Google Gemini 3.7 Flash: Price, Benchmarks and Why Its ‘Workhorse’ Strategy Matters
Google’s new Gemini 3.7 Flash is cheaper than many of its closest rivals, substantially stronger than Gemini 3.6 Flash across coding and agentic tasks, and built to be used repeatedly inside autonomous workflows. Its significance may lie less in winning the AI intelligence race than in changing the economics of putting intelligence to work.
Google launched Gemini 3.7 Flash on 13 August 2026, only three weeks after releasing Gemini 3.6 Flash. On the surface, it looks like another unusually rapid iteration in an AI market that has become accustomed to models arriving every few weeks. Look more closely, however, and the release reveals something more interesting about where Google believes the commercial AI race is heading.
Gemini 3.7 Flash is not being presented as Google DeepMind’s ultimate frontier model. Google calls it its most intelligent “workhorse” yet: a multimodal reasoning model designed particularly for coding, autonomous agents, web development and enterprise workflows, with a one-million-token context window and an introductory API price of $0.75 per million input tokens and $3.75 per million output tokens. Google says the model is based on Gemini 3.6 Flash but incorporates algorithmic improvements to its reasoning foundation, alongside configurable thinking settings that allow developers to balance quality, cost and latency.
That combination matters. The latest generation of artificial intelligence is beginning to move beyond a model in a browser answering one question at a time. Coding agents may inspect a repository, devise a plan, modify files, run tests, discover an error and try again. Enterprise agents may search documents, invoke software tools, update systems and continue through several stages before completing a job. In these environments, intelligence has to be invoked repeatedly, which means the economics of every model call begin to compound.
Gemini 3.7 Flash is therefore an attempt to answer a question that is becoming increasingly important for the AI industry: how intelligent does a model need to be before making that intelligence cheaper, faster and more reliable becomes more valuable than making it slightly smarter?
That question may ultimately matter much more to businesses than which company briefly occupies the top of a benchmark leaderboard.
Gemini 3.7 Flash at a glance
| Release date | 13 August 2026 |
|---|---|
| Developer | Google DeepMind |
| Model family | Gemini 3 |
| Based on | Gemini 3.6 Flash |
| Primary use cases | Coding, agents, enterprise workflows, knowledge work |
| Inputs | Text, images, audio and video |
| Output | Text |
| Context window | Up to 1 million tokens |
| Maximum output | 64,000 tokens |
| Introductory input price | $0.75 per 1M tokens |
| Introductory output price | $3.75 per 1M tokens |
| Introductory pricing ends | 31 December 2026 |
| Price from 1 January 2027 | $1.50 input / $7.50 output per 1M tokens |
| Knowledge cutoff | March 2026, with some domains potentially limited to January 2025 |
| Availability | Gemini API, AI Studio, Antigravity, Gemini Enterprise and Gemini Spark |
Google says Gemini 3.7 Flash supports text, images, audio and video inputs, while generating text outputs of up to 64,000 tokens. It can process a context window of up to one million tokens and allows developers to alter its thinking configuration according to the required balance between reasoning quality, latency and cost.
What is Gemini 3.7 Flash?
Gemini 3.7 Flash is the latest model in the faster and more economical branch of Google’s Gemini family. Rather than attempting to maximise intelligence regardless of computational expense, Flash models are intended to provide enough reasoning capability for demanding applications while remaining practical for workloads in which a model may need to be called repeatedly.
This distinction has become increasingly blurred as smaller and cheaper models have improved. The term “Flash” once implied a relatively obvious trade-off: businesses could choose the most capable model when quality mattered or the faster model when speed and cost mattered more. Gemini 3.7 Flash demonstrates how quickly that assumption is disappearing. Google’s own results place the model close to, and on particular evaluations ahead of, much more expensive systems from OpenAI and Anthropic.
Google says 3.7 Flash is based on Gemini 3.6 Flash rather than being a completely new architecture. The company attributes the improvement to algorithmic advances in its core reasoning foundation and developer feedback from the previous release. That also explains the extraordinarily short three-week gap between the two models. Google is presenting this not as a generational reset but as evidence that it can now improve the intelligence and behaviour of the Flash family unusually quickly.
