Gemini 3.7 Flash review, pricing and verdict
Google's mid-tier Gemini 3 workhorse, tuned for agentic coding and document-heavy business workflows at a fraction of Pro-tier pricing.
- Best for
- Agentic coding pipelines
- Not for
- Terminal-heavy or computer-use agent loops
Should you use Gemini 3.7 Flash?
Gemini 3.7 Flash lifted FrontierCode 1.1 to 43.6% from its predecessor's 34.4%, but it is no longer the model to start with. Google's deprecations page recommends a newer Flash model and says calls to this endpoint are routed there automatically. Read this page for the August 2026 figures and migration context.
Best for
- Agentic coding pipelines DeepSWE v1.1 long-horizon coding score of 65.3% against 48.6% for the prior Flash release.
- High-volume document and business-workflow automation AutomationBench score of 30.4% against Gemini 3.6 Flash's 17.0%, at a blended cost near $1.35 per 1M tokens.
- Web app generation and front-end scaffolding Rates 1588 Elo on Code Arena web development, ahead of Claude Sonnet 5 and GPT-5.6 Terra.
Not for
- Terminal-heavy or computer-use agent loops Trails GPT-5.6 Terra 65.3% to 69.6% on DeepSWE v1.1's long-horizon agent tasks.
- Air-gapped or self-hosted deployment Closed weights, API and enterprise access only.
- New projects starting today Google recommends Gemini 3.8 Flash and routes 3.7 Flash requests to it.
Gemini 3.7 Flash pricing
- Input, per 1M tokens
- $0.75/1M
- Output, per 1M tokens
- $3.75/1M
- Cached input, per 1M tokens
- $0.07/1M
- Pricing model
- per-token
- Free tier
- Available via Google AI Studio's free quota; free-tier prompts may be used to improve Google products, unlike the paid API which is not trained on.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.022 | $0.0037 | $0.026 |
| Support reply | $0.0015 | $0.0011 | $0.0026 |
| One coding agent run | $0.150 | $0.075 | $0.225 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Cost examples
- Summarize a 100K-token research PDF (introductory rate)
- $0.07
- Daily coding agent loop, 1M tokens in / 200K out (introductory rate)
- $1.50
- Same coding agent loop after Jan 1, 2027 standard pricing
- $3.00
Google priced Gemini 3.7 Flash at $0.75 per 1M input tokens and $3.75 per 1M output tokens through December 31, 2026, with cached input near $0.075 per 1M (a 90% discount). From January 1, 2027 the standard rate is $1.50 input and $7.50 output per 1M tokens. Google's model card gives Gemini 3.6 Flash the same introductory rate and the same expiry. Google routes requests for this endpoint to its recommended replacement, so check that model's pricing for new work.
What Gemini 3.7 Flash can do
Gemini 3.7 Flash is a Google DeepMind model released on August 13, 2026 for agentic coding and document workflows. It scores 65.3% on a long-horizon coding benchmark and accepts 1,048,576 input tokens. Google has since named a newer Flash model as its replacement and routes requests for this endpoint to that model.
What goes in and what comes out
Reads
- text
- image
- audio
- video
- tool-calls
Writes
- text
- tool-calls
Capabilities
- Audio inputSupported
- Code executionSupported
- Function callingSupported
- Structured outputSupported
- Tool useSupported
- Video inputSupported
- VisionSupported
- Audio outputNot supported
- Web browsingNot supported
- Reasoning modes
- low-thinking, medium-thinking, high-thinking
- Long context recall
- high
Key features
Customizable Thinking
Thinking settings let developers weigh answer quality against speed and spend on each request.
1M-Token Context Window
Accepts up to 1M tokens of input and returns as many as 65,536 tokens of output, holding 97.0% recall accuracy at 128K depth on Google's GDM-MRCR v2 test.
Multimodal Input
Reads text, images, audio, and video in a single request, with text-only output.
Native Tool Use and Function Calling
Calls tools across multi-step agent loops, which Google highlights as the focus of the algorithmic changes over the prior Flash release.
