Gemini 3.7 Flash landed on August 13, 2026 - 23 days after 3.6 Flash - and its DeepSWE v1.1 score jumped 16 points in that window. That is the fastest capability gain in the Flash line's history, and it happened without a new base model. The question for teams building agent workflows is whether those numbers transfer to real work, or whether they stay in the benchmark table.
The short version: the coding gains are real and independently confirmed. The workflow automation claim is the most interesting and the least trustworthy. The pricing looks like a deal until you read the footnotes.
What Gemini 3.7 Flash actually scored on coding
The coding improvement is the story Google told, and it largely holds up. On FrontierCode 1.1, which measures production code quality, 3.7 Flash scores 43.6% against 34.4% for 3.6 Flash. On DeepSWE v1.1, the long-horizon software engineering benchmark, it hits 65.3% versus 48.6%. That 16-point DeepSWE gain happened in less than four weeks.
At launch it scored 65.3% on DeepSWE v1.1 for long-horizon software engineering, against 49.0% for 3.6 Flash, and 43.6% on the main split of FrontierCode 1.1 - the best figure in Google's launch comparison, ahead of Claude Sonnet 5 and GPT-5.6 Terra.
Web development moved with it. The model had an Elo score of 1538 on Arena.ai's WebDev Arena for last month's model, and this release comes in at 1588.
The one honest regression: CharXiv Reasoning dropped to 84.5% without tools, down from 85.2% for 3.6 Flash. It is a small number on a narrow eval, but it matters for teams that use the Flash tier on dense mathematical reasoning tasks.
The model offers three thinking_level settings: low, medium, and high, with medium as the default. Low is intended for latency-sensitive work such as incident-response pipelines and real-time chat. That setting dial is the practical control most launch coverage buried. An agent pipeline doing a fast classification pass does not need the same thinking budget as one resolving a multi-file refactor.
The AutomationBench number deserves skepticism
This is the finding you should not route a platform migration on without further testing. On AutomationBench, which Google describes as measuring enterprise workflow automation, Gemini 3.7 Flash scores 30.4%, up sharply from 17.0% for 3.6 Flash. Google's table lists Claude Sonnet 5 at 10.7% and GPT-5.6 Terra at 23.6%.
A model outperforming Claude Sonnet 5 by 19 points on workflow automation would be significant if the benchmark were independent. It is not. AutomationBench is a private Google eval set. Its Zapier provenance deserves scrutiny before it drives a platform migration decision.
The DeepSWE and FrontierCode numbers are worth trusting because they are run by third parties against fixed holdout sets. AutomationBench is worth watching, not acting on immediately. The right call is to run your own evals against the specific workflow you care about - read-doc, call-tool, produce-output chains on real data - before treating the 30.4% as ground truth.
That combination of document comprehension, tool invocation, and output generation is relevant for enterprise agents because many practical deployments require more than generating text or code. An agent may need to interpret a long report, identify relevant information, decide which tool to invoke, update another system, and produce a document for a human reviewer. Reliability across that chain can matter more than performance on an isolated reasoning benchmark.
Where it trails and what that means for agent routing
The agentic computer-use scores are the exception to the sweep. 3.7 Flash scores 38.1% on OSWorld 2.0 and 14.9% on Terminal-Bench 3.0, both trailing GPT-5.6 Terra at release.
GPT-5.6 Terra is ahead on DeepSWE (69.6%), Terminal-bench 2.1 (87.4%), Terminal-bench 3.0 (20.8%), and OSWorld-2.0 (50.2%).