The result is a model aimed squarely at software engineering, autonomous agents and knowledge-intensive business processes. These are attractive markets because they involve enormous potential volumes of model usage, but they also expose weaknesses that are easier to hide in ordinary chatbot conversations. An agent that misunderstands an instruction early in a workflow can waste subsequent model calls, invoke the wrong tool or require human intervention. Reliability is therefore inseparable from cost.
Google’s claim for Gemini 3.7 Flash is not simply that each call costs less. It is that better planning, stronger instruction-following and fewer retries should allow developers to complete more useful work with those calls.
Gemini 3.7 Flash benchmarks: where has it actually improved?
Google’s published benchmarks show a substantial improvement over Gemini 3.6 Flash across many of the tasks most closely associated with agentic AI. On FrontierCode 1.1 Main, an evaluation of production code quality, Gemini 3.7 Flash scores 43.6 per cent compared with 34.4 per cent for 3.6 Flash. On DeepSWE v1.1, which tests longer software-engineering tasks, the improvement is from 48.6 to 65.3 per cent. Terminal-bench 2.1 rises from 78.0 to 85.8 per cent, while AutomationBench, designed around enterprise workflow automation, increases from 17.0 to 30.4 per cent.
The pattern matters more than any individual score. Google has not merely concentrated on abstract reasoning tests. Some of the largest improvements concern the ability to operate across longer sequences of actions: software engineering, tool use, workflow automation and interaction with computer environments.
| Benchmark | Gemini 3.7 Flash | Gemini 3.6 Flash | Change |
|---|---|---|---|
| FrontierCode 1.1 Main | 43.6% | 34.4% | +9.2 pts |
| DeepSWE v1.1 | 65.3% | 48.6% | +16.7 pts |
| Code Arena | 1588 Elo | 1538 Elo | +50 Elo |
| Terminal-bench 2.1 | 85.8% | 78.0% | +7.8 pts |
| Terminal-bench 3.0 | 14.9% | 5.4% | +9.5 pts |
| AutomationBench | 30.4% | 17.0% | +13.4 pts |
| GDPVal-AA v2 | 1525 Elo | 1422 Elo | +103 Elo |
| GDP.pdf | 34.0% | 22.0% | +12.0 pts |
| OSWorld-2.0 | 47.9% | 33.8% | +14.1 pts |
These figures come from Google DeepMind’s launch evaluation and should be interpreted accordingly. Benchmark results are useful evidence, but they are not equivalent to independent testing inside a live business. Different models can also be run with different reasoning settings, scaffolds and evaluation configurations, which means a headline percentage should never be read as a universal measure of superiority. Google itself notes that its evaluation encompasses reasoning, coding, agentic tool use, multimodal performance and long context, while the underlying methodology varies between tests.
The strongest conclusion the data supports is therefore narrower but still significant: Gemini 3.7 Flash appears to be a materially more capable model than the 3.6 Flash system it replaces, and the improvements are concentrated in exactly the areas Google wants developers to use it for.
Why coding is central to Gemini 3.7 Flash
Software development has become one of the most valuable proving grounds in artificial intelligence because code gives a model something ordinary language often does not: an external test of whether its work actually functions.
A model can produce a beautifully written but incorrect explanation without immediately revealing the error. Software is less forgiving. The code either compiles or it does not; the tests pass or they fail; the application behaves as intended or something breaks.
AI coding is consequently moving from isolated code generation towards longer-horizon engineering. A capable coding agent may need to understand a large repository, locate the source of a bug, decide which files require modification, write those changes, execute tests, interpret the failures and continue until a working solution emerges. This demands planning and memory as much as knowledge of programming syntax.
Gemini 3.7 Flash’s 65.3 per cent score on DeepSWE v1.1 therefore says something more interesting than whether the model can write code snippets. It suggests a substantial improvement in Google's ability to sustain software-engineering work across a sequence of steps, although GPT-5.6 Terra remains ahead on the same benchmark at 69.6 per cent in Google’s published comparison.
Google performs better on other coding evaluations. Gemini 3.7 Flash scores 43.6 per cent on Frontier Code 1.1 Main, compared with 42.7 per cent for Claude Sonnet 5 and 41.3 per cent for GPT-5.6 Terra. On Code Arena, Google reports an Elo rating of 1588 for 3.7 Flash, ahead of Claude Sonnet 5 at 1541 and GPT-5.6 Terra at 1523. Again, these cross-provider comparisons come from Google's own model card, so they should be treated as useful but not definitive evidence.
