Gemini 3.7 Flash Benchmarks [Coding, Reasoning, Price]

Gemini 3.7 Flash is Google's fastest-improving mid-tier model, and the benchmark scores explain why builders are paying attention. It launched on August 13, only three weeks after Gemini 3.6 Flash, and the coding numbers moved more in those three weeks than most models move in a full generation. If you are deciding whether to route your app's traffic to this model, the scores below tell you where it wins, where it loses, and what the independent data confirms.

What are the benchmarks of Gemini 3.7 Flash?

Gemini 3.7 Flash posts leading scores in coding, expert reasoning, and long-context retrieval, with an Artificial Analysis Intelligence Index of 56 at the high thinking level. Google benchmarked it against Gemini 3.6 Flash, Claude Sonnet 5, GPT-5.6 Terra, and Muse Spark 1.2, and 3.7 Flash tops the table on 9 of the 20 published benchmark rows.

The table below pulls the headline results directly from Google DeepMind's official model card, published August 2026. Vendor-reported figures like these reflect Google's own testing harness, so treat them as the launch baseline rather than independent confirmation.

Gemini 3.7 Flash benchmark scores from the Google DeepMind model card - source: deepmind.google, August 2026

The pattern is consistent. Where Gemini 3.6 Flash was a competent but unremarkable Flash model, 3.7 Flash closes the gap with frontier models on knowledge work and pulls ahead on several coding evals. The AutomationBench result is a good example: it nearly doubled from 17.0% to 30.4%, and since that benchmark measures how well a model completes real business workflows end to end, the jump points to a model that can be trusted with more of a multi-step task before a human has to step in. Google credits developer feedback and algorithmic changes to the reasoning core for the jump.

Gemini 3.7 Flash coding benchmark results lead its own class

Coding is where Gemini 3.7 Flash separates from its predecessor most sharply. DeepSWE v1.1, a long-horizon software engineering eval that measures how well a model handles multi-step coding tasks, climbed from 48.6% to 65.3%. In practice, a higher DeepSWE score means the model can carry a task like "add auth to this app" across many files and steps without losing the thread. That 17-point gain is the single largest movement anywhere in the model card.

The other coding signals point the same direction:

• FrontierCode 1.1 Main, which scores production code quality, reached 43.6%, the highest in the comparison group and ahead of both Claude Sonnet 5 and GPT-5.6 Terra. A higher score here means fewer bugs and less cleanup in the code the model writes the first time.
• Code Arena, an Elo-rated benchmark for web development, put 3.7 Flash at 1588, roughly 50 points clear of every other model tested. In plain terms, people comparing its web interfaces side by side preferred them over what the other models produced.
• Terminal-bench 2.1, which measures agentic terminal coding, rose to 85.8%, though GPT-5.6 Terra still edges it at 87.4%. This score tracks how reliably the model can run real commands in a terminal to get a job done, which matters for any tool that automates setup or deployment.

For teams building coding agents, the DeepSWE result matters most because it reflects sustained, multi-file work rather than single-shot completions. The FrontierCode and Code Arena scores suggest cleaner first-pass code and stronger front-end generation, two things that show up quickly when you put a model in front of real users.

There is a caveat worth naming. On DeepSWE v1.1 itself, GPT-5.6 Terra scores higher at 69.6%, and on the harder Terminal-bench 3.0 general-agent eval, 3.7 Flash lands at 14.9% against Terra's 20.8%. Terminal-bench 3.0 raises the difficulty to open-ended agent tasks, so the low scores across the board show how much headroom still remains for fully autonomous work. Flash-tier speed and price come with a ceiling, and the most demanding agentic workflows still favor the larger frontier models.

Gemini 3.7 Flash reasoning benchmark scores rival frontier models

Gemini 3.7 Flash reasoning benchmarks put it within a point of models that cost three times as much, with leading scores on expert reasoning and long-context retrieval. The gains show up across three distinct reasoning dimensions, each covered below.

1. Expert reasoning beats models that cost three times more

On HLE-Verified, a test of multidisciplinary expert reasoning, Gemini 3.7 Flash scored 53.6%, ahead of GPT-5.6 Terra at 51.1% and far ahead of Claude Sonnet 5 at 31.0%. A strong HLE score signals the model can handle hard, specialist questions across fields rather than just everyday prompts, so it holds up better on technical or research-grade work. For a model priced at $0.75 per 1M input tokens, that result reframes what a Flash-tier model is expected to do.

