Google releases Gemini 3.7 Flash for coding agents—but Gemini 3.5 Pro remains in limbo

Google unveiled Gemini 3.7 Flash on Thursday as its newest model for powering coding agents and business automation workflows, just three weeks after the release of Gemini 3.6 Flash. The rapid succession raises questions about model fatigue, but the company argues the new version is meaningfully stronger at writing and maintaining code over extended agent tasks.

The timing underscores Google's aggressive approach to competing in generative AI development velocity. However, it also means the long-promised Gemini 3.5 Pro—originally expected in June—remains absent from the lineup, with the company now focused on training Gemini 4 instead.

Substantial gains on coding benchmarks#

Gemini 3.7 Flash demonstrated sharp improvements across multiple evaluation frameworks. On DeepSWE v1.1, the model achieved a 65.3% score, up from 49% for its predecessor. FrontierCode 1.1 Main showed similar momentum, jumping from 34.4% to 43.6%. WebDev Arena performance increased modestly from 1,538 to 1,588.

The pattern echoes developments elsewhere: smaller, purpose-built models continue to outpace expectations within their class. Yet benchmarks tell an incomplete story. Real-world performance depends on how reliably these gains translate to individual codebases and whether the model can recover gracefully when it encounters unfamiliar patterns.

Pricing now, cost later#

To accelerate adoption, Google is charging $0.75 per million input tokens and $3.75 per million output tokens—half the original rate for Gemini 3.6 Flash. This promotional window lasts until December 31, 2026. Beginning January 1, 2027, both rates double to $1.50 and $7.50 respectively.

The incentive structure is deliberate: teams get time to evaluate and deploy agents at favorable pricing, but longer-context workflows or reasoning-intensive tasks will face significantly higher operational costs in 2027. Developers who build integrations now should anticipate that economics change materially in the new year.

Configurable reasoning levels balance speed and cost#

The model introduces three thinking_level settings—low, medium, and high—that control how much reasoning the model performs before responding. Low suits time-critical applications like incident response; medium balances responsiveness with coding tasks; high allocates more compute for complex logic or planning challenges.