Discounted Cached Input
System prompts and long documents that get reused across calls are billed at a fraction of the standard input rate once cached, per Google's pricing page.
Known limitations
- No native audio output, only audio input; text-to-speech needs a separate model.
- Trails GPT-5.6 Terra on long-horizon agent coding, 65.3% against 69.6% on DeepSWE v1.1.
- CharXiv Reasoning without tools is 84.5% against 85.2% for Gemini 3.6 Flash, the one tracked regression on the card.
- Closed weights: no self-hosting, air-gapped deployment, or fine-tuning of the base model.
Who it is for
- Engineering teams building agentic coding assistants on a budget
- Enterprises automating document-heavy workflows in finance, law, and life sciences
- Developers building web apps who want fast, cheap front-end scaffolding
- Existing Gemini 3.6 Flash integrators upgrading for the same price
Gemini 3.7 Flash benchmarks
Gemini 3.7 Flash against its nearest rivals
Output speed tokens per second, higher is better
Standout result
DeepSWE v1.1 long-horizon coding score, up 16.7 points from Gemini 3.6 Flash
Source: Google DeepMind model card, read October 26, 2026Where it sits
Priced around the middle of the 104 GA models with a published price (rank 46), rank 4 of 58 on output speed as cited from Artificial Analysis, and one of 42 that document a zero-data-retention option. Ranked against GA models; this record is not GA.
| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $1.50/M | #46 / 104 | peer median $1.70/M |
| Output speedtokens/s | 340 tok/s | #4 / 58 | peer median 95 tok/s |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
Measured figures
- GDPval-AA v2
- 1,525
- WebDev Arena Elo
- 1588
- AA blended price
- $1.35/M
- Output speed
- 340 tok/s
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Strengths and limits
Strengths
- Leads Claude Sonnet 5 and a leading GPT rival on the web-development arena in Google's card.
- Improves production code quality over its predecessor on Google's FrontierCode test.
- Raises enterprise workflow automation accuracy well above its predecessor on AutomationBench.
- The 1M-token window keeps recall high on Google's long-context test, so the size is usable rather than nominal.
Show 2 more
- Scores 65.3% on DeepSWE v1.1 long-horizon coding, up from 48.6% for Gemini 3.6 Flash.
- Blended per-token cost near $1.35 per 1M (80/20 input/output) at the introductory rate.
Limits
- Superseded: hidden from the AI Studio picker, with Google steering new work to its replacement.
- Trails a leading GPT rival on long-horizon agent and terminal tasks in Google's own comparison table.
- Closed weights: available only through Google surfaces, with no self-hosting or fine-tuning of the base model.
- The introductory price ends with 2026, after which list pricing doubles on both input and output.
Gemini 3.7 Flash alternatives
The nearest AI models on Hokai for the same job.
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| Gemini 3.7 Flash This page | 1M | $0.75 / $3.75 | text, image, audio, video → text | |
 Claude Fable 5Superseded by Fable 5.1 and Fable 5.2 but still sold on the Claude API: Anthropic's first generally available Mythos-class model, released June 2026, with a 1M-token context. | Anthropic | 1M | $10 / $50 | text, image, pdf → text |
| Claude Fable 5.1Superseded by Claude Fable 5.2 in November 2026 but still sold on the Claude API: Anthropic's second Mythos-class model, released Sept 1, 2026, for long-horizon coding and research. | Anthropic | 1M | $10 / $50 | text, image, pdf → text |
 Gemini Decision 1Google DeepMind's decision model scores many questions over one multimodal input of up to 262,144 tokens and bills zero output tokens. | 262K | $0.10 / $0 | text, image, audio, video, pdf → text |
Gemini 3.7 Flash specs
- Context window
- 1.0M tokens
- Max output
- 66K tokens
- Latency (P50)
- 9830 ms
- Tokens per second
- 340.1 tok/s
- Mixture of experts
- No
- Parameters
- Not disclosed
- Architecture
- Transformer (proprietary; exact architecture not publicly disclosed)
Training data
Google states a primary training-data cutoff of March 2026, though some specialized domains carry knowledge only through January 2025. The model card describes automated multilingual content-safety evaluation and external red-teaming but does not disclose training corpus composition or synthetic-data ratio.
