Gemini 3.7 Flash Review — Google's $0.75 Workhorse, Half the Price of 3.6 Flash
✅ Pros
- • Introductory pricing at $0.75/$3.75 per 1M tokens is exactly half of 3.6 Flash's launch price, and Google keeps the entry price through December 31, 2026
- • FrontierCode 1.1 Main jumps to 43.6% (from 34.7% on 3.6 Flash), with a LMArena Elo of 1588 vs 1538 — a genuine leap for a 'workhorse' tier model
- • Knowledge-work gains are concrete: GDP.pdf 34.0% vs 22.0%, AutomationBench 30.4% vs 17.0% — nearly doubling business-workflow completion rates
- • Full 1,048,576-token input context with 65,536-token output, supporting text, image, video, audio, and PDF inputs in one model
- • First-class agent stack: Antigravity, computer use (Preview), code execution, search grounding, structured outputs, and thinking levels (low/medium/high)
- • Early customer feedback cited in the launch post shows results significantly better than 3.6 Flash at a lower cost per task
⚠️ Cons
- • Introductory pricing is temporary: on January 1, 2027 the price doubles to $1.50/$7.50, and HN commenters note that two newer model generations will likely exist by then
- • AA Intelligence Index of 56 still trails frontier models (Grok 4.6 at 61, Fable 5 Max at 62), and the gain from 52 comes with 37k output tokens per task vs 26k on 3.6 Flash
- • Text-only output — audio generation and image generation are not supported, limiting some multimodal workflows
- • On pure text workloads, DeepSeek V4 Flash/Pro remains 13-26x cheaper per token with comparable intelligence, per HN analysis
- • No vision output and no minimal thinking mode (low/medium/high only), which narrows some cost-optimization paths
Developers and enterprises that want a balanced price-performance model with Google's agent tooling, multimodal input, and 1M-token context
$0.75 / $3.75 per 1M tokens (intro through 2026-12-31), then $1.50 / $7.50
Quick Verdict
Gemini 3.7 Flash is Google’s most capable small-model release to date — and at $0.75 per 1M input tokens it undercuts its own predecessor by 50%. On the headline coding benchmark, FrontierCode 1.1 Main, it scores 43.6% versus 34.7% for 3.6 Flash, and its LMArena Elo improved from 1538 to 1588. The model keeps the full 1M-token context and adds agent-first capabilities: computer use (Preview), code execution, search grounding, and three explicit thinking levels. The catch is the fine print: introductory pricing expires December 31, 2026, after which the price doubles. If you are planning a long-lived production workload, model the $1.50/$7.50 rate from day one.
Rating: 8.0/10 — a genuinely competitive workhorse, not a frontier model, and the price doubling makes the value story time-limited.
What’s New in 3.7 Flash
Google positions 3.7 Flash as “our most intelligent workhorse model,” and the benchmark deltas support that framing. The improvements cluster in three areas:
Software engineering. FrontierCode 1.1 Main rises from 34.7% (3.6 Flash) to 43.6%. Google also highlights gains in debugging, issue resolution, and first-pass code accuracy. The model “thinks more diligently,” putting more effort into multi-step planning and tool calls, which Google says translates to fewer retries in engineering workflows.
Knowledge work. The GDP.pdf benchmark — which tests complex-document processing — improves from 22.0% to 34.0%. AutomationBench, a real-world business workflow eval, nearly doubles from 17.0% to 30.4%. Google cites “significantly outperforming 3.6 Flash” on finance, law, and biosciences knowledge tasks.
Web development. Beyond raw benchmarks, Google’s demo shows a text prompt producing a fully playable 3D game, and an annual report PDF converted into an interactive data story with live charts.
