Gemini 3.7 Flash, 3.6 Flash & 3.5 Flash-Lite: Benchmarks, Pricing, Speed

Google just released Gemini 3.7 Flash, three weeks after 3.6 Flash. It scores 56 on the Intelligence Index (up from 50), runs at 340 tok/s, and costs $0.75/$3.75 per million tokens through the end of 2026. That's 50% cheaper than 3.6 Flash's regular pricing and smarter than GPT-5.6 Terra at a fraction of the cost.

The speed of iteration here is unusual. Three Flash updates in under a month (3.5, 3.6, 3.7), each with meaningful intelligence gains. Logan Kilpatrick credited "algorithmic improvements" from teams across Google DeepMind. Meanwhile, 3.5 Pro still hasn't shipped.

Update (August 13, 2026): Gemini 3.7 Flash launched with a 12% intelligence jump over 3.6 Flash (56 vs 50 on the Intelligence Index). Three effort levels (low/medium/high). Same intro pricing as 3.6 Flash ($0.75/$3.75). Available in the API, AI Studio, Antigravity, and Spark. All data below is updated to reflect 3.7 Flash. OpenRouter has it at google/gemini-3.7-flash with an additional 50% discount ($0.375/$1.875) during the intro period.

The Flash Family Side by Side

3.7 Flash (high) at Intelligence 56 is now the strongest Flash-tier model Google has released. It beats Grok 4.5 (54), matches GPT-5.6 Terra (55), and sits just one point below Kimi K3 (57). At $0.75/$3.75, it costs a fraction of any of those competitors.

The effort levels give you a useful tradeoff. 3.7 Flash (low) scores 51 on the Intelligence Index and costs $0.16 per task. That matches GPT-5.6 Luna's intelligence while being much faster (254 vs 150 tok/s) and costing less per task. The medium level at 53 beats Claude Sonnet 5 on price while nearly matching its intelligence score.

3.7 Flash: What Changed in Three Weeks

The headline: a 12% intelligence increase with no price increase. 3.7 Flash launched August 13, 2026, exactly 23 days after 3.6 Flash.

Google calls it "our most intelligent workhorse model yet for coding and agents." The improvements come from algorithmic changes, not scale. Same Flash architecture, same pricing tier, but noticeably smarter output. Here's what Artificial Analysis measured:

The cost-per-task drop deserves attention. Despite the same per-token pricing ($0.75/$3.75), 3.7 Flash costs 20% less per task than 3.6 Flash. That means the model is getting better at solving problems with fewer tokens. Smarter and more efficient.

Google specifically mentioned improvements in "debugging, issue resolution, and better first pass code accuracy." Evan Otero from the OpenRouter team confirmed: "Better intelligence for complex workflows. Tons of improvements in debugging, issue resolution, and better first pass code accuracy."

Three effort levels are available, similar to how Claude Opus 5 handles effort:

At low effort, 3.7 Flash matches Luna's intelligence (51) at a fraction of the speed cost. At high effort, it competes with models priced 4-8x higher. The low effort level is particularly useful for sub-agent routing where you need "good enough" reasoning at minimal cost.

Pricing runs through the end of 2026 at the introductory $0.75/$3.75. Starting January 1, 2027, it reverts to $1.50/$7.50. On OpenRouter, the current pricing is even lower at $0.375/$1.875 with an additional 50% promotional discount.

Available in the Gemini API, Google AI Studio, Android Studio, Google Antigravity, and through Spark (Google's personal agent for AI Pro and Ultra subscribers).

Where 3.6 Flash Improved Over 3.5 Flash

The headline number: time per task dropped from 2.7 minutes to 1.3 minutes, a 50%+ reduction measured by Artificial Analysis. That's driven by two things working together: 17% fewer output tokens per task and nearly double the output speed (304 tok/s vs 156 tok/s). On some benchmarks, the efficiency gain is much larger. Google says DeepSWE token usage dropped by 65%.

Here's what the benchmark improvements look like:

The coding gains are real. DeepSWE and MLE Bench both saw double-digit improvements. 3.6 Flash "delivers higher precision with fewer unwanted code edits and reduced execution loops," according to Google. In practice, that means less back-and-forth when using it as a coding agent.

Computer use is now a built-in client-side tool through the Gemini API. The 83% OSWorld-Verified score is the highest in Google's comparison table, ahead of GPT-5.6 Luna and Grok 4.5. If you're building browser automation or computer-use agents, 3.6 Flash is worth testing.