API endpoints
| Name | Endpoint | Auth |
|---|---|---|
| Gemini API (Google AI Studio) | https://ai.google.dev/gemini-api/docs | API Key |
| Google Vertex AI | https://cloud.google.com/vertex-ai | GCP IAM |
| Gemini Enterprise Agent Platform | https://docs.cloud.google.com/gemini-enterprise-agent-platform | GCP IAM / Enterprise SSO |
SDKs
- Python
- JavaScript
- REST
Prompting patterns
Show 2 prompting patterns
Set an explicit thinking level per request
thinkingConfig: { level: 'high' } for hard reasoning, 'low' for latency-sensitive calls3.7 Flash exposes low/medium/high thinking tiers that trade quality against cost and latency; the default medium setting under-uses the model on genuinely hard tasks.
Drop deprecated sampling parameters
Remove temperature, top_p, top_k, and prefilled model turns from Gemini 3.5-era prompts
Gemini 3.6 Flash and later, including 3.7 Flash, no longer support these legacy parameters; leaving them in a migrated prompt causes request errors.
Known quirks
Show 4 known quirks
Introductory price expires Dec 31, 2026
The launch price of $0.75/$3.75 per 1M input/output tokens is a temporary promotion tied to the calendar, not a permanent rate; it reverts to $1.50/$7.50 on January 1, 2027.
Workaround: Budget production workloads at the post-2027 standard rate, or lock in usage-based commitments before the window closes.
No longer accepts legacy sampling parameters
Requests carrying temperature, top_p, top_k, or prefilled model turns from Gemini 3.5-era integrations fail, since this behavior was already removed in Gemini 3.6 Flash.
Workaround: Strip these fields from any prompt template migrated from Gemini 3.5 Flash or earlier before calling the 3.7 Flash endpoint.
Reserved, cautious default tone
Early reviewers note the model defaults to a notably cautious, hedged tone compared to rival Flash-tier models, which can read as overly conservative in consumer-facing chat use.
Workaround: Add explicit tone instructions in the system prompt (e.g. 'answer directly, skip hedging') for consumer-facing deployments.
Small regression on tool-free visual reasoning
CharXiv Reasoning without tool access dipped to 84.5%, a 0.7-point regression from Gemini 3.6 Flash, even though most other benchmarks improved.
Workaround: For chart and figure-heavy reasoning tasks, allow the model tool access (code execution) rather than relying on pure no-tools visual reasoning.
Gemini 3.7 Flash facts
- Deployment
- API · Google Vertex AI
- Trains on your data
- No
- Licence
- Proprietary (Gemini API Terms of Service)
- Safety posture
- Balanced
- Released
- 13 Aug 2026
- Current version
- gemini-3.7-flash
- Knowledge cutoff
- 01 Mar 2026
- Family
- Gemini 3
- Variant
- Multimodal
- Data retention
- On the paid Gemini API and Vertex AI / Gemini Enterprise Agent Platform, prompts and responses are not used to train models; they're logged briefly for abuse detection and legal compliance. On the free Google AI Studio tier, submitted content may be used to improve Google products and reviewed by humans. Enterprise customers can request zero data retention, which strips prompts, responses, and identifiable metadata before logging.
Content restrictions
- CBRN (chemical, biological, radiological, nuclear) uplift content blocked under Google's Frontier Safety Framework.
- Cyber-offense content restricted per the April 2026 Frontier Safety Framework assessment.
- Child safety evaluations applied per Google's standard protections.
Safety controls
- Rlhf
- Yes
- System prompt
- Yes
- Constitutional AI
- No
- Content filter API
- Yes
- Red teaming
- Automated multilingual safety evaluations plus manual red-teaming by specialist teams external to the development team, per the August 2026 model card.
