@OfficialLoganK and @GoogleAIStudio recently shipped gemini 3.6 flash – their new workhorse. key facts:
• 17% fewer output tokens than 3.5 flash on the artificial analysis index, up to 65% on deepswe, at a lower price – $1.50/1m in, $7.50/1m out
• step up in coding, knowledge work and computer use (osworld-verified 83.0 vs 78.4); fewer reasoning steps and tool calls per task
• pitched as the efficiency sweet spot for scaling agents, not the frontier
the other side of the ring: @alibaba_qwen's qwen3.8-max-preview, unveiled july 19 at waic. 2.4t params, and in alibaba's own release the team calls it "compatible with leading frontier models, second only to fable 5"
one bets on speed and price, the other on raw capability. so we gave both the same hard job – rebuild a real building in 3d from a single photo
every model run through @aimlapi, our partner platform – one api, all the frontier labs in one place
our test – 3 prompts, single-file html, procedural @threejs via cdn, no assets. the model gets one photo and a spec, and has to return an explorable scene: massing, materials, glazing, interior, orbit, day/dusk:
1. ligar bay bach – nz cantilevered cliffside house, blonde timber + black steel + glass
2. blackwood ridge (the sleeve house) – charred-timber wedge with a raking prow
3. fallingwater – frank lloyd wright's concrete trays over a waterfall
results (summed across all 3 houses):
- cost
#1 qwen3-max – $0 (free preview)
#2 gemini 3.6 flash – $0.61
#3 kimi k3 – $1.07
#4 gpt 5.6 sol – $1.75
- tokens
#1 gpt 5.6 sol – 53,889
# qwen3-max – 65,875
#2 gemini 3.6 flash – 69,486
#4 kimi k3 – 170,803
- lines of code
#1 qwen3-max – 2,753
#2 kimi k3 – 2,793
#3 gpt 5.6 sol – 3,395
#4 gemini 3.6 flash – 3,370
overall: no single winner – kimi builds the best-looking scene, qwen the most detailed one for free and on the fewest lines, gpt the highest fidelity but at ~3x the price, and gemini flash the cheapest that still holds up
so we read the code
qwen3-max – the tidiest file of the four. zero external assets: procedural canvas textures for every surface and hand-written glsl shaders for the waterfall + pool. bakes its own pmrem env map in-scene, lerps day/dusk per-frame, and instances the forest with rejection sampling that carves a clearing around the house. the most economical programmer here – fewest lines, most packed into them
gemini 3.6 flash – leaner, but real three.js underneath: correct cantilevers, red mullions, working massing. the headline is that a flash-tier model writes this at all. 3.5 flash gave you threadbare boxes; 3.6 gives you a building. thinner material and lighting layer than qwen, but no longer reads as a budget model
our take:
• kimi k3 is the top designer of the group – better eye than qwen3.8, best composition and lighting on load
• qwen3.8 is the surprise: error-free, fast, plenty of detail – and it does it on the fewest lines of code. efficient and clean at the same time
• gemini 3.6 flash is the story for the labs – flash models are no longer a cheap tier you settle for. they're starting to compete with the full-size models outright
We’re rolling out three new models to make AI agents faster, smarter, and cheaper at scale:
— Google DeepMind (@GoogleDeepMind) July 21, 2026
🔵 Gemini 3.6 Flash: It uses fewer tokens than 3.5 Flash to deliver higher quality work at the exact same cost.
🔵 Gemini 3.5 Flash-Lite: A fast, cost-effective option for everyday tasks… pic.twitter.com/SP1eU3FCW5
Nick Trenkler