Gemini 3.6 Flash

Gemini 3.6 Flash is a proprietary, multimodal large language model released by Google on July 21, 2026. It belongs to the Gemini 3 series and uses the stable API identifier gemini-3.6-flash. Google released it as a generally available model for coding, knowledge work, visual reasoning, and multi-step agent workflows, with lower standard output-token pricing than Gemini 3.5 Flash.^[1][2]

The model accepts text, images, video, audio, and PDF documents and returns text. Its input limit is 1,048,576 tokens and its maximum output is 65,536 tokens. It supports internal reasoning, tool calls, code execution, structured responses, grounding, and context caching, although the exact feature set differs across Google's developer and enterprise platforms.^[2][3]

Google DeepMind describes Gemini 3.6 Flash as based on Gemini 3.5 Flash. It has not published the model's parameter count, weights, exact architecture, training mixture, or training compute. The public evidence therefore supports comparisons of documented behavior, prices, and evaluations, but not architectural claims about size or internals.^[4]

Release and positioning

Google announced Gemini 3.6 Flash alongside Gemini 3.5 Flash-Lite and the restricted Gemini 3.5 Flash Cyber model. It was made available the same day through the Gemini API, Google AI Studio, Android Studio, Google Antigravity, the Gemini app, and enterprise products.^[1][4] Google Cloud records the model's launch stage as generally available rather than preview.^[6]

The 3.6 designation does not mean that the model is part of a separate Gemini 3.6 family with a corresponding Pro model. Google's model card calls it an addition to the Gemini 3 series and says it is based on Gemini 3.5 Flash.^[4] At launch, Google was still testing Gemini 3.5 Pro with partners.^[1]

Google positioned the release around efficiency rather than a new maximum capability tier. Its announcement said Gemini 3.6 Flash used 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index, took fewer reasoning steps and tool calls, and cost less per output token.^[1] That statement is tied to a particular evaluation suite. It does not establish a fixed token reduction for every prompt, tool harness, or thinking level.

The migration guide also reports fewer unwanted code edits, fewer debugging loops, and less action bias on diagnostic tasks. It says the model is more likely than 3.5 Flash to run diagnostic code before editing, which can improve complex work but add unnecessary exploratory steps to simple tasks. Google separately notes that human evaluators preferred some earlier models for visual layout and styling, even though 3.6 Flash produced more functional code in those tests.^[3]

Model specifications

Gemini 3.6 Flash is a proprietary multimodal reasoning model. Its model card rounds the limits to a one-million-token input and 64,000-token output, while the API reference gives the exact binary values below.^[2][4]

Scroll sideways for more →

The context window is a capacity limit, not a claim that the model recalls every detail throughout a million-token prompt. Retrieval quality depends on the content, query, placement, and evaluation method. The model card's own long-context table reports 91.8% on the 128k cumulative GDM-MRCR v2 measure and 54.0% on the separate pointwise one-million-token measure. Because the two rows use different aggregation, their difference is not a simple degradation curve.^[4][5]

The model produces text only. It can interpret supported media but does not generate images or audio, and the Gemini Live API is not supported.^[2] The model card identifies Gemini 3.5 Flash as its dependency but delegates most architecture and training-data discussion to the earlier card. The 3.5 card, in turn, says that model is based on Gemini 3 Flash.^[8] This chain does not disclose a parameter count, mixture-of-experts layout, post-training recipe, or the exact data added for 3.6.

API behavior and tools

The Gemini Developer API lists context caching, code execution, file search, function calling, Google Search and Maps grounding, URL context, and structured output as supported. It also lists computer use as a preview capability. Batch, Flex, and Priority inference are available as consumption modes.^[2]

These features combine model inference with services outside the model. Search grounding can retrieve current web material beyond the March 2026 knowledge cutoff, and code execution can calculate or inspect data, but neither guarantees that the final response is correct. Function calls and computer actions also depend on schemas, permissions, tool results, and the surrounding agent. The model's benchmark score cannot measure every failure mode introduced by that larger system.

Google documents several migration constraints. Custom temperature, top-P, and top-K settings are ignored. Custom frequency and presence penalties cause an error. Requests that end with a model-authored turn are also rejected, affecting applications that used a prefilled assistant response.^[3][6] These behaviors make Gemini 3.6 Flash more opinionated at the API level than models that expose arbitrary sampling controls.