Data governance
- Iso 27001
- Yes
- Soc2 type2
- Yes
- GDPR compliant
- Yes
- HIPAA eligible
- Yes
- Trains on inputs
- No
- Eu AI act classification
- general-purpose AI (GPAI) model
Version history
gemini-3.8-flash · 02 Sep 2026
Recommended replacement per Google's deprecations page. Gemini 3.7 Flash requests are routed to it; no shutdown date announced.
gemini-3.7-flash · 13 Aug 2026
Algorithmic refinement of Gemini 3.6 Flash, not a new pretraining run. DeepSWE v1.1 65.3% (from 48.6%), AutomationBench 30.4% (from 17.0%), Code Arena 1588 Elo (from 1538). Introductory pricing $0.75/$3.75 per 1M tokens.
gemini-3.6-flash · 21 Jul 2026
Predecessor. DeepSWE v1.1 48.6%, AutomationBench 17.0%, Code Arena 1538 Elo on the Gemini 3.7 Flash card.
What changed
Questions about Gemini 3.7 Flash
How much does Gemini 3.7 Flash cost per 1M tokens?
Until the end of 2026 Google billed $0.75 for each million input tokens and $3.75 for each million output tokens, with cached input near $0.075. After the new year the list rate becomes $1.50 input and $7.50 output, doubling both meters.
How does Gemini 3.7 Flash compare with Claude Sonnet 5 on benchmarks?
On Google's AutomationBench for enterprise workflows, Gemini 3.7 Flash scores 30.4% against Claude Sonnet 5's 10.7% and GPT-5.6 Terra's 23.6%. Claude Sonnet 5 still wins on Agent's Last Exam desktop tasks, posting 33.3% against 26.3%. These are vendor-reported results from Google's model card, so the better choice depends on whether the workload is document-heavy or computer-use-heavy.
Is Gemini 3.7 Flash open weights or API-only?
Gemini 3.7 Flash is closed-weights: Google has not released model files, so access runs through the Gemini API, Google AI Studio, the Gemini Enterprise Agent Platform and the Gemini app. There is no self-hosting, no fine-tuning of the base weights, and no air-gapped deployment option.
Is Gemini 3.7 Flash deprecated, and what replaces it?
Yes, in the sense that Google now recommends Gemini 3.8 Flash, released September 2, 2026. The older model is hidden from the AI Studio picker, yet its endpoint is not shutting down: requests are forwarded to the newer one, and Google has announced no shutdown date.
Should a new project still pick this model?
No. New work should start on the recommended replacement, since Google forwards traffic from this endpoint there. Use this page as the August 2026 record of launch figures, and look at an open-weights model instead if self-hosting is a requirement.
About Gemini 3.7 Flash
Google DeepMind released Gemini 3.7 Flash on August 13, 2026, three weeks after Gemini 3.6 Flash, as the Flash-tier entry of the Gemini 3 lineup: a fast, low-cost model for coding and agent work from Google DeepMind. It has since been superseded. Google's deprecations page lists Gemini 3.8 Flash, released September 2, 2026, as the recommended replacement, notes that 3.7 Flash can no longer be selected in Google AI Studio, and says the endpoint is not shutting down but routes API requests to Gemini 3.8 Flash automatically. Google has announced no shutdown date. Treat the figures below as the August 2026 launch record, not the current state of the art.
Read the full background
Google describes 3.7 Flash as algorithmic improvements to the reasoning foundation of 3.6 Flash with support for agentic video understanding, not a new pretraining run, and has not disclosed a parameter count or whether it is a mixture-of-experts design.