What Google is clearly trying to establish is that developers no longer need to make a straightforward choice between frontier capability and Flash economics.
Gemini 3.7 Flash vs GPT-5.6 Terra and Claude Sonnet 5
The comparison with rival models is where the launch becomes particularly interesting. Google’s model card places Gemini 3.7 Flash alongside GPT-5.6 Terra, Claude Sonnet 5 and Meta’s Muse Spark 1.2 across a wide set of evaluations. There is no universal winner. GPT-5.6 Terra remains stronger on several demanding long-horizon agent tasks, Claude Sonnet 5 performs particularly well on knowledge work, and Gemini leads several coding, web-development and workflow-automation tests.
The major difference is price.
| Model | Input price / 1M tokens | Output price / 1M tokens | Selected strength |
|---|---|---|---|
| Gemini 3.7 Flash | $0.75* | $3.75* | Web development, production code, workflow automation |
| Claude Sonnet 5 | $2.00 | $10.00 | Knowledge work |
| GPT-5.6 Terra | $2.00 | $12.00 | Long-horizon engineering and terminal agents |
*Gemini 3.7 Flash introductory pricing runs until 31 December 2026. From 1 January 2027, pricing rises to $1.50 per million input tokens and $7.50 per million output tokens.
Google reports an Artificial Analysis Intelligence Index score of 56 for Gemini 3.7 Flash, against 57 for GPT-5.6 Terra and 55 for Claude Sonnet 5. A one- or two-point difference on a composite benchmark is not evidence that the models are interchangeable, but it illustrates the commercial proposition unusually clearly: Google is offering a model it portrays as being in roughly the same broad capability class at a substantially lower introductory token price.
The comparison becomes more nuanced at task level. GPT-5.6 Terra scores 87.4 per cent on Terminal-bench 2.1 against Gemini’s 85.8 per cent and leads DeepSWE v1.1 by 69.6 to 65.3 per cent. Gemini, however, scores 30.4 per cent on AutomationBench against 23.6 per cent for Terra and 10.7 per cent for Claude Sonnet 5, while also leading the three models on FrontierCode 1.1 and Code Arena in Google's table. Claude Sonnet 5 leads Gemini on GDPVal-AA v2, an evaluation of knowledge work, with an Elo score of 1598 against 1525.
This is a healthier way of understanding modern AI competition than asking which model is simply “best”. The answer depends increasingly on the work being performed.
For an organisation running large numbers of web-development or workflow-automation agents, Gemini's economics may be particularly compelling. For difficult terminal-based software engineering, a more expensive model that succeeds more often might prove cheaper per completed task. For certain knowledge-work applications, Claude may continue to justify its higher unit price.
Cost per token is therefore becoming less meaningful than cost per successful outcome.
That is arguably the most important idea contained in the Gemini 3.7 Flash launch.
The economics of AI agents are changing the model race
A chatbot makes model economics relatively easy to understand. A user sends a request and the model sends an answer. If the interaction consumes 10,000 tokens, the cost can be calculated fairly directly.
An agent does not necessarily work like that. It can receive one objective and make many model calls while attempting to achieve it. The system might formulate a plan, search for information, invoke an API, analyse what returns, discover that its original strategy has failed, revise the plan, call another tool and continue until the work is complete.
Every additional loop adds cost and latency. More importantly, every mistake can create further loops.
This means a model that is marginally more expensive per token can still be cheaper per job if it completes the task with fewer retries. The reverse is also true: an inexpensive model can become costly if its lower reliability causes the surrounding agent to repeatedly reconsider its work.
Google explicitly connects Gemini 3.7 Flash’s improved reasoning to this problem. Its launch announcement says the model is better at adapting to roadblocks, clarifying intent and executing multi-step plans, with fewer retries and less manual oversight across engineering workflows. Those are Google’s claims and require independent verification, but they reveal the economic metric the company is trying to improve.
For an AI industry still accustomed to comparing subscription prices and token rates, this is a significant shift. The relevant calculation for agentic AI is increasingly not “What does this model cost?” but “What does it cost this model to finish the work correctly?”