Pricing: The Fine Print Matters
The headline number is compelling: $0.75 per 1M input tokens and $3.75 per 1M output tokens — half the original 3.6 Flash price. But this is introductory pricing that runs only through December 31, 2026. From January 1, 2027, the standard rates apply: $1.50 input / $7.50 output per 1M tokens.
| Plan | 3.7 Flash Intro (to 2026-12-31) | 3.7 Flash Standard (from 2027-01-01) | 3.6 Flash (reference) |
|---|---|---|---|
| Input (per 1M tokens) | $0.75 | $1.50 | $1.50 (launch) |
| Output (per 1M tokens) | $3.75 | $7.50 | $7.50 (launch) |
| Context window | 1,048,576 tokens | 1,048,576 tokens | 1M |
| Output limit | 65,536 tokens | 65,536 tokens | 65,536 |
The API model card confirms the full spec: input token limit of 1,048,576, output limit of 65,536, inputs of text/image/video/audio/PDF, and text-only output. Thinking is supported at low, medium, and high levels — notably, “minimal” is not supported and returns an error. Batch API, Flex inference, and Priority inference are all supported, and context caching works, which matters for agentic workloads with repeated system prompts.
Benchmark Table
| Benchmark | Gemini 3.7 Flash | Gemini 3.6 Flash | Delta |
|---|---|---|---|
| FrontierCode 1.1 Main | 43.6% | 34.7% | +8.9 |
| LMArena Elo | 1588 | 1538 | +50 |
| GDP.pdf | 34.0% | 22.0% | +12.0 |
| AutomationBench | 30.4% | 17.0% | +13.4 |
| AA Intelligence Index | 56 | 52 | +4 |
An important caveat from the HN thread: the AA Intelligence Index gain from 52 to 56 comes with more output tokens per task — 37k for 3.7 Flash versus 26k for 3.6 Flash. The halved input price partially offsets this, but per-task cost is not a clean 50% reduction at the output-heavy end.
Pros & Cons
Pros: Half-price intro tier; real coding and knowledge-work gains; 1M-token context with multimodal input; strong agent stack (Antigravity, computer use, code execution); context caching supported.
Cons: Price doubles in 2027; text-only output; AA Index still 5-6 points behind Grok 4.6 and Fable 5 Max; higher output-token consumption per task; no minimal thinking level.
Community Verdict
HN response (533 points, 65 comments at the time of writing) was mixed but substantive. The pricing skepticism was sharpest: “Introductory pricing until December 2026 implies no significant competition” — spellk. Another thread commenter compared against Cognition’s FrontierCode leaderboard, noting Terra sits at roughly half the price per task, and that “Grok 4.6 appears to straight up be better AND cheaper” — bisonbear. The multimodal argument was the main counterpoint: “if you deal with text only, what is the benefit of using this over DS V4 Flash/Pro? 13-26x cheaper with comparable intelligence” — euazOn. And on positioning: “This is genuinely a competitive model, considering it beats Claude Sonnet 5 on almost all benchmarks and is more than half its price. Seems like Google is back in the game, though not leading the frontier anymore” — nickandbro.
Alternatives
- Grok 4.6 — Higher AA Intelligence Index (61 vs 56) at $2/$6 per 1M tokens. Better for coding-agent workloads per xAI’s own comparisons, but pricier on input.
- DeepSeek V4 Flash / V4 Pro — $0.435/$0.87 per 1M tokens with Fable-class averages. The cost leader for text-only workloads; no multimodal input.
- Claude Sonnet 5 — Comparable benchmark scores but more than double the price per token; the “quality incumbency” choice for teams already on Anthropic.
- GPT-5.6 Tera — Cognition’s FrontierCode shows Terra at about half the per-task price of 3.7 Flash for code-specific work.
FAQ
Q: When does the introductory Gemini 3.7 Flash pricing end? A: December 31, 2026. Starting January 1, 2027, pricing doubles to $1.50 per 1M input and $7.50 per 1M output tokens.
Q: What context window does Gemini 3.7 Flash support? A: 1,048,576 tokens (1M) input, with a 65,536-token output limit.
Q: Is Gemini 3.7 Flash multimodal? A: Yes for input — text, image, video, audio, and PDF. Output is text-only; audio and image generation are not supported.
Q: Where can I use Gemini 3.7 Flash? A: Gemini API, Google AI Studio, Android Studio, Antigravity for agent-first workflows, and Gemini Enterprise Agent Platform for enterprises.
Q: How does 3.7 Flash compare to DeepSeek V4 Flash on price? A: DeepSeek V4 Flash is roughly 13-26x cheaper per token with comparable text intelligence, per HN analysis. Google’s edge is multimodal input and the agent toolchain.