On Google's published model card, 3.7 Flash scores 65.3% on DeepSWE v1.1 (long-horizon coding) against 48.6% for 3.6 Flash, and 43.6% on FrontierCode 1.1 against 34.4%. Its Code Arena web-development rating is 1588 Elo, against 1541 for Claude Sonnet 5 and 1523 for GPT-5.6 Terra. On AutomationBench it scores 30.4%, against 10.7% for Claude Sonnet 5 and 23.6% for GPT-5.6 Terra, and on GDP.pdf document comprehension it reaches 34.0%, up from 22.0%. Those are vendor-reported figures. The card also reports an Artificial Analysis Intelligence Index of 56 (Claude Sonnet 5 at 55, GPT-5.6 Terra at 57), but that index is versioned: the Artificial Analysis page read on 2026-10-26 shows 39 on its current v4.3.2 index, and says the model is deprecated and no longer benchmarked beyond the default 10k-token workload. Do not compare the two numbers directly.
The model accepts up to 1,048,576 input tokens and returns as many as 65,536 output tokens. On Google's GDM-MRCR v2 long-context test it holds 97.0% at 128K average depth. Inputs cover text, images, audio, and video; output is text only. Google supports customizable thinking configurations that trade response quality against latency and cost.
Google priced 3.7 Flash at an introductory rate that runs to the end of 2026 and doubles on both input and output from the start of 2027; the price notes carry the exact figures. The card footnote lists the identical introductory rate and expiry for its predecessor, so the discount is not unique to this model. See the cheapest LLM API guide for how that rate compares, and Gemini 3.5 Flash-Lite for the lower tier.
The model was offered through the Gemini API and Google AI Studio, Google's Gemini Enterprise Agent Platform (see Google Vertex AI), and the Gemini app. It is a closed-weights model, so there are no downloadable weights, VRAM requirements, or quantization options. Google's card describes automated safety evaluations, manual red teaming by specialist teams outside the development group, and an assessment under its April 2026 Frontier Safety Framework, with updated safeguards for CBRN and cyber-offense misuse. Named limitations include hallucination, jailbreak risk, occasional slowness or timeouts, and a March 2026 knowledge cutoff with some domains limited to January 2025.
Where the alternatives win: new projects should start on Gemini 3.8 Flash, since 3.7 Flash can no longer be picked in AI Studio. Teams building terminal-heavy or computer-use agent loops should look at GPT-5.6 Terra, which scores 69.6% on DeepSWE v1.1 and 20.8% on Terminal-bench 3.0 against 65.3% and 14.9% here. Anyone needing self-hosted or air-gapped deployment needs an open-weights model, because this one is API-only. To pick between current options, read how to choose the right LLM, the Grok 4.7, Opus 5, Astra and Gemini 3.8 comparison and the best AI coding assistants, or browse the model directory.
Data handling, as last stored from Google's terms (not re-read in this refresh): paid usage through the Gemini API and the enterprise platform is excluded from model training, with brief abuse and legal-compliance logging, while the free AI Studio tier may use submitted content to improve Google products. For a Claude-side view, see which Claude model to use.
Sources and guides
Official sources
- Gemini 3.7 Flash model page Vendor
- Licence text Licence
- System card Safety
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Google DeepMind's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.
Hokai guides that mention Gemini 3.7 Flash
Lists that include Gemini 3.7 Flash
The provider
More AI Models on Hokai
Shieldstral 1.0Mistral's open-weight multimodal safety classifier, released 4 August 2026, that screens text and images against policies you write.
- Claude Mythos 5.2Anthropic's November 2026 Mythos model: Fable 5.2's weights with relaxed cyber and biology safeguards for verified teams, with a 1M-token context.
ERNIE-5.5-Flash-PreviewBaidu's text-only preview mixture-of-experts model, served through Qianfan on OpenRouter with a 1M-token context window and tool calling.
- Gemini 4 CarbonGoogle DeepMind's Gemini 4 Carbon, open to paying Gemini API developers since 10 November: a reasoning model for agentic coding, research and cyber defense.
JevTypeSafe's first System One Model, launched in 2026: skips text generation for typed, calibrated decisions software can act on directly.Nex-N2.6Nex AGI's hosted agentic model for computer use and coding, with image input, tool calling and about 321B parameters, served first-party on OpenRouter.