Gemini 3.7 Flash Is the Quiet Value Upgrade

OpenAI’s GPT-5.6 Luna is the clear quality-per-dollar leader in WhatModel’s latest figures. Its input price is $0.1, its quality measure is 83.7, and its quality-per-dollar rating is 837. That is a wide lead over the next model in the value ranking, DeepSeek V4 Flash 0423, at 579.3.

The catch is that this ranking uses input price. Output costs still matter for applications that generate long answers, code, or documents, so the list is best read as a starting point rather than a complete bill estimate.

The quiet upgrade

The most interesting new buy is Google’s Gemini 3.7 Flash. It arrived with a score of 88, a quality measure of 87.1, and input and output prices of $0.375 and $1.875. That puts it third in the current value table, with a quality-per-dollar figure of 232.3.

What makes the model stand out isn’t just its ranking. It looks materially better than Google’s Gemini 3.6 Flash on the numbers supplied here: Gemini 3.6 Flash has a quality measure of 80.3, a score of 83, and an input price of $0.75. Gemini 3.7 Flash is cheaper on input while posting the higher quality measure and score.

I’d look at Gemini 3.7 Flash first for general-purpose workloads that need a large context window. Its context is 1048576, matching the batch version. The batch model cuts the listed prices to $0.1875 input and $0.9375 output, but it has no score yet. For scheduled or noninteractive work, that discount is meaningful; for a live application, I’d want evidence that the batch offering behaves as expected before treating it as a direct substitute.

DeepSeek keeps the pressure on

DeepSeek V4 Flash 0423 remains a strong value choice at $0.14 input, with a quality measure of 81.1, a score of 86, and quality per dollar of 579.3. Its advantage is simple: the price is low enough to make a high-quality model affordable at scale.

DeepSeek V4 Pro is also in the value table, with a quality measure of 82.5 and a score of 86. Its input price is $0.435, producing quality per dollar of 189.7. A new listing, DeepSeek V4 Pro 0813, is priced at $0.435 input and $0.87 output, but it has no score. I wouldn’t assume the newer listing is better until the quality data catches up.

OpenAI’s GPT-5.6 Terra remains the quality pick among the value entries, with a quality measure of 90.5 and a score of 90. But its $1 input price gives it quality per dollar of 90.5, far below Luna. That’s the trade-off in this market: the best absolute model and the best buy are often different products.

Watch the price hikes

Qwen’s recent changes weaken several previously attractive options. Qwen3 Coder 30B A3B Instruct rose from $0.07 to $0.2925 on input, a 317.9% increase. Qwen3 Coder 480B A35B rose from $0.3 to $0.975, up 225%. GLM 5.2 moved from $0.49 to $1.4, with the database showing an increase of 185.7% on one recorded change. Qwen3 30B A3B Instruct 2507 rose from $0.04815 to $0.13, a 170% increase.

Those moves matter more than a small quality difference when a model sits in a high-volume pipeline. I’d recheck any Qwen-based cost estimate rather than relying on an older comparison.

For the lowest possible input bill, Mistral Nemo is listed at $0.019 through DeepInfra, followed by GPT-5 Nano at $0.025 through OpenAI. Solar Pro 4 is $0.03 input and $0.12 output, with a score of 71. Cheap hosting can be useful, but without a comparable quality-per-dollar entry, price alone doesn’t tell me which model will do the job best.

My practical shortlist is GPT-5.6 Luna for maximum value, DeepSeek V4 Flash 0423 for another low-cost high-quality option, and Gemini 3.7 Flash for the strongest new all-rounder. Gemini is the model that quietly changed the buying conversation this week.