Gemini 3.7 Flash: Benchmarks, Pricing, and API Guide
Twenty-three days separate Gemini 3.6 Flash and Gemini 3.7 Flash. That is barely enough time for most engineering teams to finish a serious migration test. Google used it to ship a materially stronger agent model, make it generally available on day one, switch Antigravity's default model, and set an introductory $0.75/$3.75 rate for both 3.7 and 3.6 through December 31.
The benchmark jump is real: Artificial Analysis scores 3.7 at roughly 56 versus roughly 52 for 3.6 while measuring about 340 output tokens per second, and Google reports larger gains on long engineering loops, computer use, automation, and PDFs. But the easy headline—“a faster, smarter model for half the price”—leaves out three details.
First, Google's model card says 3.7 is based on 3.6. This is an algorithmic reasoning and execution upgrade, not a disclosed new foundation architecture. Second, the $0.75 input and $3.75 output rates expire after December 31, 2026; every listed inference rate doubles on January 1. Third, Google also applied the promotion to 3.6 Flash. The two models have the same unit price, so the cheaper completed task is determined by token use, retries, tool calls, and acceptance rate—not the version number.
The useful way to think about this launch is as an agent-runtime release: Google improved the worker, kept the context and tool surface, preserved its provider-defined API rules, and moved a managed agent to a new default model beneath an unchanged outer ID. Model routing, version logging, and canary evaluation are now product architecture.
Google's official Gemini 3.7 Flash announcement positions it as the new workhorse for coding and agents. Unlike many Gemini launches, this is not a preview endpoint waiting for a production label. The dedicated Gemini API specification lists gemini-3.7-flash as stable and generally available.
Developers can use it through Google AI Studio and the Gemini API, Google Antigravity, Android Studio, Gemini Enterprise Agent Platform, and the Gemini Enterprise app. For consumers, Google says Gemini Spark—available to Google AI Pro and Ultra subscribers in more than 160 supported countries—began using 3.7 that day. This does not mean every free Gemini conversation switched models.
The model accepts text, images, audio, video, and PDFs and produces text. It supports 1,048,576 input tokens and 65,536 output tokens, function calling, structured output, code execution, File Search, URL context, Search and Maps grounding, caching, and Computer Use in preview. It does not support native image or audio generation, and it does not support the Live API.
Thinking is adjustable at three levels:
low for latency-sensitive work such as chat, incident triage, drafting, and quick analysis;medium, the default, for most coding and multi-step agent tasks;high for difficult reasoning, hard coding, math, and longer tool loops.minimal is not a quiet alias for low; the model page says it returns an error. Higher thinking can improve completion, but reasoning tokens are billed as output. Treat the level as part of the deployed SKU, not a cosmetic prompt option.
The lineage matters. Google's model card says 3.7 Flash is based on 3.6 and refers back to the prior card for architecture, training-data processing, hardware, and software. Google describes algorithmic improvements to the reasoning foundation but discloses no parameter count or architecture change. The consequential result is that a derivative update produced a substantial agent uplift in 23 days without a publicly described new base-model training run.
This article is a direct update to RohitAI's Gemini 3.6 Flash analysis. That release shifted control from sampling knobs toward thinking, tools, interaction state, and managed routing. Gemini 3.7 does not reverse the shift. It shows how quickly Google can now change the worker inside that system.
Google's launch table mixes public benchmarks, private evaluations, live leaderboards, and an independent composite. The pattern is still clear: 3.7 improves most when it must stay oriented across several actions, recover from roadblocks, manipulate a UI or repository, and finish an artifact.
Google reports FrontierCode 1.1 rising from 34.4% to 43.6%, WebDev Arena moving from 1538 to 1588 Elo, and Harvey's legal evaluation reaching 90.7%. The public Cognition FrontierCode page did not yet expose a 3.7 row when checked. Artificial Analysis independently measured Terminal-Bench 2.1 at 85.77%.
It is not a sweep. In Google's launch table, GPT-5.6 Terra remains ahead on DeepSWE and both Terminal-Bench versions, while Sonnet 5, Terra, and Muse Spark 1.2 lead GDPVal-AA. Agents' Last Exam reaches only 26.3%. CharXiv chart reasoning slips slightly with and without tools, and long-video understanding moves only 1.2 points.