Gemini 3.7 Flash is also a multimodal model
Coding dominates Google's launch narrative, but Gemini 3.7 Flash is not a coding-only system. The model accepts text, images, audio and video within its one-million-token context window, allowing a single workflow to combine information that conventional enterprise systems may store in very different forms.
Google’s own evaluations highlight this breadth. The model records 85.4 per cent on LVBench for long-video understanding and 97.0 per cent on Google DeepMind’s long-context MRCR evaluation at 128,000 tokens. It scores 34.0 per cent on GDP.pdf, which evaluates expert comprehension of difficult PDF documents, compared with 22.0 per cent for Gemini 3.6 Flash. In the company’s Harvey LAB-AA evaluation of complex legal workflows, 3.7 Flash reaches 90.7 per cent.
The significance is practical rather than merely multimodal. A business agent might need to read a contract, inspect a chart, analyse a spreadsheet exported as a PDF, listen to recorded material and update a software system as part of one objective. Using a single model capable of reasoning across those formats can simplify the surrounding architecture, provided its performance is good enough for the task.
This is one area where Google’s historical strengths matter. Its AI strategy sits inside a company that already controls enormous ecosystems of documents, email, search, video, maps, browsers, mobile devices and cloud software. A sufficiently capable multimodal agent has far more opportunities to become useful when the company deploying it also owns many of the environments in which that agent can act.
Gemini Spark shows what Google ultimately wants Flash to become
Gemini 3.7 Flash is being rolled into Gemini Spark, Google’s personal AI agent for Google AI Pro and Ultra subscribers in more than 160 countries. Google describes Spark as an agent that can continue working on behalf of a user, using Workspace tools to consolidate files, draft emails and update status documents across more complicated workflows.
This is strategically more important than it might initially appear. Google does not merely need developers to select Gemini inside an API console. It has the opportunity to insert agentic AI into Gmail, Drive, Docs and other software that users already depend upon every day. Reuters also notes that Gemini 3.7 Flash is being positioned as a lower-cost model for businesses building autonomous systems capable of planning tasks, operating software tools and completing workflows with less human intervention.
The distribution advantage is formidable. OpenAI and Anthropic can build extraordinarily capable AI systems, but Google already owns much of the digital environment in which knowledge work happens. If Gemini can reliably move from interpreting that environment to acting within it, the contest becomes about something larger than model intelligence.
It becomes a contest over who owns the working layer of AI.
Google Antigravity turns the same idea towards developers
For developers, one of the primary destinations for Gemini 3.7 Flash is Google Antigravity, the company's agentic development platform. Google presents the model as particularly suited to complex workflows involving coding, multimodal understanding, external tools and longer-running objectives. Demonstrations include agents coordinating across mobile development environments, analysing datasets distributed across cloud systems and modernising websites while checking their performance against external measurements.
These demonstrations should not be confused with neutral evidence of real-world reliability. They are carefully selected examples created to show the model at its best. Nevertheless, they demonstrate what Google now considers a normal use case for its Flash line. The object is no longer simply to help a programmer write a line of code faster. It is to give an AI system an objective and allow that system to coordinate much of the work required to achieve it.
That is a fundamentally more demanding role for the underlying model, and it explains why modest-seeming differences in planning, tool selection and first-pass accuracy have become commercially important.
How much does Gemini 3.7 Flash cost?
Gemini 3.7 Flash is available until 31 December 2026 at an introductory price of $0.75 per million input tokens and $3.75 per million output tokens. From 1 January 2027, Google says standard pricing will rise to $1.50 per million input tokens and $7.50 per million output tokens.
An important detail is easily lost in Google's launch messaging. The company says the introductory rate represents half the original cost of Gemini 3.6 Flash, but Google's current model card lists 3.6 Flash at the same promotional $0.75/$3.75 price. In other words, developers choosing between the two models today should not assume that 3.7 Flash currently costs half as much as 3.6. The significant point is that Google is providing the newer, substantially stronger model at the same discounted price during the promotional period.
For anyone modelling production costs beyond 2026, the future price matters. A project that appears economical at the introductory rate should be tested against the January pricing rather than built around a temporary discount.
That distinction is exactly the sort of detail businesses need when AI moves from experimentation to infrastructure.