Platform support is not identical. The Google Cloud Gemini Enterprise Agent Platform page lists Search and Maps grounding, code execution, function calling, caching, chat completions, and a retrieval-augmented generation engine. It marks tuning, the Live API, and Computer Use as unsupported.^[6] The apparent conflict with the Developer API's preview Computer Use feature is channel-specific: a preview tool in one API should not be assumed to exist in another product.

The Cloud page also documents request limits such as up to 3,000 images, 10 videos, and one audio file per prompt. Video is limited to about 45 minutes with audio or one hour without it, while an audio prompt can reach about 8.4 hours or the token ceiling.^[6] These are service limits and may change independently of the underlying model.

Access and pricing

The model is hosted by Google rather than distributed as downloadable weights. The model card lists access through the Gemini app, Gemini Enterprise app, Gemini Enterprise Agent Platform, Google AI Studio, Gemini API, and Google Antigravity.^[4] Product entitlements, quotas, data-handling terms, and tools vary among these surfaces.

Google's Developer API pricing page listed the following rates on July 24, 2026. Prices are in US dollars per one million tokens. Output charges include thinking tokens.^[7]

Scroll sideways for more →

Standard access has a limited free tier for input, output, and cached input. Batch and Flex do not list free-tier access. The pricing page lists cache storage at $1.00 per one million tokens per hour. Search and Maps grounding each include 5,000 paid-tier prompts per month shared across Gemini 3 models, followed by $14 per 1,000 search queries. A prompt can trigger more than one billable query.^[7]

The launch comparison of $7.50 per million output tokens for 3.6 Flash against $9.00 for 3.5 Flash refers to standard paid API pricing.^[1][7] It does not include input volume, cache storage, grounding, priority service, or the number of tokens a particular agent consumes. Cost per completed task therefore depends on both the price table and model behavior.

Reported evaluations

Google's July 2026 model card reports gains over Gemini 3.5 Flash across its selected coding, agent, multimodal, and long-context evaluations.^[4] The figures below are not one uniform test suite. Some were run internally by Google, some came from public leaderboards, and several use distinct harnesses or tool configurations.^[5]

Scroll sideways for more →

The evaluation report says Gemini scores are pass@1 unless noted and that single-attempt settings do not use majority voting or parallel test-time compute. Default sampling was used except where the report says otherwise.^[5]

DeepSWE is an important exception to a direct model comparison. The public leaderboard used high reasoning for Gemini 3.6 Flash and medium reasoning for Gemini 3.5 Flash. The 49% and 37% scores therefore compare available leaderboard configurations rather than holding reasoning effort constant. MLE-Bench used the Partial 30 subset, two runs per problem, an H100-equipped sandbox with internet access, a one-million-token context, and a budget of up to 256 tool calls. OSWorld used five runs, 1080p resolution, and a maximum of 100 steps.^[5]

The with-tools CharXiv row used Search and code execution, while the no-tools row did not. That difference measures a system configuration as well as the base model. Similarly, coding-agent results depend on repository state, command permissions, timeouts, harness prompts, and tool implementations. The published table should not be read as a hardware-independent latency or reliability guarantee.

Independent measurements

Artificial Analysis evaluated Gemini 3.6 Flash at high reasoning before release. Its composite Intelligence Index v4.1 score was 50, the same as Gemini 3.5 Flash, despite changes within individual tests. The organization reported average time per index task of 1.3 minutes for 3.6 Flash versus 2.7 minutes for 3.5 Flash, and estimated average cost per task of $0.50 versus $0.59.^[9]

Its pre-release article measured 304 output tokens per second. The live model page displayed 251.3 output tokens per second and a 14.02-second time to first token when checked on July 24.^[10] These are not necessarily contradictory results. Endpoint load, measurement window, thinking configuration, and service updates can change observed latency. They do show why a single speed figure needs a date and test method.

The Intelligence Index score is a composite, not a percentage. Version 4.1 combines nine evaluations covering knowledge work, terminal use, science, reasoning, long context, and hallucination behavior.^[10] Matching a predecessor on that index is compatible with better results on some Google-selected benchmarks and regressions or smaller changes elsewhere.

Axios independently confirmed the release and its focus on lower deployment cost, but described the performance and token figures as Google claims.^[11] Launch-day reporting is useful for chronology and product context; it is not an independent reproduction of Google's full benchmark table.