That distribution is more useful than a victory count. Google appears to have improved the model's ability to execute, recover, and finish more than its static perception across every modality. A product dominated by repository work, browser actions, design parity, or document workflows has a stronger upgrade case than one dominated by chart interpretation or low-latency single-turn answers.
Google publishes a model evaluation methodology document alongside the table. It is not one controlled head-to-head: Gemini results mix Google-computed runs and public leaderboards; competitor figures may come from provider reports or leaderboards; AutomationBench is private; LVBench uses different frame limits across providers; OSWorld was run before an August 8 patch; and Google made minor compatibility changes to two Terminal-Bench 3.0 tasks. Treat the values as a release map, then test the actual harness, tools, prompts, and failure costs in your system.
DeepSWE's independent v1.1 board largely validates the direction of Google's claim. Under the same mini-swe-agent harness across 113 tasks, 91 repositories, and five languages, Gemini 3.7 Flash High scored 65% ±2 at $2.18 per task, versus 47% ±4 and $4.42 for 3.6 High. But it was not the clear winner: Gemini Luna Max scored 67% ±4 at $0.61, Kimi K3 69% ±5, GPT-5.6 Sol 73% ±3, and Claude Opus 5 74% ±4. The confidence intervals around 3.7 and Luna overlap. The honest conclusion is a major Gemini upgrade, not universal coding leadership. DeepSWE v1.1 leaderboard
The current Artificial Analysis 3.7 Flash page provides an independent snapshot of quality and a measured view of Google's API serving. When checked on launch day, the high-thinking model scored 56.03 on Intelligence Index v4.1.1 and 45.10 on the Agentic Index. Gemini 3.6 Flash scored 51.58 and 40.51 respectively.
Serving improved even more sharply. Artificial Analysis measured 3.7 at about 340.1 output tokens per second, up from roughly 220.8 for 3.6—a 54% increase. On its standard latency workload, median time to first chunk fell from about 18.2 seconds to 9.8 seconds, while the calculated time to return 500 tokens fell from 20.5 seconds to 11.3 seconds.
Flash is exceptionally fast once generation starts, but a reasoning request can still wait nearly ten seconds for a visible chunk. That trade works for background agents; autocomplete, live support, voice, and instant-feedback interfaces must still optimize first-chunk latency.
Long context changes the picture. With a 100,000-token input, Artificial Analysis measured 395.4 output tokens per second but a 16.1-second first-chunk delay and 17.38-second end-to-end time. Gemini 3.6 completed the equivalent workload in 17.35 seconds—effectively unchanged. A one-million-token window is capacity, not a latency guarantee: prompt processing can dominate even when decoding becomes much faster. Launch-day serving also varied widely, from roughly 226 to 743 tokens per second at the fifth and 95th percentiles, while first-chunk latency ranged from 5.8 to 49.5 seconds.
Artificial Analysis also tested Medium thinking. High scored 56.03 on Intelligence Index versus 53.42 for Medium—a 4.9% advantage—but their Agentic Index scores were effectively identical at 45.10.
| Launch-day measurement | Medium | High | High's change |
|---|---|---|---|
| Agentic Index | 45.10 | 45.10 | Effectively flat |
| Cost per weighted task | $0.263 | $0.402 | +53% |
| Time to first chunk | 4.22 sec | 9.83 sec | +133% |
| End-to-end latency | 6.05 sec | 11.30 sec | +87% |
| Output tokens per task | 23,311 | 36,847 | +58% |
High materially won Humanity's Last Exam, 47.87% to 38.97%, and Terminal-Bench 2.1, 85.77% to 78.28%. Medium narrowly won SciCode, long-context recall, and tau3 Banking. That makes Medium the sensible default for routine tool loops, extraction, high-volume coding, and parallel workers. High earns its extra latency and measured task cost when ambiguity, difficult planning, or failure cost justifies it. “Use High everywhere” is not supported by the independent agent score.