The long-context performance jump is the most dramatic. GDM-MRCR v2 at 1M tokens went from under 27% to 54%. That means 3.6 Flash can actually use its full 1M context window without quality collapsing. The previous generation couldn't say that.

Knowledge cutoff moved from January 2025 to March 2026. Fourteen months of fresher training data. That matters for anything involving recent APIs, libraries, or world events.

One caveat: early users are reporting that the knowledge cutoff doesn't always behave as expected. When asked about the best frontier model, 3.6 Flash confidently answered "Claude 3.5 Sonnet," a model from late 2024. A March 2026 cutoff should know better. This kind of confident-but-wrong factual recall was one of the issues that delayed Gemini 3.5 Pro, where internal checkpoints showed "frequent knowledge cutoff hallucinations" according to leaked test results. If your workflow depends on the model knowing recent facts, use Search Grounding or RAG rather than trusting the base weights.

3.5 Flash-Lite: The Speed and Cost Story

Flash-Lite isn't trying to compete with frontier models. It's built for the parts of your pipeline where you need fast, cheap, good-enough responses at scale.

Artificial Analysis measured time per task at 0.6 minutes, nearly half the 1.0 minutes of Gemini 3.1 Flash-Lite. Average output tokens per task dropped from 20K to 13K, meaning Flash-Lite is doing the same work with fewer tokens. The cost per task still went up ($0.04 to $0.09) because the per-token pricing increased, but the Intelligence gain is massive: +11 points on the Intelligence Index, with the biggest jumps in agentic evaluations (GDPval-AA went from 642 to 1140, TerminalBench from 31 to 53.6).

The comparison that matters here is against models in the same price bracket:

Flash-Lite is a third of the input cost of GPT-5.4 mini. Less than a third of Claude Haiku 4.5. And it lands within a few points on most benchmarks while leading on some.

What caught my attention: Flash-Lite beats Gemini 3 Flash on SWE-Bench Pro (54.2% vs 49.6%) and OSWorld-Verified (74.0% vs 65.1%). The "Lite" label is misleading. This model is genuinely better than last generation's main Flash at agentic and coding tasks.

Google designed configurable thinking levels for Flash-Lite. You can dial reasoning down to "minimal" for high-volume batch work, or push it to higher thinking for complex sub-agent tasks. That flexibility matters when you're routing different parts of a pipeline through the same model.

Gemini 3.5 Flash Cyber: Google's Security-Focused Model

This one's different from the other two. It's not a general-purpose model you can call through the API.

Gemini 3.5 Flash Cyber is fine-tuned from 3.5 Flash for one purpose: finding and fixing security vulnerabilities. It doesn't run standalone. It works inside CodeMender, Google's code security agent, where multiple Flash Cyber agents coordinate to scan code, validate findings, and produce a combined vulnerability report.

The architecture is multi-agent. Several Flash Cyber instances work in parallel on different aspects of a codebase, each handling discovery or verification. The combined report goes through validation before surfacing results. Google says this reaches "competitive frontier-level performance" on CyberGym at a cheaper cost than throwing larger models at the same problem.

Why this matters beyond security: Flash Cyber shows that fine-tuning a cheap, fast model for a specific domain and wrapping it in purpose-built agent infrastructure can match frontier performance. The model itself costs less per token than larger alternatives. The agent architecture (CodeMender) is doing the heavy lifting on coordination and accuracy.

The access model is deliberately restricted. Governments and trusted partners only, via a limited pilot. No public API. No timeline for broader access. Google is being cautious, and for good reason. A model optimized for finding vulnerabilities could also help exploit them. The restricted release is the right call.

For most developers, this doesn't change your workflow today. But it signals Google's playbook for specialized AI: take a cheap Flash model, fine-tune it for a domain, wrap it in an agent framework, and deploy it to targeted users. Expect similar specialized models for healthcare, legal, finance, and other verticals. If Kimi K3's unrestricted security capabilities make you nervous, Google's gated approach is the opposite end of that spectrum.

Gemini 3.5 Pro: When Is It Coming Out?

Google announced Gemini 3.5 Pro at I/O 2026 in May. It was supposed to ship in June. Then an intermediate deadline passed. Then July 17. It's missed all of them.

