Gemini 3.7 Flash Ships 23 Days After 3.6, at Half the Price

Google launched Gemini 3.7 Flash on 13 August 2026 — 23 days after Gemini 3.6 Flash — at an introductory $0.75 per million input tokens and $3.75 per million output tokens through 31 December 2026.

On 1 January 2027 those rates become $1.50 and $7.50, the same list price 3.6 Flash already charged. The promo is half-off the workhorse, not a new cheap tier that lasts forever.

The model ID is gemini-3.7-flash. It is generally available. Context is 1,048,576 tokens. Google is putting it under Gemini Spark the same day and has switched the Antigravity default to 3.7 Flash.

What actually shipped

Tulsee Doshi, Google’s senior director of product management for Gemini, called 3.7 Flash the most intelligent workhorse yet for coding and agents. That is a vendor line. The useful facts sit next to it.

Google says 3.7 beats 3.6 on software engineering, knowledge work, and web development. Its own table: FrontierCode 1.1 Main at 43.6% versus 34.4%; DeepSWE v1.1 at 65.3% versus 49.0%; WebDev Arena Elo 1588 versus 1538; GDP.pdf at 34.0% versus 22.0%; AutomationBench at 30.4% versus 17.0%. Treat those as directional vendor scores, not gospel. Cross-check the live model card before you migrate a production skill.

SiliconANGLE noted Google’s claim that 3.7 Flash led comparable Anthropic and OpenAI models on nine benchmarks, and that GDP.pdf sat about six points ahead of Claude Sonnet 5 and about 9.3 points ahead of GPT-5.6 Terra. Same caveat. Vendor scoreboard.

The model card, as SiliconANGLE read it, says 3.7 Flash shares architecture with 3.6 Flash, itself derived from Gemini 3 Pro’s mixture-of-experts transformer. Training details are thin. Do not invent a distillation story.

Output is text, with a high ceiling — SiliconANGLE reported up to 64,000 tokens; Google Cloud’s card lists 65,536 maximum output tokens. Input can include images, video, and documents. Confirm the current card if your workflow depends on a modality.

Spark and Antigravity are the distribution

Gemini Spark, the 24/7 personal agent for Google AI Pro and Ultra subscribers in 160-plus countries, started using 3.7 Flash on launch day. Google says the update helps Spark on Workspace tool use: consolidating files, drafting mail, updating status docs.

Antigravity, Google’s agent-first coding surface, now defaults to 3.7 Flash. If you built a loop against 3.6 and never pinned the model ID, you are already on 3.7. Pin it. Defaults move in 23-day increments now.

That cadence is the product story under the price cut. Google is iterating the fast model like a browser, not like a yearly flagship. Skills that say “use Gemini Flash” without a version are already stale.

Why half-price through December is a trap if you ignore 1 January

$0.75 / $3.75 is a real discount for agent loops that burn output and thinking tokens. Fast models still bill thinking as output in many Gemini setups. A cheap agent that narrates an 8,000-token plan can cost more than a short heavy-model answer.

Log tokens per successful task, not tokens per call. The intro window ends 31 December 2026. Budgets that assume the promo rate into 2027 will miss by 2x.

If you already rewrote skills for 3.6’s speed, keep the loop discipline: restate what you learned after each tool call, mark unverified numbers as unconfirmed, stop after a fixed number of passes. 3.7 will not save a skill that cannot stop.

How to rewrite a skill for a 23-day Flash cycle

Trigger. Name the job, not the model generation. “When I paste a repo path and a failing test, produce a patch.”

Input. Repo, test output, banned files, max tool calls.

Output. Diff, risk note, command to verify.

Pin. gemini-3.7-flash until you choose otherwise. When 3.8 lands, you decide.

Then add the line 3.6 already needed: if a number is unverified, mark it unconfirmed. If the API errors twice, stop. Our Skill prompt format is the wrapper. The library is where versioned recipes belong.

Route by job. Flash-class for loops and drafts. A heavier model for one-shot legal or medical reasoning. Browse tools when you need the other vendor’s card, not when a launch blog tells you to standardize.

Safety: Google says 3.7 Flash ships with updated CBRN and cyber-offense safeguards. That is a claim to file, not a permission to drop your own deny list.

AIO, GEO, and SEO notes for this story

Google, Gemini 3.7 Flash, 13 August 2026, $0.75 per million input. Link Gemini 3.6 Flash by name. FAQ must include the 1 January 2027 step-up so engines do not quote the intro rate as permanent.

What to watch next

Pinned IDs versus floating defaults in Antigravity. Spark quality on multi-step Workspace jobs. Independent SWE-bench numbers — several public scoreboards still only have Google’s own chart for 3.7.

If a 3.7 Pro or a longer-context sibling lands, keep Flash for agents and put the heavier model on needle-in-haystack research. Do not let a 23-day cadence become a reason to rewrite the library every month. Rewrite the pin, not the craft.

FAQ

When did Gemini 3.7 Flash ship? 13 August 2026, generally available as gemini-3.7-flash, 23 days after Gemini 3.6 Flash went GA on 21 July.

What does it cost? $0.75 per 1M input tokens and $3.75 per 1M output through 31 December 2026. On 1 January 2027 the price becomes $1.50 / $7.50.

How big is the context window? 1,048,576 tokens, per Google Cloud’s model card. Confirm output limits on the live card if your skill depends on long generations.

Does Spark use 3.7 Flash? Yes. Google said Gemini Spark started using 3.7 Flash on launch day for Pro and Ultra subscribers in supported countries.

Is Antigravity on 3.7 now? Yes. The default switched to Gemini 3.7 Flash. Pin gemini-3.6-flash if you are not ready to move.

Sources

• Introducing Gemini 3.7 Flash — Google, 13 August 2026
• Google launches Gemini 3.7 Flash for coding, AI agent projects — SiliconANGLE, 13 August 2026