Gemini's independent position is strong but specialized. Its 56.03 Intelligence score sits below GPT-5.6 Sol and Kimi K3 and around Terra, Muse Spark, and Sonnet-class results. More importantly, its 45.10 Agentic Index trails Luna, DeepSeek V4 Pro, Muse Spark, Terra, Kimi, and Sonnet in the same launch-day snapshot. Gemini's case is exceptional throughput plus multimodality, one-million-token context, Google tools, and distribution—not an undisputed claim to the best agent model.
Capability and calibration did not move together. Artificial Analysis reports 55.32% accuracy and a 64.53% hallucination rate on its specialized unknown-answer test. That does not mean 64.5% of ordinary Gemini answers are false. It means the model frequently answered when that evaluation expected abstention—and the figure worsened from roughly 55.6% for 3.6. Better knowledge coverage does not remove the need for grounding, citations, and explicit abstention tests.
There is a second cost signal in the same evaluation. On a weighted-task basis, 3.7 emitted 36,847 output tokens versus roughly 25,700 for 3.6, an increase of about 43%. Reasoning tokens barely changed—14,048 versus roughly 13,800—but visible answer tokens rose about 91%, from roughly 11,900 to 22,799. The new model is spending much of its price reduction on longer answers, not dramatically more hidden reasoning.
Through December 31, 2026, Gemini 3.7 Flash costs $0.75 per million input tokens and $3.75 per million output tokens, including thinking tokens. Google applied the same promotional rates to 3.6 Flash. On January 1, 2027, those Standard prices double to $1.50 input and $7.50 output.
The promotion spans Google's serving modes, not only Standard inference:
| Paid tier | Through Dec. 31: input / output | From Jan. 1: input / output | Cache read now → later |
|---|---|---|---|
| Standard | $0.75 / $3.75 | $1.50 / $7.50 | $0.075 → $0.15 |
| Batch | $0.375 / $1.875 | $0.75 / $3.75 | $0.0375 → $0.075 |
| Flex | $0.375 / $1.875 | $0.75 / $3.75 | $0.0375 → $0.075 |
| Priority | $1.35 / $6.75 | $2.70 / $13.50 | $0.135 → $0.27 |
Cached-context storage rises from $0.50 to $1.00 per million tokens per hour. Batch suits asynchronous bulk work; Flex trades predictable latency for lower-cost capacity; Priority costs more for latency-sensitive workloads.
Grounding has a separate meter. Google lists 5,000 free Search requests per month shared across all Gemini 3.x models, then $14 per 1,000 requests, and warns that one Gemini request can fan out into multiple individually billable Search queries. Maps lists 5,000 prompts per month, then $14 per 1,000 search queries. Include those tool-side charges in agent economics.
Google's pricing page says free-tier content may be used to improve its products, while paid-tier content is not. The Gemini API Additional Terms add that unpaid content may be reviewed by humans; paid prompts and responses may still be logged temporarily for abuse prevention, security, and legal compliance; and Search or Maps grounding data may be stored for 30 days. “Not used for training” is therefore not the same as zero retention.
Now for the non-obvious calculation. Artificial Analysis estimates Gemini 3.7 Flash High at $0.402 per weighted task under the promotional tariff. Gemini 3.6 High's measured task cost was roughly $0.557 under its original $1.50/$7.50 rate. Mechanically halving only that historical run's token prices produces a counterfactual cost near $0.279 per task—not a new evaluation.
Against that like-for-like promotional counterfactual, 3.7 costs about 44% more because it emits substantially more billable output. It may still be cheaper per accepted outcome if the extra work eliminates retries or reviewer intervention. Identical token rates do not produce identical task costs.
This is the same scheduling problem RohitAI identified in the DeepSeek V4 Pro 0813 analysis, expressed differently. DeepSeek makes the clock explicit with peak and off-peak windows. Google uses Batch, Flex, Priority, caching, thinking level, and a promotional expiry. Either way, a production agent needs a router and a cost model, not one hard-coded endpoint.
The simplest raw API example uses Google's Interactions API. The stable model string is gemini-3.7-flash, and medium thinking is the default.
from google import genai
client = genai.Client()
response = client.interactions.create(
model="gemini-3.7-flash",
input=(
"Audit this payment workflow for retry races, propose a safe locking "
"strategy, and return the patch plus tests."