The Artificial Analysis Intelligence Index, which runs models through a common independent harness, backs this up. It placed Gemini 3.7 Flash (high) at 56, one point above Claude Sonnet 5 and one point below GPT-5.6 Terra, the closest reference available for a same-harness comparison. Note the "(high)" label: it reflects the top thinking-level configuration, not the default.

2. Long-context retrieval leads the comparison group

On GDM-MRCR v2 at the 128k window, 3.7 Flash hit 97.0%, the top score in the group and a meaningful jump from 3.6 Flash's 91.8%. This benchmark checks whether a model can find specific facts buried deep inside a very long input, so a high score means it won't lose track of details in a large document or long chat. Models that retrieve reliably across long inputs are easier to trust in document-heavy workflows, and this is where the improvement is most usable day to day.

3. Document comprehension jumped 12 points over 3.6 Flash

On GDP.pdf, which tests a model's ability to pull structured answers out of dense PDFs, 3.7 Flash scored 34.0% against 22.0% for its predecessor. If your product parses contracts, financial filings, or research papers, that gain is worth testing directly against your own documents.

How does Gemini 3.7 Flash pricing compare to its benchmarks?

Gemini 3.7 Flash keeps the same introductory price as 3.6 Flash while delivering higher scores, which is the core of its value case. Google lists $0.75 per 1M input tokens and $3.75 per 1M output tokens, roughly a third of what Claude Sonnet 5 and GPT-5.6 Terra charge for input. For a full breakdown of what actually drives your bill, see our Gemini 3.7 Flash pricing guide.

Gemini 3.7 Flash pricing versus benchmark scores - pricing as of August 2026, source: Google DeepMind model card

One detail deserves a flag before you commit. The $0.75 and $3.75 rates are introductory and expire on December 31, 2026. Starting January 1, 2027, input rises to $1.50 per 1M tokens and output to $7.50, doubling your token cost overnight. If you are modeling long-term unit economics, plan around the post-January price, not the launch price.

Even at the higher 2027 rates, Gemini 3.7 Flash stays cheaper than the frontier models it competes with on reasoning. The value case does not disappear when the discount ends; it just narrows.

Where Gemini 3.7 Flash still loses

Gemini 3.7 Flash does not win every benchmark, and an honest read of the model card shows GPT-5.6 Terra leading on the hardest agentic tasks. On OSWorld-2.0, which measures agentic computer use, Terra scored 50.2% to Flash's 47.9%. A higher OSWorld score means the model is better at actually operating a computer, clicking, typing, and navigating apps the way a person would, which is the core skill behind any hands-off automation. On Agent's Last Exam, a multimodal desktop and OS agent test, Claude Sonnet 5 led at 33.3% against Flash's 26.3%. That benchmark stresses the same skill under harder, mixed-media conditions, so the gap shows where 3.7 Flash still needs a human in the loop.

These gaps cluster in one place: long, tool-heavy agentic workflows that run many steps without human input. For those, the larger frontier models still hold an edge. For coding, reasoning, document analysis, and long-context retrieval, Gemini 3.7 Flash is competitive with or ahead of models that cost far more.

The takeaway for builders is practical. Match the model to the job. A support bot, a code assistant, or a document parser will likely run better and cheaper on 3.7 Flash. A fully autonomous computer-use agent may still justify a frontier-tier model. To see how Flash stacks up against Google's own Pro tier, our 3.7 Flash vs 3.1 Pro comparison breaks it down.

Beyond the benchmark scores

Gemini 3.7 Flash earns its benchmark headlines on coding and reasoning, and its price makes those gains hard to ignore for anyone shipping AI features. The scores confirm a model that punches above its tier, with the clearest wins in long-horizon coding, expert reasoning, and long-context retrieval, and honest limits in fully autonomous agentic work.

If you are building an app and want to use Gemini 3.7 Flash without wiring up API keys or juggling billing across providers, Emergent gives you access to Gemini, Claude, and GPT through a single Universal LLM Key. You describe what you want to build and ship a working app without managing infrastructure.

Start Building on Emergent.