Google AI stands its ground: Gemini 3.8 achieves near-Opus5 performance at just 15% of the price, securing first place in 8 programming categories.

Google's Gemini Flash model proves that sheer scale works wonders. The newly released Gemini 3.8 Flash is impressive, surpassing the current flagship model Opus 5 in multiple performance benchmarks.

As one of the top three AI giants, Google has clearly fallen behind in large language models this year. The repeated delays of Gemini 3.5 Pro have undermined confidence in the company, so it has essentially shelved the Pro lineup for now. Instead, it has focused on rapidly upgrading the Flash series of low-cost, fast models, which has jumped from 3.5 all the way to Gemini 3.8 Flash, with performance growing increasingly formidable.

The latest Gemini 3.8 Flash maintains its low-price advantage: $0.75 per million input tokens and $3.75 per million output tokens, with a 1 million token context window and a 65,000 token output cap. However, the promotional pricing only lasts until December 31, after which it will double starting January 1, 2027.

Compared to Opus 5, Gemini 3.8 Flash's overall pricing is only about 15% of the latter's, meaning you could run seven Gemini 3.8 Flash tasks for the cost of one Opus 5 task. Combined with its multimodal strengths, it's well-suited for handling miscellaneous workloads with high efficiency and low cost.

That said, Gemini 3.8 Flash's performance is nothing to scoff at. It secured first place in 8 out of 14 benchmark suites. On DeepSWE v1.1, it improved from 65.3% to 71.0%, approaching Opus 5's 74.0%.

On tests like Terminal-bench 2.1, it achieved 89.4%, slightly edging out Opus 5's 89.1% and GPT-5.6 Sol's 88.8%, demonstrating top-tier large model performance.

It also scored highest in financial agent, legal agent, complex chart reasoning, and long video understanding benchmarks.

Of course, there are weak spots too. On Terminal-bench 4.0, it only managed 19.1%, compared to Opus 5's 51.8%, and on OSWorld-2.0 it scored just 59.0%, significantly trailing Opus 5's 75.4%.

It's also worth noting that while Gemini 3.8 Flash offers a 1 million token context window, it only provides 64K tokens for output—a rather stingy limit that makes it unlikely to handle large-scale tasks.

Gemini 3.8 Flash is now available through the Gemini app, AI Studio, and the Gemini API.