GLM-5.2 on AMD MI355X: Cost Impact
Wafer reports GLM-5.2 on AMD MI355X at 2,626 tok/s/node and over 2x lower cost than Blackwell. Here is the pricing impact.
By AI Pricing Guru Editorial Team
AI Pricing Guru articles are maintained by the editorial workflow behind the site: daily pricing snapshots, provider source checks, and review passes for model launches, subscription limits, and billing changes.
Wafer published a useful GLM-5.2 infrastructure result on July 3: the team says it served GLM-5.2 on AMD MI355X at 2,626 tokens per second per node on a cached long-context workload, while claiming the hardware route is over 2x cheaper than Blackwell.
This is not a Z.ai API price cut. GLM-5.2 still sits in our live pricing data at $1.40 per million input tokens, $0.26 per million cached input tokens, and $4.40 per million output tokens. The pricing impact is about what happens behind the API bill: if AMD serving gets easier and faster for frontier open-weight models, hosted providers can put more pressure on Nvidia-backed inference margins.
For current model rates, keep the Z.ai pricing page, Together AI pricing page, OpenAI pricing page, and AI token calculator open. For the broader GLM-5.2 buyer context, read our earlier GLM-5.2 API pricing breakdown and AI API pricing comparison.
What Changed
Wafer’s post, “Performance per dollar is getting faster and cheaper,” describes a GLM-5.2 deployment on AMD MI355X capacity from TensorWave. The headline numbers are:
| Metric | Wafer result |
|---|---|
| Model | GLM-5.2 |
| Hardware | AMD MI355X |
| Aggregate throughput | 2,626 tok/s/node |
| Workload | 20K input / 1K output, 60% cache hit rate |
| Saturation point | 2.4 requests per second |
| TTFT at saturation | 0.81s p50 / 2.22s p95 |
| Single-stream decode | 213 tok/s |
| Claimed cost angle | Over 2x cheaper than Blackwell |
Wafer also says MI355X is around 2.75x cheaper per GPU on average than Nvidia B300 while offering comparable hardware specs. The company reports that its GLM-5.2 MI355X result reached roughly 80% of measured B200 performance on the same aggregate workload.
Those details matter because GLM-5.2 is an open-weight model with a 1M-token context window. Open weights create the option to self-host or use specialized inference providers, but only if the serving stack can run the model efficiently. Wafer’s result is a sign that AMD support for large open models is getting less theoretical.
Pricing Comparison
The public token price for GLM-5.2 has not changed. The live buyer comparison still looks like this:
| Model | Input | Cached input | Output | Buyer read |
|---|---|---|---|---|
| GLM-5.2 | $1.40 / 1M | $0.26 / 1M | $4.40 / 1M | Mid-priced open-weight long-context route |
| DeepSeek V4 Pro | $0.435 / 1M | $0.003625 / 1M | $0.87 / 1M | Cheaper raw token floor for coding and reasoning |
| Qwen3.7 Max via Together | $1.25 / 1M | $0.13 / 1M | $3.75 / 1M | Hosted open-model alternative |
| GPT-5.4 | $2.50 / 1M | $0.25 / 1M | $15.00 / 1M | Premium closed model at lower input price than GPT-5.5 |
| Claude Sonnet 5 | $2.00 / 1M | $0.20 / 1M | $10.00 / 1M | Main Claude coding route |
| GPT-5.5 | $5.00 / 1M | $0.50 / 1M | $30.00 / 1M | Premium OpenAI route |
GLM-5.2 remains much cheaper than GPT-5.4, GPT-5.5, and Claude Sonnet 5 on output tokens. It is still not as cheap as DeepSeek V4 Pro, but its 1M context window, open-weight control, and improving serving story make it more interesting for long-context coding agents.
The Wafer result changes the infrastructure side of the discussion. If a provider can serve GLM-5.2 on AMD at a lower hardware cost while preserving enough throughput, it has more room to discount, absorb traffic spikes, or offer subscription-style coding plans without burning as much margin.
Cost Example
Take a long coding-agent workload with:
- 1M fresh input tokens
- 4M cached input tokens
- 500K output tokens
At current listed token prices:
| Model | Estimated API cost |
|---|---|
| DeepSeek V4 Pro | ~$0.88 |
| Qwen3.7 Max via Together | ~$3.65 |
| GLM-5.2 | ~$4.64 |
| Claude Sonnet 5 | ~$11.70 |
| GPT-5.4 | ~$11.00 |
| GPT-5.5 | ~$22.00 |
This is why performance-per-dollar serving matters. GLM-5.2 is already priced well below premium closed models for output-heavy work. If AMD hosting keeps improving, the model can become a stronger middle route between ultra-cheap DeepSeek traffic and premium Claude or OpenAI escalation.
What This Means
For buyers, the immediate takeaway is not “move everything to AMD.” Most teams buying AI APIs will never touch the GPU layer directly. The takeaway is that infrastructure competition can show up later as better API pricing, more stable capacity, and more credible hosted open-weight routes.
For providers, the result is a margin story. If Nvidia capacity stays scarce and expensive, providers that can serve large open models on AMD get another lever. They can target high-volume coding-agent traffic without depending entirely on Blackwell availability.
For self-hosting teams, the result is encouraging but still not a shortcut. Wafer’s post describes quantization, SGLang changes, speculative decode fixes, kernel tuning, and topology changes from TP8 to TP4 x DP2. That is real engineering work. GLM-5.2 may be open-weight, but high-throughput serving is not plug-and-play.
Practical Advice
API buyers should keep benchmarking GLM-5.2 against Claude Sonnet 5, GPT-5.4, GPT-5.5, DeepSeek V4 Pro, and hosted Qwen routes on accepted-task cost. Do not route based only on list price.
Teams with enough volume to negotiate should ask providers how they handle capacity, cache pricing, and peak-load routing. A provider with efficient AMD serving may be able to offer better terms for predictable long-context workloads.
Self-hosting teams should benchmark hosted GLM-5.2 first. If the API does not beat your current route on quality and accepted-task cost, infrastructure tuning will not save the project. If it does beat your current route and the volume is high enough, AMD capacity becomes a serious procurement question.
Bottom Line
Wafer’s GLM-5.2 on AMD MI355X result is a pricing story because it attacks the cost base behind model serving. The public GLM-5.2 token rate is unchanged at $1.40 input, $0.26 cached input, and $4.40 output per million tokens, but the infrastructure story is moving in buyers’ favor.
If AMD serving continues to close the gap with Blackwell for large open-weight models, GLM-5.2 and similar models could pressure premium API prices from below: not just by being open, but by becoming cheaper to serve at useful throughput.
Sources: Wafer: Performance per dollar is getting faster and cheaper, Z.ai pricing docs, Z.ai GLM-5.2 technical post, and AI Pricing Guru’s live pricing dataset.