Developer
Google

Gemini 3.5

Google's Gemini 3.5 generation. Gemini 3.6 Flash, shipped July 21, 2026, is the flagship: 1M token context, a March 2026 cutoff, $1.50/$7.50 per 1M.

Updated

Released
May 19, 2026
Type
Multimodal Language Model
Context window
1M tokens
Pricing
$1.50 / $7.50 per Mtok
Knowledge cutoff
March 2026
License
Proprietary
🌐 Official Website📚 Documentation💰 Pricing & API
On this page

Overview

Gemini 3.5 is Google's model generation announced on May 19, 2026 at Google I/O. Its current flagship is Gemini 3.6 Flash, a point release that went generally available on July 21, 2026, replacing Gemini 3.5 Flash as the default Flash model.

Google frames Flash as its "most intelligent model for agentic and coding tasks," a positioning that inverts the usual Flash-is-the-cheap-one story: it is the frontier model of this line in practice, not a distillation of one. Gemini 3.6 Flash sharpens that pitch in the one direction that is hard to argue with — it scores higher than its predecessor and costs less. Output fell from $9.00 to $7.50 per million tokens while input held at $1.50, and Google reports the model spending 17% fewer output tokens on equivalent work (measured across the Artificial Analysis benchmark suite at the time of publication, with up to 65% observed on Datacurve's DeepSWE).

The release also closed the generation's most obvious weakness. Gemini 3.5 Flash launched with a January 2025 knowledge cutoff — already 16 months stale on its release day. Gemini 3.6 Flash moves it to March 2026.

Gemini 3.5 Pro still has not shipped. Announced at I/O with a June rollout, it was described on July 21, 2026 as "currently testing with partners." It has no API model ID and no pricing entry. Treat it as forthcoming, not available.

The prior generation's flagship, Gemini 3.1 Pro, was released on February 19, 2026 and remains available — still carrying a Preview label, and still the more expensive option per token despite scoring below 3.6 Flash on every benchmark Google publishes for both.

Capabilities

  • Agentic loops: Google tunes Flash for "agentic loops involving complex coding cycles." The 3.6 improvement is framed as precision rather than raw power — fewer unwanted code edits, fewer execution loops, fewer tool calls per task.
  • Native multimodal input: Text, image, video, audio, and PDF in a single request. Output is text.
  • Thinking: Internal reasoning before answering, with structured outputs supported alongside it.
  • Computer use: Now a built-in client-side tool in the Gemini API and Gemini Enterprise rather than a separate model path, though the feature itself remains preview-labelled.
  • Grounding: Google Search grounding and Google Maps grounding, plus URL context and file search.
  • Code execution: Runs code as a tool during a response.
  • Long-context retrieval: The 1M-token window is now backed by a usable retrieval score — 54.0% on GDM-MRCR v2 at 1M tokens, against under 27% for both 3.5 Flash and 3.1 Pro.

Technical Specifications

  • Model ID: gemini-3.6-flash (stable). Predecessor gemini-3.5-flash remains callable.
  • Alias: gemini-flash-latest tracks the newest Flash release and is hot-swapped on each launch — pin an explicit ID in production.
  • Input limit: 1,048,576 tokens
  • Output limit: 65,536 tokens
  • Knowledge cutoff: March 2026
  • Input modalities: Text, image, video, audio, PDF
  • Output modalities: Text
  • Supported: Thinking, structured outputs, function calling, Search grounding, Google Maps grounding, code execution, Batch API, flex inference, priority inference, caching, file search, URL context, computer use (preview)
  • Not supported: Audio generation, image generation, Live API

Model Family

Gemini 3.6 Flash — generally available

Released July 21, 2026. 1,048,576 input / 65,536 output tokens. $1.50 / $7.50 per million. March 2026 cutoff. The default choice for agentic and coding work on the Gemini API.

Gemini 3.5 Flash — superseded, still callable

Released May 19, 2026. Same context limits, January 2025 cutoff, $1.50 / $9.00 per million. Scores below 3.6 Flash on every benchmark Google publishes for both, at a higher output price — there is no reason to start new work on it.

Gemini 3.5 Flash-Lite — generally available

Released July 21, 2026. gemini-3.5-flash-lite. Roughly 350 output tokens per second at $0.30 in / $2.50 out per million, aimed at document processing, high-volume automation, and subagent roles inside larger agent systems. It scores 54.2% on SWE-Bench Pro against 49.6% for Gemini 3 Flash, a full-size model two Flash releases back. Context caching is not available on it.

