A bright AI route accelerating through a dark production gateway

Release Watch

Gemini 3.7 Flash Release: What Shipped on August 13, 2026

Q: What is the Gemini 3.7 Flash model ID?

gemini-3.7-flash, with dots, on both the official Gemini API and EvoLink. The dashed form only appears in page URLs.

Q: Which parameters break when migrating from 3.6?

Custom temperature, topp, topk, and candidatecount are deprecated, and numeric thinkingbudget gives way to the thinking_level string.

EvoLink Team

Product Team

August 14, 2026

7 min read

The Direct Answer: What Is Confirmed Today?

Gemini 3.7 Flash is generally available. Google released it on August 13, 2026, describing it as "our most intelligent workhorse model yet for coding and agents." These facts are confirmed against Google's official documentation as of August 14, 2026:

Model ID: gemini-3.7-flash — a stable release, no preview suffix.
Context window: 1,048,576 tokens; max output: 65,536 tokens.
Input: text, image, video, audio, PDF. Output: text only.
Official introductory pricing: $0.75 input / $3.75 output per 1M tokens through December 31, 2026, then $1.50 / $7.50 from January 1, 2027. Thinking tokens are billed at the output rate.
Same price as Gemini 3.6 Flash — the upgrade case is capability and token efficiency, not a lower rate card.
It is not a new pretrained model: Google's model card describes it as algorithmic improvements on the Gemini 3.6 Flash reasoning foundation, shipped three weeks after 3.6.

For pricing on EvoLink, code examples, and production guidance, see the Gemini 3.7 Flash API page.

What Changed vs Gemini 3.6 Flash

Google's published benchmark deltas against Gemini 3.6 Flash (all numbers are Google's own, from the official model card — treat them as vendor-reported until independent replications land):

Benchmark	Gemini 3.7 Flash	Gemini 3.6 Flash
FrontierCode 1.1 (production coding)	43.6%	34.4%
DeepSWE v1.1 (long-horizon SWE)	65.3%	~49% (baseline figures vary by source)
Terminal-bench 2.1 (terminal execution)	85.8%	78.0%
AutomationBench (agentic)	30.4%	17.0%
WebDev Arena	1588 Elo	1538 Elo
GDM-MRCR v2 128k (long context)	97.0%	91.8%

If you are weighing whether to move an existing 3.6 workload, the decision guide is here: Gemini 3.7 Flash vs Gemini 3.6 Flash.

The launch narrative centers on agentic reliability: fewer failed agent loops, better recovery from roadblocks, and stronger design-to-code and codebase-auditing behavior. Independent launch-day coverage broadly repeats these numbers; the only independent measurements so far come from Artificial Analysis, which recorded top-tier output speed alongside a higher hallucination rate than 3.6 Flash — worth verifying on your own tasks before promoting the model to production.

Thinking Levels: `minimal` Is Gone

Gemini 3.7 Flash introduces tunable thinking levels — low, medium (default), and high. Two consequences for existing Flash users:

minimal no longer exists. Sending it to the Gemini API returns an error, so low is now the cheapest reasoning floor. High-volume classification and extraction pipelines that relied on minimal should benchmark Gemini 3.5 Flash-Lite as the low-cost route instead.
The lowest available reasoning level is low, and thinking tokens bill at the output rate — so output cost depends on the thinking level you select, not just answer length.

On EvoLink, requests that still send reasoning_effort: "none" or "minimal" are set to be downgraded automatically to low instead of failing, so migrated code keeps working while you update it.

Breaking API Changes and the Migration Checklist

Migrating from Gemini 3.6 Flash is a model-ID swap plus parameter hygiene:

Change the model to gemini-3.7-flash (the dashed form is a page URL, not an API parameter).
Remove custom temperature, top_p, top_k, and candidate_count — they are deprecated on 3.7 Flash.
Replace numeric thinking_budget values with the string thinking_level parameter.
Replace minimal with low, or route that traffic to a Lite-tier model.
Re-run representative prompts: same rate card does not mean same token consumption, and some launch-day reports observed higher per-task token usage than 3.6.