Safety evaluation

Google's model card lists hallucination, occasional slowness, timeouts, and the March 2026 knowledge cutoff as known limitations. It says Google strengthened safeguards against chemical, biological, radiological, nuclear, and cyber-offense misuse while training the model to reduce refusals of beneficial requests.^[4]

The card reports relative changes on internal automated safety evaluations against Gemini 3.5 Flash. Text-to-text safety errors decreased by 1.35 percentage points and multilingual safety errors decreased by 5.45 points, with lower defined as better. Image-to-text safety was unchanged. Refusal tone fell by 3.31 points, a regression under the higher-is-better convention, and unjustified refusals rose by 0.25 points, a regression under the lower-is-better convention.^[4]

These figures are differences, not absolute safety rates. Google also warns that it revised its automated evaluations, so the values cannot be compared directly with older model cards. It reports specialist human red teaming, satisfaction of child-safety launch thresholds, and no egregious concerns in the described review scope.^[4]

The frontier-safety conclusion is partly inherited from another model. Google directly evaluated Gemini 3.1 Pro against its Critical Capability Levels, then stated that 3.6 Flash did not show meaningful new capabilities or material increases in the covered domains. It performed additional cyber testing and reported that the model remained below the cyber threshold.^[4] This supports Google's launch decision, but it is not equivalent to publishing full 3.6-specific results for every frontier-risk domain.

Limitations

Gemini 3.6 Flash remains subject to the general limits of hosted foundation models and agent systems.

• It can produce plausible but false text. Search grounding may provide newer material, but it does not guarantee correct synthesis or reliable citations.^[4]
• The March 2026 knowledge cutoff predates the model's July release. Unaugmented answers cannot be assumed to include later events.
• A million-token input does not guarantee complete recall or equal attention across the window. The published long-context results vary substantially by length and scoring method.^[4][5]
• Coding and computer-use scores depend on external harnesses and tools. Real deployments can fail because of permissions, ambiguous interfaces, network state, repository differences, or irreversible actions.
• Computer Use is preview-only in the Developer API and unavailable in the documented enterprise channel. Applications need platform-specific tests, constrained permissions, and confirmation gates for consequential actions.^[2][6]
• Custom sampling values are unavailable, limiting direct control and complicating migration from systems that rely on temperature or assistant-prefill behavior.^[3][6]
• The model is proprietary. The missing weights, parameter count, training mixture, compute, and detailed architecture limit independent auditing of memorization, data provenance, scaling, and environmental cost.^[4]
• Public prices and rate limits can change. End-to-end cost also depends on thinking tokens, tool loops, grounding queries, cached context, retries, and the chosen service mode.^[7]

A reproducible evaluation should record the model ID, date, API surface, thinking level, tool set, prompt and media preprocessing, number of attempts, limits, and pricing used. Without those details, the label Gemini 3.6 Flash identifies the model family but not the system that produced a result.

References

1. Tulsee Doshi. "Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber." *Google*, July 21, 2026. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/ ↩
2. Google. "Gemini 3.6 Flash." *Gemini API Documentation*, updated July 21, 2026. https://ai.google.dev/gemini-api/docs/models/gemini-3.6-flash ↩
3. Google. "Using the latest Gemini models." *Gemini API Documentation*, July 2026. https://ai.google.dev/gemini-api/docs/latest-model ↩
4. Google DeepMind. "Gemini 3.6 Flash Model Card." July 21, 2026. https://deepmind.google/models/model-cards/gemini-3-6-flash/ ↩
5. Google DeepMind. "Gemini 3.6 Flash: Model evaluation approach, methodology and results." July 2026. https://deepmind.google/models/evals-methodology/gemini-3-6-flash ↩
6. Google Cloud. "Gemini 3.6 Flash." *Gemini Enterprise Agent Platform Documentation*, updated July 21, 2026. https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-6-flash ↩
7. Google. "Gemini Developer API pricing." *Gemini API Documentation*. https://ai.google.dev/gemini-api/docs/pricing ↩
8. Google DeepMind. "Gemini 3.5 Flash Model Card." May 19, 2026. https://deepmind.google/models/model-cards/gemini-3-5-flash/ ↩
9. Artificial Analysis. "Gemini 3.6 Flash and Gemini 3.5 Flash-Lite: Halving Time per Task." July 21, 2026. https://artificialanalysis.ai/articles/gemini-3-6-flash-3-5-flash-lite-halving-time ↩
10. Artificial Analysis. "Gemini 3.6 Flash Intelligence, Performance and Price Analysis." July 2026. https://artificialanalysis.ai/models/gemini-3-6-flash ↩
11. Madison Mills. "Google releases series of new cheaper Gemini models." *Axios*, July 21, 2026. https://www.axios.com/2026/07/21/google-gemini-ai-models ↩