Gemini 3.7 Flash

Gemini 3.7 Flash is a proprietary, multimodal large language model released by Google on August 13, 2026. It is a generally available model in the Gemini 3 series and uses the stable API identifier gemini-3.7-flash. Google developed it for coding, tool-using agents, multimodal analysis, and other multi-step tasks.^[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 low, medium, and high thinking levels, with medium as the default. The minimal setting is not supported.^[3][4]

Google DeepMind describes Gemini 3.7 Flash as based on Gemini 3.6 Flash, with algorithmic changes to its reasoning foundation. Google has not published the 3.7 model's weights, parameter count, exact architecture, training mixture, or training compute. The public documentation supports comparisons of observable behavior, service features, prices, and named evaluations, but not claims about undisclosed internals.^[6][8]

Release and relationship to Gemini 3.6 Flash

Google released Gemini 3.7 Flash 23 days after Gemini 3.6 Flash, a gap the company described as three weeks. The launch post attributed the short cycle to developer feedback and algorithmic work that Google expected to use in later models. The Gemini API release notes and both model pages classify the new endpoint as stable and generally available rather than preview.^[1][2][3][9]

The 3.7 label does not identify a separate family with a corresponding 3.7 Pro model. The model card calls it the next iteration in Gemini 3 and states that it is based on 3.6 Flash. The earlier card says 3.6 Flash is based on Gemini 3.5 Flash. Most of the 3.7 card's architecture, data, hardware, and software sections refer readers back to the 3.6 documentation instead of providing new technical detail.^[6][8]

For developers moving from earlier endpoints, Google recommends gemini-3.7-flash as the target for Gemini 3.6 Flash, Gemini 3.5 Flash, Gemini 3 Flash Preview, and Gemini 3.1 Pro workloads. The migration guide says applications should remove deprecated temperature, top-p, and top-k settings, replace thinking budgets with named thinking levels, remove candidate count, and avoid prefilled model turns. These are API compatibility requirements, not evidence that every older prompt will behave identically after migration.^[4]

Model specifications

Scroll sideways for more →

The model card rounds the input capacity to one million tokens and the output capacity to 64,000 tokens, while the API pages give the exact values in the table. A context limit describes how much material a request can contain. It does not guarantee perfect recall, equal attention across the prompt, or correct synthesis of every document.^[3][6]

Google lists March 2026 as the knowledge cutoff. The same model card adds an unusual qualification: some domains may contain updates through that date, while knowledge in other domains can be limited to January 2025, in line with the wider Gemini 3 family. Search grounding can provide newer material, but it does not remove the model's documented hallucination risk.^[6]

Gemini 3.7 Flash is a text-output model. It can analyze supported media, but it does not generate images or audio and does not support the Gemini Live API. The Gemini Developer API lists caching, code execution, file search, function calling, Search and Maps grounding, structured output, URL context, and computer use in preview. Batch, Flex, and Priority inference are also supported.^[3]

Access and product integration

Developers can use the model through the Gemini API and Google AI Studio. Google's launch announcement also names Android Studio and Google Antigravity. The API guide says 3.7 Flash became the default underlying model for the Antigravity agent in Gemini Managed Agents and the Antigravity SDK.^[1][4]

On Google Cloud, the model is available in Gemini Enterprise Agent Platform Model Garden and the Gemini Enterprise app. The Cloud model page uses the same gemini-3.7-flash identifier, marks it GA, and lists global, United States, and European Union service regions. For users familiar with Google Vertex AI, the current 2026 documentation places this model under the Gemini Enterprise Agent Platform namespace rather than assigning it a separate Vertex-specific model ID.^[1][9]

Consumer access was narrower than a general model-picker rollout. At launch, Google said Gemini Spark in the Gemini app would run on 3.7 Flash for Google AI Pro and Ultra subscribers in supported countries. Google described Spark as available in more than 160 countries and said the update was intended to improve its use of Google Workspace tools. This does not establish that every Gemini app conversation or subscription tier used 3.7 Flash.^[1][6]

The model remains hosted by Google. The model card lists no downloadable weights or self-hosting channel. Product entitlements, quotas, data handling, tool availability, and service-level terms differ among the consumer app, Developer API, and enterprise platform.^[5][6][9]

Pricing

Google introduced temporary pricing for both Gemini 3.7 Flash and Gemini 3.6 Flash. The standard paid rate is $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. On January 1, 2027, the scheduled standard rates become $1.50 for input and $7.50 for output. Output billing includes thinking tokens.^[4][5][10]

Scroll sideways for more →

The launch post described the introductory price as half the original 3.6 Flash cost. That comparison concerns the earlier list price. During the promotion, Google's pricing table applies the same rates to 3.6 and 3.7 Flash. Standard cached input costs $0.075 per million tokens during the promotion and $0.15 afterward, with a separate cache-storage charge. A limited free tier is also listed, subject to Google's access and data-use terms.^[1][5]

Token price is not the same as cost per completed task. Agent cost also depends on thinking tokens, tool calls, cached context, retries, grounding queries, and whether the application selects Standard, Batch, Flex, or Priority service. The temporary rates should therefore be dated rather than presented as permanent model characteristics.

