What Are the Latest Gemini Models? 2026 Google Gemini Lineup, Ranking, and Differences

Q: What is the latest Gemini model in August 2026?

By ship date, the newest public workhorse is Gemini 3.7 Flash (13 Aug 2026). By reasoning ceiling, the flagship is still Gemini 3.1 Pro (Deep Think optional). There is no shipped 3.7 Pro.

Q: Is there a Gemini 3.5 Pro?

As of writing, DeepMind still says 3.5 Pro coming soon. Do not draw a 3.5 / 3.6 / 3.7 Pro on the architecture diagram.

As of 19 August 2026, Google Gemini is no longer “one model, two sizes.” The Flash line shipped Gemini 3.7 Flash on 13 August; the Pro line is still Gemini 3.1 Pro from February. DeepMind’s product page still says 3.5 Pro coming soon. So “latest” is two sentences: the newest workhorse is 3.7 Flash; the strongest public reasoning flagship is still 3.1 Pro (with optional Deep Think).

This piece follows the official product pages, the Firebase / Gemini API model tables, and Google’s 13 August blog post on 3.7 Flash. Do not pick a model by version number. 3.7 does not automatically beat 3.1 Pro. What belongs in a routing table is a role, not a digit.

3.7 Flash

Newest workhorse (13 Aug 2026) for coding / agents

3.1 Pro

Current reasoning flagship (since February)

Typical context window (2M on Vertex for enterprise)

What “latest” actually names

In summer 2026 Google split Gemini across several calendars. DeepMind’s homepage currently pins:

Gemini 3.7 Flash: complex agent work and coding at scale. Introduced 13 August, about three weeks after 3.6 Flash (21 July).
Gemini 3.5 Flash-Lite: high throughput when you still want some intelligence.
Gemini 3.1 Pro: hard tasks and turning ideas into deliverables; the sidenote is still 3.5 Pro coming soon.
Gemini 3.1 Deep Think: harder science, research, and engineering problems with extended reasoning.

On the consumer side, Gemini Spark (Google AI Pro / Ultra, 160+ countries) moved to 3.7 Flash on 13 August. The free Gemini app still often defaults to 3.6 Flash. Do not treat the name in the app and the API id gemini-3.7-flash as the same SLA.

Three lines before you read the tables

There is no public Gemini 3.5 Pro / 3.6 Pro / 3.7 Pro — Pro is still 3.1
3.7 Flash is a workhorse upgrade, not a new flagship that replaces 3.1 Pro
The image line is Gemini Image (nickname Nano Banana); it is not the same endpoint as text Flash

A ranking by job, not an IQ leaderboard

A single IQ ranking will send you to the wrong model. Rank by job:

Hardest reasoning / long documents / Deep Think: Gemini 3.1 Pro (Deep Think when needed). That is the public flagship, not 3.7.
Production agents, coding, knowledge-work throughput: Gemini 3.7 Flash. Google highlights gains vs 3.6 on coding, web generation, PDFs, and workflow automation.
Near-Pro intelligence at Flash price: Gemini 3.5 Flash (stable; Firebase says shutdown no earlier than 2027-05-19).
Free-app default / conservative: Gemini 3.6 Flash.
Mass classify / extract / sub-agents: Gemini 3.5 Flash-Lite.
Legacy: Gemini 2.5 Pro / Flash / Flash-Lite — Firebase lists shutdown as early as 2026-10-16. Do not pin new work to them.

Official 3.7 vs 3.6 numbers (inside Flash, not vs 3.1 Pro): FrontierCode 1.1 Main 43.6% vs 34.4%; DeepSWE v1.1 65.3% vs ~49%; GDP.pdf 34.0% vs 22.0%; AutomationBench 30.4% vs 17.0%; WebDev Arena Elo 1588 vs 1538. That means Flash got stronger and cheaper. It is not a claim that Flash now beats Claude or GPT flagships.

Why Flash and Pro no longer ship together

Late 2025 opened the Gemini 3 generation: 3 Pro preview, Deep Think, 3 Flash. In 2026, Flash shipped 3.5 → 3.6 → 3.7 while Pro stayed on 3.1. Product translation:

Flash owns latency, price, agent loop count, and tool-call discipline.
Pro owns “this problem is worth extra reasoning” — long context, hard multimodal work, Deep Think.
Do not route everything to latest. Default traffic on 3.7 Flash; escalate to 3.1 Pro on failure or low scores. That is usually an order of magnitude cheaper than Pro everywhere.

3.7 Flash introductory price (through 2026-12-31): $0.75 / 1M input and $3.75 / 1M output, about half of the original 3.6 list price; from 2027-01-01 it returns to $1.50 / $7.50. Context: ~1M in, 64k out. Tools: function calling, Search as a tool, Computer use. Surfaces: Gemini API / AI Studio, Android Studio, Antigravity, Gemini Enterprise Agent Platform.

Code editor screen for routing Gemini 3.7 Flash versus 3.1 Pro coding agents — Use 3.7 Flash as the default coding loop and keep 3.1 Pro for hard cases. That matches real bills better than pinning the whole repo to “latest Gemini.”

The rest of the family: Lite, image, Deep Think, niches

Flash-Lite: volume, speed, cost

3.5 Flash-Lite is the current Lite workhorse: receipt extraction, translation, large fan-out. DeepMind demos show lower latency than 3.5 Flash, and it can draft many designs while 3.6 Flash acts as the master agent. Older 3.1 Flash-Lite still appears in some consoles; move new traffic to 3.5 Lite unless a frozen eval set says otherwise.