),
generation_config={"thinking_level": "medium"},
)
print(response.output_text)
The JavaScript equivalent is similarly small:
import { GoogleGenAI } from "@google/genai";
const client = new GoogleGenAI({});
const response = await client.interactions.create({
model: "gemini-3.7-flash",
input:
"Audit this payment workflow for retry races, propose a safe locking strategy, and return the patch plus tests.",
generation_config: { thinking_level: "medium" },
});
console.log(response.output_text);
The short request hides a migration contract. Google's 3.7 guide tells teams moving from older Gemini routes to remove temperature, top_p, and top_k. They were already unsupported by 3.6; leaving them in shared configuration creates only the appearance of control.
Other changes carry forward:
thinking_budget with thinking_level;candidate_count, which Gemini 3.x does not support;previous_interaction_id for multi-turn state;call_id and name in FunctionResponse objects when using generateContent;Malformed_Function_Call handling when a model emits prose before a tool call.Gemini 3.7 upgrades the model without returning the old sampler. Google steers applications toward thinking level, schemas, tools, interaction state, and managed-agent configuration. Multi-model layers should expose those provider-specific semantics instead of pretending one temperature slider maps across vendors.
Google's managed Antigravity agent still uses the outer identifier antigravity-preview-05-2026. On July 21, the company made 3.6 Flash the default beneath it. On August 13, the same guide made 3.7 Flash the new default. The underlying model can be configured through agent_config, but an application relying on the default changed without changing its agent string.
That is the second default move in 23 days.
antigravity-preview-05-2026The minimum useful production log is therefore larger than model or agent:
effective execution identity =
requested agent ID
+ resolved model and version
+ thinking level
+ enabled tools and permissions
+ agent configuration
+ context and compaction policy
+ SDK/client version
Store that manifest with prompts, tool traces, outputs, approvals, and tests. Canary raw gemini-3.7-flash separately from Antigravity: the managed route adds planning, tools, a remote environment, artifacts, and context management. One eval cannot validate both the checkpoint and harness.
Spark extends the lesson to consumer work: better reasoning can improve Workspace actions, but permissions, sources, execution policy, and approval boundaries still decide whether the result is safe.
At the same promotional token price as 3.6, the default recommendation is straightforward: test 3.7 now, but move with canary traffic rather than a blind string replacement. The benchmark and serving gains are large enough to justify the evaluation. The higher observed output volume and changed behavior are large enough to justify controls.
Start with repository repair, browser and computer use, design-to-code, PDF analysis, and workflows where 3.6 stalls or needs repeated correction. Measure completion and regressions before expanding traffic.
Hold a known-good route temporarily when deterministic output shape, latency, or a validated tool loop matters more than immediate benchmark gains. Pin it and define an exit date.
Classification, extraction, routing, translation, and reversible subagent work may not need 3.7. Compare accepted-task cost at the deployed thinking level instead of promoting every call.
Terra, Sonnet, a higher reasoning tier, or another specialist can still win on difficult software tasks, knowledge work, or costly failures. Flash should be the workhorse, not the only horse.
The right production metric is cost per accepted outcome. Record all input, cache reads and writes, reasoning, answer tokens, search queries, tool calls, retries, latency, reviewer corrections, and rollback events. RohitAI's GPT-5.6 Codex rate-card analysis reaches the same conclusion from another pricing system: agent economics live at the task and workflow layers.
The model card points to 3.6 for architecture, training data, processing, hardware, and software. The release remains proprietary, with no downloadable weights or local recipe.
Google gives a March 2026 cutoff but warns that some domains may behave as though knowledge ends in January 2025. Treat freshness as uneven. Search grounding adds current information alongside source-quality, cost, latency, and citation checks.
The card acknowledges hallucinations and occasional slowness or timeouts. Automated safety deltas versus 3.6 are small and mixed, including slight regressions in text safety, tone, and unjustified refusals, while human red-teaming found no egregious new concern. No tracked critical-capability threshold was reached, although cyber uplift and one CBRN uplift category triggered alert thresholds. Google says 3.7 ships updated safeguards targeting cyber offense and CBRN risk.
One disclosure is still pending: the model card says a dedicated Gemini 3.7 Frontier Safety Framework report will be published shortly. That statement may age quickly. Link the eventual report when it appears rather than treating today's absence as permanent.