The July 21 blog post said "Gemini 3.5 Pro is currently testing with partners and we plan to make it broadly available as soon as it's ready." The August 13 blog post (for 3.7 Flash) doesn't mention 3.5 Pro at all. That silence says more than any update would.

The backstory is worse than a simple delay. According to reporting from Bloomberg and The Verge, Google DeepMind scrapped the original Gemini 3.5 Pro entirely after internal testers found structural failures in recursive tool-calling and SVG generation. They did a full pretraining restart. The rebuilt model (internally called "Rev25") reportedly still has weak coding performance and frequent knowledge cutoff hallucinations. Some older "Rev24" checkpoints perform better on coding than the newer ones, which suggests the regression isn't fully resolved.

The hallucination issue deserves emphasis. Internal testing found that 3.5 Pro confidently generates answers about events it should acknowledge uncertainty about. For anyone building retrieval-augmented pipelines or agent workflows that depend on the model knowing what it doesn't know, that failure mode is worse than simply getting lower benchmark scores.

I'll be honest: this is becoming a pattern. Google ships Flash variants on time and delays Pro. It happened with 2.5 Pro too. Multiple Google employees told Bloomberg they're frustrated, concerned the company is losing ground to Anthropic and OpenAI. Google's response: "We're shipping quickly across models while keeping them effective for customers."

The Flash updates are solid and production-ready. But if you were planning API strategy around Gemini 3.5 Pro benchmarks, don't. No confirmed benchmark numbers exist for the rebuilt model. Build against what you can test today.

And here's the bigger picture. While everyone waits for 3.5 Pro, Google confirmed Gemini 4 pre-training has started. The question developers are starting to ask: does anyone still need 3.5 Pro if 3.7 Flash covers most production use cases at Intelligence 56 and Gemini 4 is the next real leap?

How 3.7 Flash Compares to the Competition

3.7 Flash changes the competitive picture. At Intelligence 56, it's no longer just a "fast cheap model." It's genuinely competitive with mid-frontier models at a fraction of the cost.

3.7 Flash vs GPT-5.6 Terra: 3.7 Flash scores higher (56 vs 55) and costs 63% less on input ($0.75 vs $2) and 69% less on output ($3.75 vs $12). Speed isn't even close: 340 tok/s vs 75 tok/s. The only reason to pick Terra now is if you need the OpenAI ecosystem or max/ultra reasoning modes.

3.7 Flash vs Grok 4.5: 3.7 Flash scores higher (56 vs 54) and runs 5x faster (340 vs ~70 tok/s). Grok costs more on input ($2 vs $0.75) but less on output ($6 vs $3.75). Grok's token efficiency (fewer tokens per task) narrows the cost gap. For latency-sensitive work, 3.7 Flash wins. For token-constrained budgets, Grok is competitive.

3.7 Flash vs Kimi K3: K3 still leads on intelligence (57 vs 56) and has open weights. 3.7 Flash is 5.5x faster (340 vs 62 tok/s) and costs 75% less on output ($3.75 vs $15). For speed-sensitive production workloads, 3.7 Flash wins. For self-hosting and peak open-weight quality, K3 is still the pick.

3.7 Flash vs GPT-5.6 Luna: Luna costs less ($0.20/$1.20 vs $0.75/$3.75) but scores 5 points lower (51 vs 56). 3.7 Flash (low) matches Luna's intelligence at 51 while running 70% faster (254 vs 150 tok/s). Luna wins on pure cost. 3.7 Flash wins on intelligence-per-dollar.

3.7 Flash vs 3.6 Flash: Upgrade immediately. +6 points on Intelligence (50 to 56), 12% faster (304 to 340 tok/s), 20% cheaper per task ($0.50 to $0.40), same pricing. No reason to stay on 3.6 Flash.

Real-World Cost Comparison

Using our cost calculator assumptions: 500 tasks/month at 20,000 average tokens per task.

3.7 Flash at ~$25/month for 500 tasks is remarkable. It scores 56 on the Intelligence Index while costing less than Sonnet 5, Terra, or Grok 4.5. Only Luna is cheaper, and it scores 5 points lower. On the OpenRouter promotional pricing ($0.375/$1.875), the monthly cost drops to ~$12.

Artificial Analysis measured 3.7 Flash (high) at $0.40 per task. 3.7 Flash (low) at $0.16 per task. For comparison, GPT-5.6 Sol costs $7.08 per task. The cost difference is 17x.