Nano Banana / Gemini Image

Images are not a hidden parameter on the text Flash models. Common tiers:

Gemini 3.1 Flash Image (Nano Banana 2): conversational edits, multi-image fusion, character consistency.
Gemini 3 Pro Image: higher fidelity, readable text, up to ~14 reference images.
Gemini 3.1 Flash-Lite Image (Nano Banana 2 Lite): high-volume, latency-sensitive asset pipelines.

Google’s 3D-game demo used Flash to orchestrate and Nano Banana to paint characters and textures. Split those into two calls. Do not expect gemini-3.7-flash to carry generative vision alone.

Deep Think, Omni, Audio, Robotics, Embedding

3.1 Deep Think: extended reasoning on Pro for science / math / long-horizon agents. Costlier and slower; not a chat default.
Gemini Omni: video and “make anything from anything.” 3.7 demos orchestrated Omni for parallax landing pages.
Gemini Audio: realtime speech.
Gemini Robotics / Robotics ER 2: vision–language–action. Do not mix into ordinary chat APIs.
Gemini Embedding 2: multimodal vectors for retrieval / RAG. Do not embed with a generator.
3.5 Flash Cyber: find-and-fix vulnerabilities. Evaluate as a security product, not a general coder.

Cheat sheet: pick this, not that

Model	Role	Use for	Do not use for
3.1 Pro	Flagship reasoning	Hard cases, long docs, Deep Think	High-QPS chat, Pro on every step
3.7 Flash	Newest workhorse	Coding agents, workflows, most production traffic	Pretending a 3.7 Pro already exists
3.6 Flash	Free-app default	Consumer consistency, transition baselines	New production default unless evals win
3.5 Flash	Stable near-Pro	Contracts that need a long support window	Ignoring the 2027 shutdown clause
3.5 Flash-Lite	Cheap and fast	Extract, classify, parallel sub-agents	Single-shot ultra-hard reasoning
2.5 family	Legacy	Buffer for old integrations	New pins after October 2026
Image / Nano Banana	Generate and edit pixels	Assets, UI refs, character lock	Replacing a text model for architecture

API IDs, price windows, shutdown risk

Put stable IDs in config: gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash, gemini-3.1-pro, gemini-2.5-pro. Newer 3.x Flash builds are marked short-term availability (shutdown often TBD). 3.5 Flash is no earlier than 2027-05-19. 2.5 may go as early as 2026-10-16. Preview IDs (historically gemini-3-flash-preview) do not belong in production aliases.

Prices: only the 3.6 / 3.7 Flash intro window is discounted through end of 2026; New Year’s Day 2027 removes it. Re-run TCO at the post-intro rates before you sign a contract.

How developers should write the routing table

Default to 3.7 Flash. Coding, tool loops, docs-to-pages, most Workspace agents start here.
Escalate hard cases to 3.1 Pro / Deep Think. Trigger on eval scores or human sampling, not every user sentence.
Sidecar bulk work on 3.5 Flash-Lite. Fields, receipts, 20 drafts in parallel.
Pixels on Image models. Text models only write prompts and orchestrate.
Leave 2.5 before October. Keep a month of paired logs. Do not wait for the shutdown notice.
Isolate the eval machine. Model swaps blow up local GPU/RAM folklore. A cloud Mac is a better place to run the same harness in parallel. When you need a dedicated compile and agent loop, split evals off the laptop with Cloud Mac.

Three questions before you choose

Is this one climb-to-the-summit reason, or ten thousand cost-controlled loops?
Is the failure a wrong sentence, or a bad merge? Only the latter is worth Pro
Is the output code/JSON or pixels? Pixels go to Nano Banana

FAQ

What is the latest Gemini model in August 2026?

By ship date, the newest public workhorse is Gemini 3.7 Flash (13 Aug 2026). By reasoning ceiling, the flagship is still Gemini 3.1 Pro (Deep Think optional). There is no shipped 3.7 Pro.

Is there a Gemini 3.5 Pro?

As of writing, DeepMind still says 3.5 Pro coming soon. Do not draw a 3.5 / 3.6 / 3.7 Pro on the architecture diagram.

Is the free Gemini app on 3.7?

Not necessarily. The free tier often still shows 3.6 Flash; 3.7 is prioritized for paid Spark, the API, and enterprise agent platforms. Trust the in-product picker.

Is Nano Banana the same as Gemini 3.7?

No. Nano Banana is the Gemini Image family nickname. It is billed and called separately from the 3.7 Flash text workhorse. In 3.7 demos they are orchestrated together.

Can I still use 2.5?

Yes, until the documented shutdown window. Firebase lists 2.5 Pro/Flash/Flash-Lite as early as 16 Oct 2026. New projects should use 3.x.

Wrap-up

Three sentences: the newest workhorse is 3.7 Flash; the strongest public reasoner is still 3.1 Pro; Flash and Pro no longer share a version number, so the number is not a ranking. Write roles first (default loop / hard-case upgrade / extract sidecar / images), then fill in IDs. When 3.5 Pro actually ships, you only change the escalate row.

If you care about Apple’s Siri training partnership with Gemini, see the 18 August fall-event software preview. That is AFM distillation, not the Gemini app on iPhone.