Gemini 3.5 Flash Cyber — restricted pilot

A security-specialised variant that finds and patches software vulnerabilities. It is not on the public API: access runs through CodeMender, in a limited-access pilot for governments and trusted partners. Google gates it explicitly on misuse risk.

Gemini 3.5 Pro — announced, not shipped

Announced May 19, 2026. No API model ID, no pricing entry, no changelog entry as of July 21, 2026. Google's latest statement is that it is testing with partners and will launch when ready.

Gemini 3.1 Pro Preview — available

Released February 19, 2026. Still labeled Preview. Tiered pricing: $2.00 in / $12.00 out per million for prompts up to 200K tokens, $4.00 in / $18.00 out above 200K. Context caching at $0.20 (≤200K) or $0.40 (>200K) per million plus $4.50 per hour.

Gemini Omni Flash Preview — public preview

Released to developers June 30, 2026. A high-performance multimodal model for high-speed video generation — 3-10 second clips at 720p from text or animated still images, with conversational editing across turns. Input $1.50 per million; output $9.00 per million (text) or $17.50 per million (video), roughly $0.10 per second at 720p.

Gemini 3.5 Live Translate — preview

A low-latency, real-time speech-to-speech translation model supporting 70+ languages. Listed in Google's model catalog and on the pricing page, still labeled Preview.

Use Cases

  • Agentic coding: Multi-step coding cycles with tool calls, which is the workload Google names first and the one 3.6 Flash was tuned for.
  • Computer-use automation: 83.0% on OSWorld-Verified is the best score in Google's own comparison set, and the tool is now callable directly from the API.
  • Long-document analysis: A 1M token input window covers large codebases, contract sets, and research corpora, with PDFs accepted natively — and, as of 3.6, a retrieval score that makes the full window worth using.
  • Grounded question answering: Search and Maps grounding for questions past the March 2026 cutoff.
  • Multimodal understanding: Video, audio, and image inputs into a text answer — transcription-adjacent work, chart reading, screen understanding.
  • High-throughput batch jobs: Batch API at 50% of interactive pricing, or Flash-Lite at a fifth of the input price when quality demands are lower.
  • Video generation and editing: Via gemini-omni-flash-preview, with conversational refinement instead of full re-prompting.

Performance / Benchmarks

Google-reported results for Gemini 3.6 Flash at its July 21, 2026 announcement, against the two models it succeeds:

BenchmarkGemini 3.6 FlashGemini 3.5 FlashGemini 3.1 Pro
SWE-Bench Pro58.7%55.1%54.2%
DeepSWE v1.149%37%12%
Terminal-Bench 2.178.0%76.2%73.8%
MLE-Bench63.9%49.7%42.6%
GDPval-AA v21421 Elo1349 Elo965 Elo
OSWorld-Verified83.0%78.4%—
GDM-MRCR v2 (1M tokens)54.0%under 27%under 27%

It does not lead the field. On Google's own comparison table, GPT-5.6 Luna leads DeepSWE (67%) and Terminal-Bench 2.1 (84.7%), Grok 4.5 leads SWE-Bench Pro (64.7%), and Claude Sonnet 5 leads MLE-Bench (66.9%) and GDPval-AA v2 (1607 Elo). Gemini 3.6 Flash's argument is price-performance, not the crown: it lands within a few points of models charging several times more per output token. All figures above are vendor-run; independent evaluations may differ.

Limitations

  • Not the top scorer on most published benchmarks — GPT-5.6 Luna, Grok 4.5 and Claude Sonnet 5 each lead it somewhere, per Google's own charts.
  • Text output only. Image and audio generation and the Live API are unsupported; use Omni Flash for video.
  • Gemini 3.5 Pro still does not exist as a callable model, two months after announcement and one missed rollout date.
  • Gemini 3.1 Pro is still Preview — preview models can change or be withdrawn — and is now both slower and more expensive than a Flash model that outscores it.
  • Output cap of 65,536 tokens is one-sixteenth of the 1,048,576-token input window; very long generations must be chunked.
  • Computer use is preview-quality despite the strong OSWorld number, and should not be assumed stable.
  • gemini-flash-latest is hot-swapped on every Flash release, so an unpinned alias silently changes model under a running application.
  • Grounding costs extra beyond 5,000 prompts per month (shared across the Gemini 3 line), at $14 per 1,000 search queries.
  • Flash-Lite drops context caching entirely, which removes the main lever for cheap repeated-prefix workloads.