Best Setup for Agent Workloads

Google clearly built these models for multi-agent architectures. Here's what the pricing and performance data suggests:

Two-model setup within Gemini:

• Master agent: Gemini 3.7 Flash at high effort ($0.75/$3.75). Handles orchestration, complex reasoning, computer use.
• Sub-agents: Gemini 3.7 Flash at low effort ($0.75/$3.75, $0.16/task). Handles routing, data extraction, summarization, bulk processing.

You don't even need two different models. Just switch effort levels. One model, two modes. Estimated monthly cost for moderate usage: $20-40.

Mixed-provider setup:

• Hard problems: GPT-5.6 Sol or Claude Opus 5 for tasks that need 59+ Intelligence scores.
• Production agents: Gemini 3.7 Flash (high) for speed-sensitive agent loops that need strong intelligence.
• Bulk sub-tasks: Gemini 3.7 Flash (low) or 3.5 Flash-Lite for anything high-volume.

This splits the workload by what each model does best. Sol or Opus handles the 10% of tasks that need frontier intelligence. 3.7 Flash handles the 60% that need speed and solid quality. Flash (low) or Flash-Lite handles the 30% that just need to be fast and cheap.

Use our model selector to find the right fit for your workload.

What Developers Are Saying

Community feedback on 3.7 Flash is noticeably more positive than the 3.6 reception. The intelligence jump from 50 to 56 addressed the main criticism of 3.6 (that it scored below Grok 4.5, Sonnet 5, and even Luna on raw intelligence). At 56, it now sits above all of those.

The rapid iteration is getting attention. Going from 3.5 to 3.6 to 3.7 in under a month suggests Google's Flash team has found a productive optimization loop. Whether they can sustain this pace is an open question, but the results speak for themselves.

The 3.6 Flash criticism was valid and worth revisiting for context. Some developers testing it on coding tasks, particularly web development and 3D work, found Grok 4.5 performing better. One common complaint: for the same budget, you could run GLM-5.2 or Grok 4.5 and get higher intelligence scores. With 3.7 Flash scoring 56, that complaint no longer applies.

The knowledge cutoff issue from 3.6 Flash (recommending "Claude 3.5 Sonnet" as the best frontier model despite a March 2026 cutoff) hasn't been confirmed as fixed in 3.7. If your workflow depends on the model knowing recent facts, use Search Grounding or RAG.

Flash-tier models are designed for speed and cost, not peak intelligence. Comparing them against full-sized frontier models on raw quality is the wrong framing. At Intelligence 56, though, 3.7 Flash is blurring that line. It's within 1 point of Kimi K3 and above GPT-5.6 Terra. That's not a Flash model you can dismiss as "budget."

Benchmark Gaps Worth Knowing

A few things Google doesn't cover that affect your evaluation.

No SWE-Bench Verified numbers for 3.7 Flash yet. Google reports SWE-Bench Pro but not Verified, which has historically been Anthropic's strongest benchmark. Every lab leads with the benchmarks where they shine.

Artificial Analysis has published full Intelligence Index results across all three effort levels, confirming the intelligence gains. Their independent testing measured 340 tok/s for 3.7 Flash (high), 274 tok/s for medium, and 254 tok/s for low.

The introductory pricing is important to understand. $0.75/$3.75 applies through December 31, 2026. Starting January 1, 2027, the price doubles to $1.50/$7.50. At that point, the cost comparison changes significantly. Plan accordingly if you're building something that depends on the current pricing.

Should You Switch to 3.7 Flash?

Switch from 3.6 Flash immediately. +6 Intelligence points, faster, cheaper per task. Same pricing. This is a free upgrade.

Switch from 3.5 Flash immediately. 3.7 Flash is better on every dimension. No discussion needed.

Switch from GPT-5.6 Terra if you don't need the OpenAI ecosystem. 3.7 Flash scores higher (56 vs 55), runs 4.5x faster, and costs 63-69% less.

Consider switching from Grok 4.5 if speed matters more than token efficiency. 3.7 Flash scores higher (56 vs 54) and runs 5x faster, but Grok uses fewer tokens per task.

Stay on Sol/Opus 5/Fable 5 if you need frontier intelligence above 56 on the Index, or you depend on OpenAI/Anthropic-specific tooling.

Use Flash-Lite if you're running high-volume batch processing, sub-agent tasks, or cost-sensitive pipelines where 36 Intelligence is sufficient. At $0.30/$2.50, it's still the cheapest option for bulk work.