The practical routing table that falls out of this:
| Task type | 3.7 Flash vs peers | Verdict |
|---|---|---|
| Production code quality (FrontierCode) | Leads Sonnet 5, Terra | Strong choice |
| Long-horizon engineering (DeepSWE) | Trails Terra by 4 pts | Competitive |
| Web UI generation (WebDev Arena) | Top of Google's comparison | Strong choice |
| Business workflow automation (AutomationBench) | Leads - private eval | Test before trusting |
| Computer use / terminal work | Trails Terra by 5-6 pts | Use Terra here |
| Broad professional knowledge (GDPval-AA v2) | Trails Sonnet 5, Muse | Not the right tier |
The pattern: 3.7 Flash is the strongest Flash-tier option for coding and document-heavy pipelines. It is not a computer-use model, and teams with requirements that sit in Pro territory - sustained complex reasoning, long-context document analysis - still have no Gemini option at that tier.
Gemini 3.5 Pro is still missing. Google promised it for June, later said it was coming "soon," and has since started training Gemini 4.
Beagle in action#eng-ops, 10:32amThe pricing catch most launch posts skipped
Gemini 3.7 Flash costs $0.75 per million input tokens and $3.75 per million output tokens at launch, with those introductory rates doubling to $1.50 and $7.50 on January 1, 2027. That's half of what 3.6 Flash launched at three weeks ago.
The piece that almost no coverage printed: thinking tokens bill as output tokens. An agent that reasons hard on a "high" thinking budget can burn far more output than the visible response suggests, so model your costs on real traces, not on response length.
There is a second catch. Google quietly cut 3.6 Flash to the identical introductory rate at the same time, so the two Flash workhorses bill the same until December 31. That means the migration decision right now is entirely about capability and reliability, not price. During the promotional period, the migration decision is therefore about reliability and completed work, not a cheaper list price.
A practical cost model for an agent doing 10,000 turns per day, each consuming roughly 2,000 input tokens and generating 800 output tokens (visible response + medium thinking):
- Daily input: 20M tokens → $15.00
- Daily output: 8M tokens → $30.00
- Total: ~$45/day at introductory rates, doubling to ~$90/day from January 1
Run that against your specific thinking budget before locking in a design that assumes the introductory rate holds.
The 23-day release cycle itself is information. Three Flash-tier upgrades between May and August, against a Pro flagship untouched since February, made clear which tier Google sees as the centre of its lineup. If you are building on Gemini, Flash is where the iteration is happening. Plan your architecture around that.
Gemini 3.7 Flash for agents: common questions
Is Gemini 3.7 Flash better than Gemini 3.6 Flash for agent workflows?
For coding and document-heavy pipelines, yes - the 16-point DeepSWE gain and the GDP.pdf jump from 22% to 34% are meaningful. For computer-use and terminal tasks, 3.6 and 3.7 score similarly and both trail GPT-5.6 Terra. If you are on 3.6 already, test before migrating; Google has not announced a shutdown date for 3.6.
What does Gemini 3.7 Flash cost per million tokens?
Introductory pricing is $0.75 per 1M input tokens and $3.75 per 1M output tokens through December 31, 2026. On January 1, 2027, those rates rise to $1.50 and $7.50. Thinking tokens count as output, so high-reasoning agent turns cost more than the visible response length suggests.
How does Gemini 3.7 Flash compare to Claude Sonnet 5 on coding?
3.7 Flash beats Claude Sonnet 5 on FrontierCode, AutomationBench, and WebDev Arena while costing roughly a third as much per output token. Sonnet 5 leads on broad professional knowledge (GDPval-AA v2) and complex agentic reasoning benchmarks like Agent's Last Exam.
What is the thinking_level setting in Gemini 3.7 Flash?
The model offers three thinking_level settings - low, medium, and high - with medium as the default. Low is intended for latency-sensitive work such as incident-response pipelines and real-time chat. Setting it to high on every agent call is the fastest way to turn a reasonable token budget into an expensive surprise.
Should I trust the AutomationBench score for picking a workflow model?
Not alone. AutomationBench is a private Google eval set, not a third-party benchmark. The +13-point lead over GPT-5.6 Terra is worth investigating but not worth acting on without running your own evals on representative tasks from your specific pipeline.