 65.3%; Copilot claims web/app and agentic gains	GPT-5.6 Terra still leads DeepSWE at 69.6% on Google’s own table

---

Agents, surfaces, and reproducibility

Axis	Reason to try 3.7 Flash now	Reason to wait
Agents	AutomationBench 30.4%; Terminal-bench 2.1 at 85.8%	Terminal-bench 3.0 is 14.9%; long-horizon work needs a separate eval
Surfaces	API, Studio, Antigravity, Copilot, Enterprise	Regular Gemini chat has not switched; computer use is Preview
Reproducibility	Artificial Analysis and BenchLM corroborate list price and launch date	Most scores are provider-reported; do not adopt without your own tasks

---

Sources for the decision table

• DeepMind Performance table
• Ars Technica
• GitHub Changelog
• Artificial Analysis (as of August 2026)

---

Pre-switch checklist

Pre-switch checklist:

• ✓ You picked one surface (API / Antigravity / Copilot / Spark)
• ✓ For a Copilot org, an admin enabled the Preview policy
• ✓ Thinking is not locked to high before you measure cost versus quality
• ✓ You modeled both intro pricing and the 2027 standard rate
• ✓ You compared 3.6 and 3.7 on at least 10 identical prompts in your repo
• ✓ You kept another model if Live API or image generation is required
• ✓ Production keys and customer data stayed off the first trial

---

If the call is still unclear

If the call is still unclear, run 3.6 and 3.7 in parallel on an isolated project during the intro window and watch failure rate, retries, and cache-hit rate only. A vendor-table gap is a weak reason to change a production default before the 2027 price reset.

---

FAQ

Q1: When did Gemini 3.7 Flash ship?

The official blog and the Gemini API changelog both date the launch to 13 August 2026. The changelog marks it generally available.

Q2: Is the price really half, and when does that end?

Intro pricing is $0.75 / $3.75 through 31 December 2026. From 1 January 2027 it becomes $1.50 / $7.50. The blog compares that intro pair with the original 3.6 Flash price. Vertex and OpenRouter may advertise additional discounts; check the billing backend’s list.

Q3: Can I use it in the Gemini app today?

Google limits the individual chat path to Spark (AI Pro / Ultra, supported countries). Ars Technica records that regular chat stayed on 3.6 Flash. Developers should use the API, Studio, Enterprise, or Antigravity.

Q4: How does GitHub Copilot expose it?

GitHub’s 13 August 2026 changelog describes a staged rollout. Business and Enterprise tenants will not see it in the model picker until an admin enables the Gemini 3.7 Flash Preview policy. Billing is usage-based at provider list prices.

Q5: Should every 3.6 Flash workload move up?

Official tables improve on coding, PDFs, and automation. The 3.6 DeepSWE baseline is 49.0% in the blog and 48.6% in the DeepMind table. Do not switch without your own tasks and a 2027 unit-cost model.

Q6: What are the context and thinking limits?

Input is about 1M tokens and output is 65,536 tokens. Thinking accepts low, medium, or high. minimal errors. Live API and image generation are unsupported.

---

Related articles:

• 934
• 11300
• 11301

Summary

Gemini 3.7 Flash is the 13 August 2026 coding-agent workhorse GA. The operational story is the intro discount — $0.75 / $3.75 versus original 3.6 Flash $1.50 / $7.50 — plus published coding and agent gains, not a claim that every Gemini surface already moved.

The difference that matters is which lane you put it on — Gemini API, Antigravity, or Copilot — and whether you spend the window through 31 December 2026 measuring that choice.

Pick one surface among Gemini API, Antigravity, or Copilot, run the same 3.6 Flash tasks in isolation, and decide whether 3.7 Flash still earns its keep at the 2027-01-01 list price.