Gemini 3.7 Flash Goes Half-Price as Alibaba Open-Sources a 2.4T Model

Two moves in mid-August reshaped the cheap end of the AI market, and both come with a catch worth reading before you re-plan your stack. Google shipped Gemini 3.7 Flash at half price on August 13, and Alibaba open-sourced Qwen3.8, a 2.4-trillion-parameter model.

Gemini 3.7 Flash: half price, with an expiry date

Google priced 3.7 Flash at $0.75 per million input tokens and $3.75 per million output, and called it half price. Read the asterisk: those rates run only through December 31, 2026, then double to $1.50 and $7.50 on January 1, 2027. The "half" is measured against the workhorse tier's standard rate, which is exactly where it lands again next year. It is a time-boxed discount, not a permanent cut, and standardizing on it now risks meeting doubled rates in January.

The capability gain looks real, though. The model scored 65.3% on the DeepSWE v1.1 coding benchmark, up from 49.0% for the prior Flash, and it carries a 1M-token context. Google did not release open weights.

Qwen3.8 is open, but open is not local

Alibaba went the other way and published the Qwen3.8 checkpoint on August 12: 2.4 trillion parameters, free to download. Before you plan to self-host it, note the hardware. Nvidia's day-zero serving test ran it on a rack of 72 Blackwell GPUs. Open weights at this scale are a licensing headline, not a home-lab option, and the open checkpoint is text-only. The managed sibling, Qwen3.8-Max (2.4T mixture-of-experts, roughly 95B active, 1M context, with vision), reached general availability on August 3 at $2 and $6 per million tokens, undercutting Gemini Flash's standard $7.50 output rate.

What it means for builders

The value tier is now a genuine fight, and Google's cut was itself a response to cheap Chinese models. OpenAI recently cut GPT-5.6 Luna by up to 80%, and models like Moonshot's Kimi K3, GLM 5.2, and Qwen3.8-Max are delivering near-frontier quality at a fraction of the cost. That competitive pressure, not generosity, is what put Flash on sale.

Two takeaways. The cheaper winner depends entirely on your input-to-output token mix, so test both on your actual workload instead of the headline rate. That is the same measure-first discipline behind the AI productivity proof gap, and it applies to model choice too. And whether you rent a closed model by the token or download open weights and rent the GPUs, at this scale you are still renting the rack.