FAQ

Is Gemini 3.7 Flash better than 3.6 Flash?

Yes, across the board. 3.7 Flash scores 56 on the Intelligence Index vs 3.6 Flash's 50. It runs faster (340 vs 304 tok/s), costs 20% less per task ($0.40 vs $0.50), and has the same per-token pricing ($0.75/$3.75 intro). Three effort levels (low/medium/high) give you more granular control. There's no reason to stay on 3.6 Flash.

How much does Gemini 3.7 Flash cost?

$0.75 per million input tokens and $3.75 per million output tokens (introductory pricing through December 31, 2026). Starting January 2027, prices double to $1.50/$7.50. On OpenRouter, there's an additional 50% promotional discount bringing it to $0.375/$1.875. Cache hits get a 90% discount.

Is Gemini 3.7 Flash better than GPT-5.6 Luna?

3.7 Flash (high) scores 56 vs Luna's 51 on the Intelligence Index and runs 2.3x faster (340 vs 150 tok/s). Luna is cheaper ($0.20/$1.20 vs $0.75/$3.75). 3.7 Flash (low) matches Luna's intelligence at 51 while running faster. Pick based on whether you need intelligence or cost savings.

How fast is Gemini 3.5 Flash-Lite?

350 tokens per second, making it one of the fastest models available. For context, GPT-5.6 Luna runs at 150 tok/s and Claude Sonnet 5 at 78 tok/s. Flash-Lite is built for workloads where latency matters more than peak intelligence.

What happened to Gemini 3.5 Pro?

It missed its June deadline, an intermediate date, and a July 17 target. The August 13 blog post for 3.7 Flash doesn't mention it at all. Internal reports suggest the model was rebuilt from scratch after failures in recursive tool-calling and still has hallucination issues. Don't plan your API strategy around it.

Can I use Gemini 3.5 Flash Cyber?

Not yet through the public API. Flash Cyber is available only to governments and trusted partners through a limited-access pilot via CodeMender, Google's code security agent. There's no timeline for broader availability.

How does Gemini 3.7 Flash compare to Kimi K3?

K3 scores slightly higher on the Intelligence Index (57 vs 56) and has open weights. 3.7 Flash is 5.5x faster (340 vs 62 tok/s) and costs 75% less on output ($3.75 vs $15). For speed-sensitive production workloads, 3.7 Flash wins. For peak coding quality and self-hosting, K3 is the better pick.

Is Gemini 3.5 Flash-Lite good enough for coding?

Yes, for many tasks. It scores 54.2% on SWE-Bench Pro, which beats Gemini 3 Flash (49.6%). It won't replace Sol or Fable 5 for complex debugging. But for code explanation, simple refactoring, test generation, and agent sub-tasks, Flash-Lite handles the work at a fraction of the cost.

When is Gemini 3.5 Pro coming out?

No confirmed date. Google announced it at I/O 2026 in May. It's missed every deadline since. The rebuilt checkpoints reportedly still show weak coding performance and knowledge cutoff hallucinations. With 3.7 Flash now scoring 56, the gap between Flash and what Pro was expected to deliver keeps shrinking.

Why is Gemini 3.5 Pro delayed?

Google DeepMind scrapped the original model after finding structural failures in recursive tool-calling and SVG generation tasks. A full pretraining restart followed. The rebuilt checkpoints reportedly still show weak coding performance and knowledge cutoff hallucinations. Bloomberg reported that Google employees are frustrated, concerned about losing ground to Anthropic and OpenAI. Rather than ship a model with known regression issues, Google chose to delay and ship Flash updates instead.

What is Gemini 4?

Google confirmed they've started "the most ambitious pre-training run yet" for Gemini 4. No specs, no timeline, no benchmarks. It's a signal that Google's next major model generation is underway, separate from the 3.5 Pro delays. Expect more details later in 2026.

Benchmark data from Artificial Analysis Intelligence Index v4.1.1. Model details from Google's official blog posts for 3.7 Flash and 3.6 Flash. 3.5 Pro delay reporting from Bloomberg, The Verge, and ChatForest. Additional analysis from officechai.com. Updated August 13, 2026. See the LLM Leaderboard for live rankings, Best LLM in 2026 for the full comparison, or the Benchmark Dashboard for side-by-side model analysis.