Release Does Not Mean API Availability Everywhere

Gemini 3.7 Flash model route being verified across three API channels

GA at Google does not mean the model is live on every channel you use. Status as of August 14, 2026:

Official channels: Gemini API / AI Studio, Android Studio, Google Antigravity, Gemini Enterprise Agent Platform, and the Gemini app's Spark experience.
Aggregators: OpenRouter listed it on day zero; kie.ai followed within a day. Most other aggregation platforms focus on open-weight models and are unlikely to carry it at all — this is a closed, API-only model with no weights to self-host.
EvoLink: gemini-3.7-flash is available on both the OpenAI-compatible /v1/chat/completions endpoint and the native Gemini API. Verify the route on your own key before switching production traffic:

curl -s https://api.evolink.ai/v1/models \
  -H "Authorization: Bearer $EVOLINK_API_KEY" | grep gemini-3.7-flash

gemini-3.6-flash remains available at the same price on the same endpoints, so you can stage the migration and roll back instantly if a workload regresses.

Pricing: The Introductory Window Matters

The headline $0.75 / $3.75 is an introductory price with an expiry date. Google's published schedule:

	Through Dec 31, 2026	From Jan 1, 2027
Input / 1M tokens	$0.75	$1.50
Output (incl. thinking) / 1M	$3.75	$7.50
Cache read / 1M	$0.075	$0.15

Cache reads are the lever that matters for agent workloads: at $0.075 per 1M tokens a cache hit costs one tenth of fresh input, so stable system prompts, repository instructions, and tool schemas pay for themselves whenever the prefix stays identical across calls.

Batch runs at half price on both schedules. If you are building cost models for 2027 budgets, use the standard rates, not the launch numbers. Current EvoLink rates for the model are shown live on the Gemini 3.7 Flash pricing section.

Signals Worth Monitoring

Independent benchmark replications — only one independent tracker has published measurements so far.
Hallucination-rate follow-ups — the launch-day regression signal needs more data.
Your own token-per-task deltas — the efficiency claim is workload-dependent.
Aggregator rollout — whether any platform beyond OpenRouter and kie.ai lists the model.
A future Pro release — Google's newest Pro-line model is still gemini-3.1-pro-preview.
December 31, 2026 — the introductory price expiry.

FAQ

When was Gemini 3.7 Flash released?

August 13, 2026, as a generally available (GA) stable model — confirmed in Google's official API changelog and model card.

What is the Gemini 3.7 Flash model ID?

gemini-3.7-flash, with dots, on both the official Gemini API and EvoLink. The dashed form only appears in page URLs.

Is Gemini 3.7 Flash a new model or an update?

Google's model card describes it as algorithmic improvements on the Gemini 3.6 Flash reasoning foundation rather than a new pretrained model.

Does Gemini 3.7 Flash cost more than 3.6 Flash?

No — the two share the same rate card, including the introductory window through December 31, 2026.

What happened to the minimal thinking level?

It was removed. Levels are now low, medium, and high; sending minimal to the Gemini API errors. EvoLink is set to downgrade legacy minimal requests to low automatically.

Which parameters break when migrating from 3.6?

Custom temperature, top_p, top_k, and candidate_count are deprecated, and numeric thinking_budget gives way to the thinking_level string.

Is there a Gemini 3.7 Pro?

No. No 3.5 or 3.7 Pro has been announced; the Pro line's newest model remains gemini-3.1-pro-preview.

Where can I use the Gemini 3.7 Flash API today?

Google's Gemini API / AI Studio, Antigravity, and the Enterprise Agent Platform; OpenRouter and kie.ai among aggregators; and EvoLink per the rollout note above — verify the ID appears in your key's /v1/models list.

All Posts

#Gemini 3.7 Flash #Gemini API #Google AI #Flash models #model release