The Model Menu Grows: A Frontier Showcase, a Faster Flash Model, and Cheaper Diffusion

What builders made with Gemini 4 Argon

Google says it released a new frontier model called Gemini 4 Argon, designed for complex tasks in coding, finance, legal, cybersecurity, and science, and rolled out broadly last week, with access for developers and enterprises through the API in AI Studio and for subscribers through the Gemini app. m:…1406

The company's showcase of five builders gives a concrete picture of early use: one builder rendered a 3D simulation of a seaplane, another built an interactive 3D globe from NASA and USGS data, a third modeled a jet engine with Blender and Three.js, a fourth built a 3D carriage from a single image, and a fifth built a multiplayer browser game called Backyard Rascals. m:…1406

Google says the model outperforms other frontier models on benchmarks across domains, a company claim that builders will have to test against their own workloads, since the showcase emphasizes 3D and interactive builds rather than the finance, legal, and cybersecurity tasks the model is also designed to handle. m:…1406

MiniMax M3.1 Flash: faster, cheaper, benchmarked

SambaNova says MiniMax M3.1 Flash is now available on SambaCloud, delivering over 200 tokens per second, running 3 times faster than its predecessor while reducing usage costs by 70 percent, with existing MiniMax M3 traffic on the platform being rerouted to the new model. m:…2627

The benchmark scores SambaNova reports span coding, agentic, and tool-use tasks: 73.7 percent on DeepSWE version 1.1, 82 percent on CyberGym, 61.7 on OSWorld version 2, 76.2 on BankerToolBench, and 29.6 on Agents' Last Exam, a spread that suggests the model is stronger on some agentic benchmarks than others rather than uniformly ahead. m:…2627

Pricing is explicit: cached input tokens cost 0.03 dollars per million tokens, regular input tokens cost 0.30 dollars per million, and output tokens cost 1.20 dollars per million, which gives builders a clear arithmetic for comparing the flash model against frontier models on tasks where speed matters more than peak capability. m:…2627

Diffusion models: cost depends on what you ask

A preprint study of diffusion language models adds a different kind of cost curve: these models can write several tokens in one forward pass, but the number of forward passes per token depends heavily on the task, with highly tokenized languages requiring more passes per token than others. m:…1646

The study finds that diffusion models spend more computation on novel, rare, or incorrect outputs, that text copied or lightly edited from the input is cheaper to generate than open-ended writing, and that generating a token for story writing in DiffusionGemma requires more than three times as many forward passes as for grammar correction. m:…1646

Two practical findings follow for builders: wrong answers on math problems take more passes per token than correct ones, and weighting candidate answers by decoding speed beats majority voting by up to 4.2 points, which means decoding speed itself carries signal about answer quality in these models. m:…1646

Three cost curves, one choice

The three model stories share a question but give different answers: Gemini 4 Argon offers a frontier showcase through an API and subscription apps, MiniMax M3.1 Flash offers measured speed and explicit token prices, and diffusion models offer a cost that varies with the task itself, so the right model depends on whether the work is showcase, throughput, or task-sensitive generation. m:…1406 m:…2627 m:…1646

Benchmarks alone do not settle the choice: Google says its model outperforms other frontier models across domains, SambaNova reports a spread of scores across agentic benchmarks, and the diffusion study shows decoding speed itself beating majority voting, which means builders need their own workload measurements alongside vendor numbers. m:…1406 m:…2627 m:…1646