The 8-Credit Tier: Seven Open Models, One Price
DeepSeek V4 Pro, Kimi K3, GLM 5.2, Qwen 3.7 Plus, MiniMax M3, MiMo V2.5 and Tencent HY 3 all cost the same in FrameTide. When price is equal, you choose on character.
Seven models in FrameTide cost exactly 8 credits per message. They come from seven different labs, most ship open weights, and several post benchmark numbers that would have been frontier results a year ago.
When the price is identical, the usual comparison collapses. You are not asking which is better value — they are the same value. You are asking which one is shaped like your problem.
The seven
| Model | Lab | Notable | Context |
|---|---|---|---|
| DeepSeek V4 Pro | DeepSeek | 1.6T MoE, 49B active | 1M |
| Kimi K3 | Moonshot | 2.8T, largest open-weight release | 1M |
| GLM 5.2 | Z.ai | 744B MoE / 40B active, MIT | 1M |
| Qwen 3.7 Plus | Alibaba | multilingual breadth | — |
| MiniMax M3 | MiniMax | long context, mixed media | — |
| MiMo V2.5 | Xiaomi | 310B, natively omnimodal | 1.05M |
| Tencent HY 3 | Tencent | 295B MoE / 21B active, Apache-2.0 | 256K |
Three of them carry a million-token context at 8 credits. That is the single most underrated fact in this table — it means "cheap" and "reads a lot" are no longer opposites.
What each one is actually for
DeepSeek V4 Pro — the price-performance benchmark. A 1.6T mixture-of-experts firing 49B parameters per token, released April 24, 2026 alongside the V4 Flash we default to at 1 credit. Its SWE-bench Verified score is among the highest available from open weights. If you want one default for hard reasoning at this tier, start here.
Kimi K3 — the agentic specialist. Moonshot announced it on July 16, 2026 and published the full 2.8T weights ten days later, making it the largest open-weight release to date. It leads open models on frontend code specifically, and it is built for long tool-call pipelines. Reach for it when the job is a sequence rather than a question.
GLM 5.2 — the bilingual reasoner. 744B MoE with 40B active under an MIT licence, and consistently near the top of open-model intelligence rankings. Strong on tool use and, in particular, on Chinese-language reasoning, where it frequently outperforms models several times its price.
Qwen 3.7 Plus — the multilingual generalist. Alibaba's line has the broadest language coverage in this group. If your work spans several languages rather than two, this is the safe pick.
MiniMax M3 — the long-input and mixed-media model. MiniMax's speciality is workflows that feed the model large documents, audio, or video rather than short prompts. Choose it when the input is big and heterogeneous.
MiMo V2.5 — natively omnimodal. Released April 22, 2026 at 310B under MIT, with a ~1.05M context and native audio, image, and video input. Xiaomi built it as the model behind their device ecosystem, which shows in its bias toward fast, practical responses.
Tencent HY 3 — the efficiency play. A 295B mixture-of-experts firing only 21B parameters per token, Apache-2.0 licensed, released July 6, 2026. The smallest context here at 256K, but unusually capable at general reasoning for its active-parameter count, and notably good at image, 3D, and video tasks.
How to choose without testing all seven
Route by the shape of the input and the shape of the output:
- Long input, one answer → Kimi K3, DeepSeek V4 Pro, or GLM 5.2 for the 1M context. Or drop to Gemini 3.6 Flash at 5 credits if the reasoning is local.
- Mixed media in → MiniMax M3 or MiMo V2.5.
- Chinese-language reasoning → GLM 5.2 first, Qwen second.
- Many languages → Qwen 3.7 Plus.
- Long chains of tool calls → Kimi K3.
- Hard code you will review → DeepSeek V4 Pro.
- No strong signal → DeepSeek V4 Pro. It is the most general of the seven.
Why this tier matters more than the frontier
The gap between this tier and the 50–100 credit models is 6x to 12x. The capability gap is real but much smaller than the price gap, and it is concentrated in places that may not be yours: prose quality, judgement on ambiguous problems, and holding accuracy over very long unreviewed outputs.
For work you are going to read anyway — which is most work — the 8-credit tier is where the sensible default lives. The frontier models earn their price at the end of a task, on the parts nobody is going to check.
The honest caveat: none of these are the best model at anything in absolute terms. They are the best value at nearly everything, which is a different and usually more useful property.
A test worth running once
Take five real tasks. Run each on DeepSeek V4 Pro at 8 credits and on whichever 50-or-100-credit model you normally reach for. Review the pairs blind.
Most teams discover that the expensive model wins clearly on one or two of the five, and that those one or two share a characteristic — usually "nobody reads the output carefully" or "the answer is a judgement rather than a fact". That characteristic is your escalation rule. Everything else can live at 8 credits permanently.
The short version
Seven models, one price, seven different shapes. Default to DeepSeek V4 Pro, switch to K3 for agentic chains, GLM 5.2 for Chinese, Qwen for many languages, and MiniMax or MiMo when the input is not text. Then spend the 50-credit models only where you would not have checked the answer.
Sources
Create with DeepSeek V4 Pro
Open FrameTide Agent with DeepSeek V4 Pro already selected, and use it on your own creative task.