OpenCode Updated usage/models (28/8/2026)
The GOATs have arrived.
Both very affordable and better than DeepSeek V4 Pro and GLM-5.2 (if benchmarks mean anything at this point).
- (New) GLM-5.3-Flash: competitive with GPT-5.6 Sol (high) with bonus vision! even at only $15 of usage, it's still affordable
- (New) Qwen3.8 Flash: competitive with GPT-5.6 Terra (max) and also with vision. $30 usage from the get go (and fast)
- (New) Hy4 Preview: Tencent's first frontier model, needs more work but looks very capable already. $30 usage
| Model | 5HOURS | 1WEEK | 1MONTH | CREDITS |
|---|---|---|---|---|
| GLM-5.3-Flash | 1580 | 3950 | 7900 | $15 |
| Qwen3.8 Flash | 5400 | 13500 | 27,000 | $30 |
| Hy4 preview | 1350 | 3380 | 6770 | $30 |
Remember how happy we were with DeepSeek V4 Flash 0731? this inference and capability acceleration of local-friendly open weights models is a little scary..
I'm personally most impressed by Qwen3.8-Flash-Next, this has a new architecture meant to power Qwen4 that allows it to partially utilize SSD storage to offload about 30% of its parameters (n-gram weights). Even though this model is supposed to be for preview/research, it's already capable enough to beat GPT-5.6 Terra (xhigh). I predict the first Qwen4 Max will at least match Opus 5!
Imagine a new 27B or 35B using this same architecture..! only one way to go from here 🏹
1 reply
Beware of Hy4 Preview... they weren't joking about its "known limitations" 😅
I tested it with the single prompt to build a solar system model in three.js, it spent many millions of in/out tokens testing, debugging, and verifying every single assumption and even every single calculation.. it even calculated and enforced accurate orbits for each celestial object, optimized the rendering, and procedurally built a milky way background 🤯 this model is just too much in its current state. I ran this test on almost every model out there, never seen any LLM go this far from a single prompt.