Is Gemini 3.7 Flash better than Gemini 3.6 Flash?
On Google’s published evaluations, the answer is overwhelmingly yes. Gemini 3.7 Flash improves substantially on its predecessor across software engineering, agentic tool use, enterprise automation, knowledge work, PDF comprehension, computer use and several scientific-reasoning evaluations. The Artificial Analysis composite score listed by Google rises from 52 for Gemini 3.6 Flash to 56 for 3.7 Flash.
There are exceptions. The model is marginally lower on the no-tools version of CharXiv in Google's table, for example, demonstrating why it is unhelpful to describe any model upgrade as uniformly better at everything. But for the tasks around which Google is explicitly positioning the release, coding, agents and enterprise workflows, the improvement is substantial enough to make 3.7 the obvious model to evaluate before its predecessor.
The speed with which the update arrived may be equally notable. Three weeks between 3.6 and 3.7 suggests Google increasingly treats model development as a continuous product optimisation process rather than a sequence of large annual launches. That may make version numbers feel increasingly temporary, but it also raises the competitive tempo for everyone else.
What are the limitations of Gemini 3.7 Flash?
Gemini 3.7 Flash remains a foundation model and retains the familiar limitations of the category. Google explicitly states that it can hallucinate and may occasionally experience slowness or timeouts. Its stated knowledge cutoff is March 2026, although Google warns that users may encounter domains where effective knowledge extends only to January 2025. Applications dependent on current information therefore need external grounding or tools rather than assuming the model knows what has happened recently.
These weaknesses matter more in an agent than they do in a conversational assistant. A false statement produced in an ordinary chat can be corrected by the user. An incorrect premise produced midway through a multi-step automated workflow can influence what the system does next. Stronger agent models reduce this risk; they do not remove the need for permissions, logging, validation and appropriate human control.
Google’s safety evaluations also reflect the increasing capability of the model. DeepMind says Gemini 3.7 Flash reached an alert threshold in part of its cybersecurity assessment but remained below the relevant critical capability level. Its broader frontier assessments did not find that the model had reached Google's tracked critical thresholds.
The fact that a model marketed partly on low cost and high volume now requires frontier-safety evaluation is itself revealing. “Workhorse” no longer means unsophisticated.
What about Gemini 3.5 Pro?
Gemini 3.7 Flash arrives while a different Google model remains conspicuously absent. Reuters reported at launch that Google had still provided no release date for Gemini 3.5 Pro, its anticipated premium model, despite saying in July that it was being tested with partners and would arrive soon. The delay has drawn attention because investors and developers are watching Google's flagship systems as a measure of whether DeepMind can keep pace with OpenAI and Anthropic at the highest end of the market.
That context could make Gemini 3.7 Flash look like a distraction from a delayed flagship. It may be more useful to regard the two products as answering different questions. Pro is expected to demonstrate how far Google can push model capability. Flash is demonstrating how much of that capability can be made economical enough for widespread deployment.
If AI agents become embedded throughout companies, the latter could prove just as commercially important.
The AI industry has spent enormous energy determining who possesses the smartest model. Businesses may ultimately spend far more determining which model produces the best economics.
The workhorse may matter more than the show horse
Gemini 3.7 Flash arrives at an interesting stage in the development of generative AI. Frontier capability is still improving, but the gap between the most powerful models and the models considered economical enough for high-volume use is narrowing. Intelligence that was once expensive is being compressed into cheaper tiers at remarkable speed.
That changes what competition looks like.
A business running an agent thousands of times does not necessarily need the world's cleverest model sitting behind every action. It needs enough intelligence to complete the task reliably, enough speed to avoid turning automation into delay, and a low enough cost that using the system repeatedly still produces a positive economic outcome. If a model succeeds more often on the first attempt, the saving can extend beyond tokens to employee supervision, tool calls, compute and time.
Gemini 3.7 Flash is Google's clearest expression of that idea so far. Its own data shows a model only one point behind GPT-5.6 Terra on the Artificial Analysis Intelligence Index while carrying a much lower introductory token price, yet the details reveal a more complicated and useful picture: Terra remains stronger on some long-horizon engineering tasks, Claude retains strengths in knowledge work, and Gemini leads on several production-code, web-development and automation evaluations.
There is no single winner because that is increasingly the wrong question.