Reported evaluations

Google's August 2026 model card reports improvements over Gemini 3.6 Flash on many coding, agent, document, computer-use, and long-context tests. The table below reproduces a limited selection rather than treating the launch suite as one combined score.^[6][7]

Scroll sideways for more →

The CharXiv results are slightly lower for 3.7 Flash, so the published evidence does not support a claim that the new model improved on every evaluation. The launch post gives the 3.6 DeepSWE result as 49.0%, while the model card and methodology report list 48.6%. The difference appears to be rounding or a snapshot change; the table uses the model card value.^[1][6][7]

Google's methodology says Gemini scores are pass@1 unless otherwise noted and use default sampling unless a test specifies another setting. Some results came from public leaderboards, while others were computed by Google. For DeepSWE, Google ran 3.7 Flash with high thinking, a mini software-engineering agent, and LiteLLM 1.96, while the 3.6 number came from Datacurve's public leaderboard. Terminal-bench and several multimodal, long-context, and scientific results were also self-computed.^[7]

The benchmark operators measure different systems. FrontierCode uses maintainer-authored repository tasks and evaluates whether a proposed change is mergeable, including correctness, testing, scope, style, and repository conventions.^[12] DeepSWE focuses on original long-horizon engineering tasks and reports verifier audits intended to reduce grading errors.^[13] AutomationBench tests end-to-end business workflows across applications and scores the final environment state; Google's table uses its held-out private set.^[14] Code Arena derives a live score from human pairwise preferences between generated web applications, so its Elo is date-sensitive.^[15]

These methodological differences limit direct conclusions. A score can reflect the model, thinking setting, agent prompt, tool harness, repository state, media preprocessing, step limit, and scoring rule. Google also states that most non-Gemini comparison figures were provider-reported, often at the highest available reasoning setting. The launch table is therefore a configuration-specific snapshot, not a hardware-independent reliability guarantee.^[7]

Independent measurements

Artificial Analysis independently evaluated the high-thinking version through Google's API. Its live page reported an Intelligence Index v4.1.1 score of 56.03, rounded to 56, matching the rounded value in Google's table. The index combines nine evaluations and is not a percentage.^[11]

Artificial Analysis measured 340.1 output tokens per second and a 9.83-second time to first token. It also reported 64 million output tokens for the full index run and an estimated weighted cost of about $0.40 per task at the introductory price. Those values describe its test date and API configuration. Endpoint load, prompt length, streaming behavior, and thinking level can change observed speed and latency.^[11]

As of the day after release, there was no peer-reviewed, model-specific study reproducing Google's complete evaluation table. The independent Artificial Analysis run and the benchmark creators' published methods provide useful checks, but neither establishes performance for every production workload.

Safety and limitations

Google's model card says Gemini 3.7 Flash can hallucinate and may occasionally be slow or time out. It reports mixed changes against 3.6 Flash on internal automated safety tests and characterizes the overall result as similar across safety and refusal tone. Google also cautions that revised test sets make those figures unsuitable for direct comparison with older model cards.^[6]

Specialist teams outside the model-development group conducted human red teaming. Google says the model met its child-safety launch thresholds and showed similar or improved content-safety performance relative to 3.6 Flash within the tested scope. These are Google assessments, not an external audit.^[6]

Under Google's Frontier Safety Framework, Gemini 3.7 Flash did not reach any tracked or critical capability level. It reached alert thresholds in parts of the chemical, biological, radiological, nuclear, and cybersecurity assessments but remained below the corresponding critical levels. Google shipped updated safeguards for CBRN and cyber-offense misuse.^[1][6]

The model's proprietary status limits independent analysis of training-data provenance, memorization, parameter scaling, and training compute. Its million-token input capacity does not guarantee complete long-context recall, and its tool support can introduce failures outside the language model itself. Computer-use and agent deployments need scoped permissions, logs, confirmation steps for consequential actions, and tests in the actual application environment.

References

1. ^Google. *Introducing Gemini 3.7 Flash*. August 13, 2026. blog.google/...introducing-gemini-3-7-flash
2. ^Google AI for Developers. *Gemini API release notes*. August 13, 2026. ai.google.dev/...changelog
3. ^Google AI for Developers. *Gemini 3.7 Flash*. ai.google.dev/...gemini-3.7-flash
4. ^Google AI for Developers. *What's new in Gemini 3.7 Flash*. ai.google.dev/...latest-model
5. ^Google AI for Developers. *Gemini Developer API pricing*. ai.google.dev/...pricing
6. ^Google DeepMind. *Gemini 3.7 Flash Model Card*. August 13, 2026. deepmind.google/...gemini-3-7-flash
7. ^Google DeepMind. *Gemini 3.7 Flash: Model evaluation approach, methodology and results*. August 2026. deepmind.google/...gemini-3-7-flash
8. ^Google DeepMind. *Gemini 3.6 Flash Model Card*. July 21, 2026. deepmind.google/...gemini-3-6-flash
9. ^Google Cloud. *Gemini 3.7 Flash*. docs.cloud.google.com/...3-7-flash
10. ^Google Cloud. *Agent Platform pricing*. cloud.google.com/...pricing
11. ^Artificial Analysis. *Gemini 3.7 Flash (high): Intelligence, Performance and Price Analysis*. August 2026. artificialanalysis.ai/...gemini-3-7-flash
12. ^Cognition. *FrontierCode 1.1*. July 7, 2026. cognition.com/...frontier-code-1.1
13. ^Datacurve. *DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks*. 2026. arxiv.org/...2607.07946
14. ^Zapier. *AutomationBench*. 2026. arxiv.org/...2604.18934
15. ^Arena.ai. *WebDev Arena: A Live LLM Leaderboard for Web App Development*. arena.ai/...webdev-arena