Pricing & Access

Gemini 3.6 Flash (per 1M tokens)

InputOutput
Standard$1.50$7.50
Batch (50% off)$0.75$3.75

Context caching: $0.15 per million tokens ($0.075 on batch), plus $1.00 per million tokens per hour of storage.

Gemini 3.5 Flash-Lite (per 1M tokens)

InputOutput
Standard$0.30$2.50
Batch (50% off)$0.15$1.25

Context caching is not available.

Gemini 3.1 Pro Preview (per 1M tokens)

Prompt sizeInputOutput
≤ 200K tokens$2.00$12.00
> 200K tokens$4.00$18.00

Batch: $1.00 / $6.00 (≤200K) and $2.00 / $9.00 (>200K). Context caching: $0.20 or $0.40 per million plus $4.50 per hour.

Grounding

5,000 prompts per month free, shared across the Gemini 3 line, then $14 per 1,000 search queries.

Where Gemini 3.6 Flash is available

The Gemini app, AI Mode in Google Search, Google Antigravity, the Gemini API, Google AI Studio, Vertex AI, Android Studio, the Gemini Enterprise Agent Platform, and Gemini Enterprise.

Ecosystem & Tools

Google platforms

  • Google AI Studio - Prototyping and API keys
  • Google Antigravity - Agentic development platform
  • Vertex AI - Enterprise deployment and MLOps
  • Gemini Enterprise Agent Platform - Agent hosting for enterprises
  • Android Studio - In-IDE model access

Open-weight sibling

  • Gemma 4 - Google's open-weight family, Apache 2.0 licensed, for local and on-device deployment

Developer

Community & Resources

Frequently Asked Questions

Gemini 3.6 Flash, released July 21, 2026 and generally available. It replaces Gemini 3.5 Flash, which shipped on May 19, 2026 and remains callable.
1,048,576 input tokens and 65,536 output tokens — unchanged from Gemini 3.5 Flash.
March 2026, up from January 2025 on Gemini 3.5 Flash.
$1.50 per million input tokens and $7.50 per million output tokens. Batch is half that — $0.75 and $3.75. Context caching costs $0.15 per million tokens plus $1.00 per million tokens per hour of storage.
Yes. Input is unchanged at $1.50, but output fell from $9.00 to $7.50. Google also reports the model spending 17% fewer output tokens on the same work, so the effective saving is larger than the price cut alone.
Text, image, video, audio, and PDF as input; text as output. It does not generate images or audio, and does not support the Live API.
Google reports 58.7% on SWE-Bench Pro, 78.0% on Terminal-Bench 2.1, 83.0% on OSWorld-Verified, 63.9% on MLE-Bench, and 54.0% on GDM-MRCR v2 at 1M tokens. All are vendor-run figures.
Not as of July 21, 2026. Google said on that date that Pro is "currently testing with partners" and will launch when ready — a second slip past the June 2026 rollout described at announcement.
A high-throughput, low-latency model released July 21, 2026 at $0.30 in / $2.50 out per million tokens, running at roughly 350 output tokens per second. It does not support context caching.
The Gemini app, AI Mode in Google Search, Google Antigravity, the Gemini API, Google AI Studio, Vertex AI, Android Studio, and Gemini Enterprise.

Related Models

GPT-5.6

OpenAI's flagship family — Sol, Terra and Luna — generally available since July 9, 2026. All three share a 1,050,000 token context and a February 2026 cutoff.

Claude Opus 5

Anthropic's Opus-tier flagship — near-Fable capability at half the price, with adaptive thinking on by default and a 1M token context window.

Claude Sonnet 5

Anthropic's best mix of speed and intelligence — the most agentic Sonnet yet, reaching near-Opus quality on coding and agentic work at Sonnet cost.

Grok 4.5

xAI's flagship model, a mixture-of-experts system trained jointly with Cursor. A 500K token context at $2.00 in / $6.00 out per million tokens.

Gemma 4

Google's open-weight Gemma 4 family — five Apache 2.0 models from 2.3B to a 30.7B dense flagship, spanning 140+ languages and native audio, vision, and text.

Llama 4

Meta's last open-weight frontier release: Scout and Maverick, natively multimodal Mixture-of-Experts models with 17B active and up to a 10M token context.

Explore More Models

Discover other AI models and compare their capabilities.

Browse Models← Back to Models