For enterprise deployments, the practical safety boundary is the agent session. Computer Use and tool access can turn a plausible mistake into an action. Apply least privilege, isolate execution, restrict secrets and network egress, require approval for consequential writes, retain tamper-evident traces, and test rollback. Better benchmark scores do not substitute for those controls.
Three conclusions matter beyond this release.
Flash is becoming continuous delivery for model behavior. A derivative rose 4.45 points on Artificial Analysis and improved serving sharply in 23 days. Teams need model evals that run like integration tests, not quarterly procurement exercises.
“Flash” now describes an economic and throughput role. It generates large artifacts quickly, but the first chunk can take nearly ten seconds and output volume may exceed 3.6. Optimize total workflow performance, not the label.
Price is an adoption window. The temporary tariff gives builders time to move traffic and gather traces before January's doubling. Another Flash may arrive first, but procurement should budget the published future rate and architecture should assume continued model movement.
My near-term prediction is a three-level Gemini routing pattern:
Flash-Lite -> routine, parallel, reversible volume
3.7 Flash -> default coding, multimodal, and agent workhorse
Pro/other -> expensive uncertainty, final review, specialized failure
Inside the 3.7 route, thinking level, serving tier, cache state, and grounding turn model selection into job scheduling. Google can now improve the worker faster than many teams can re-certify its surrounding system.
Gemini 3.7 Flash is the first recent Flash update to deliver a clear capability jump and serving upgrade together. It is GA, fast once generation begins, broadly multimodal on input, strong on several coding and execution benchmarks, and temporarily inexpensive. For builders using 3.6, it deserves an immediate canary.
It does not deserve an unmeasured fleet-wide switch. Ship it as a routed dependency with an execution manifest, an eval set, and a rollback threshold.
The winning question is not “Is 3.7 better?” It is: Which tasks finish more reliably, how quickly, and at what fully loaded cost?
Gemini 3.7 Flash is Google's GA multimodal reasoning model for coding, agents, knowledge work, and high-throughput execution. It is based on 3.6 Flash.
Use gemini-3.7-flash. Google lists it as a stable GA endpoint rather than a preview ID.
Through December 31, Standard paid pricing is $0.75 input, $3.75 output, and $0.075 cached input per million tokens. On January 1, 2027, those rates double. Batch and Flex cost half of Standard; Priority costs more.
It is half of 3.6 Flash's original $1.50/$7.50 rate, but Google also applied the same temporary $0.75/$3.75 promotion to 3.6. The two models currently have the same unit price. Actual task cost depends on token use, tools, retries, and completion rate.
It clearly improves on 3.6 across several agent benchmarks, and Google's migration guide names 3.1 Pro as a route to move from. In Google's comparison, 3.7 beats Sonnet 5 and Terra on FrontierCode but trails Terra on DeepSWE and Terminal-Bench and trails both on GDPVal. There is no universal winner.
The API supports 1,048,576 input tokens and 65,536 output tokens. Test retrieval and agent memory at real production lengths.
Gemini 3.7 Flash supports low, medium, and high. Medium is the default. minimal is not supported and returns an error. Thinking tokens count toward billed output.
It accepts text, images, audio, video, and PDFs but returns text. It has no native image or audio generation and no Live API support.
It supports function calling, schemas, grounding, code execution, File Search, URL context, caching, and Computer Use in preview. Probe the exact deployed surface.
Google lists Gemini app availability and specifically says Gemini Spark uses 3.7 Flash for Google AI Pro and Ultra subscribers in supported countries. That does not establish that every free Gemini conversation uses the same model.
It is now the default model under the existing antigravity-preview-05-2026 agent ID. The underlying model can be configured through agent_config. Log the resolved checkpoint because the outer agent ID did not change.
No. It is a proprietary hosted model. Google has not released weights or a local self-hosting recipe.
Google states March 2026 but warns that some domains may remain limited to January 2025-era knowledge. Ground and verify current facts.
Run a canary now, especially for coding, computer use, PDF analysis, and complex tool loops. Migrate broadly only after comparing accepted-task cost, latency, output shape, tool reliability, regressions, and safety in your own workload. Keep a pinned 3.6 rollback route during the transition.