The next phase of AI will be shaped by models performing millions of individual pieces of useful work inside software, companies and eventually physical systems. Under those conditions, capability has to be considered alongside latency, reliability, token consumption, tool usage, human intervention and the cost of failure.
Google calls Gemini 3.7 Flash a workhorse because it wants the model to become infrastructure rather than spectacle. That may prove to be the more consequential ambition. The frontier AI race asks how intelligent a machine can become. The workhorse race asks something businesses are likely to care about far more:
How much useful work can that intelligence actually do for the money?
Frequently Asked Questions
What is Gemini 3.7 Flash?
Gemini 3.7 Flash is a multimodal reasoning model released by Google DeepMind on 13 August 2026. It is the successor to Gemini 3.6 Flash and is primarily designed for coding, AI agents, web development, knowledge work and enterprise workflows. It supports text, image, audio and video inputs, a context window of up to one million tokens and text outputs of up to 64,000 tokens.
When was Gemini 3.7 Flash released?
Google released Gemini 3.7 Flash on 13 August 2026, three weeks after the launch of Gemini 3.6 Flash.
How much does Gemini 3.7 Flash cost?
Until 31 December 2026, Gemini 3.7 Flash costs $0.75 per million input tokens and $3.75 per million output tokens under Google's introductory pricing. From 1 January 2027, the published rates rise to $1.50 per million input tokens and $7.50 per million output tokens.
Is Gemini 3.7 Flash better than Gemini 3.6 Flash?
Google's published benchmark results show significant improvements across many of the model's intended workloads. FrontierCode increases from 34.4 to 43.6 per cent, DeepSWE v1.1 from 48.6 to 65.3 per cent, Terminal-bench 2.1 from 78.0 to 85.8 per cent and AutomationBench from 17.0 to 30.4 per cent. Individual results vary, but the evidence published by Google indicates a substantial overall upgrade.
Is Gemini 3.7 Flash better than GPT-5.6 Terra?
Neither model leads every evaluation. Google's model card gives GPT-5.6 Terra higher scores on DeepSWE v1.1, Terminal-bench 2.1, Terminal-bench 3.0 and OSWorld-2.0, while Gemini 3.7 Flash leads on FrontierCode 1.1, Code Arena and AutomationBench. Gemini's introductory API price is considerably lower. The best choice therefore depends on workload and total cost per successful task rather than a single benchmark.
Is Gemini 3.7 Flash better than Claude Sonnet 5?
Google reports stronger Gemini 3.7 Flash results on several coding and automation evaluations, including FrontierCode, DeepSWE, Terminal-bench and AutomationBench. Claude Sonnet 5 leads Gemini on GDPVal-AA v2 knowledge work and Agent's Last Exam in Google's comparison. Claude is also substantially more expensive per token at the prices listed in Google's August 2026 model card.
What is Gemini 3.7 Flash best used for?
Google positions Gemini 3.7 Flash particularly for coding agents, software engineering, autonomous workflows, web development, enterprise automation and knowledge-intensive tasks. Its multimodal support also makes it suitable for workflows combining documents, images, audio and video.
What is the context window of Gemini 3.7 Flash?
Gemini 3.7 Flash supports a context window of up to one million tokens, with a maximum text output of 64,000 tokens.
Is Gemini 3.7 Flash multimodal?
Yes. Gemini 3.7 Flash accepts text, image, audio and video inputs. Its output is currently text.
Where can developers access Gemini 3.7 Flash?
Google distributes Gemini 3.7 Flash through the Gemini API, Google AI Studio, Google Antigravity, the Gemini Enterprise Agent Platform and Gemini Enterprise. It also powers Gemini Spark for eligible Google AI Pro and Ultra subscribers.
Does Gemini 3.7 Flash have a knowledge cutoff?
Google lists March 2026 as the model's knowledge cutoff, while warning that some domains may effectively have knowledge only up to January 2025. Applications requiring current information should therefore use appropriate grounding, search or external tools.
Is Gemini 3.7 Flash Google's most powerful model?
Google describes Gemini 3.7 Flash as its most intelligent workhorse model rather than its ultimate premium model. The company has separately been developing Gemini 3.5 Pro, although Reuters reported at the time of the Flash launch that Google had not provided a new release date for that system.