Already running DeepSeek or GLM? We serve DeepSeek V4 Flash and GLM-5.3 Flash behind an OpenAI- and Anthropic-compatible API, and output costs the same as input. Change the base URL and the key. The rest of your code stays.
base_url="https://proxy.trustedrails.com/v1"
Want to test on your own traffic first? Ask for free test tokens. No card needed.
Pricing
Set your monthly volume and output share. The default mix is three input tokens for every output token, and every price is a list price per 1M tokens.
Z.ai sells GLM-5.3-Flash, the same model we serve, at $0.15 in and $0.50 out. DeepSeek now sells V4.1 Flash in place of V4 Flash, at $0.15 in and $0.60 out off-peak. We charge $0.10 in and $0.10 out.
Input costs the same, $0.10. Output costs $0.50 there and $0.10 here, five times less. They come with a DPA and SOC 2. We don’t have a DPA yet.
Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6.1 Sol are in the table for reference. Keep your hardest work on them.
Talk to us
Book 30 minutes. Tell us which models you run, how many tokens a month and what worries you about switching. If a test makes sense, you leave with free test tokens and a plan for the first 5–10% of traffic.
No slot that works? Write to andrii@trustedrails.com and we’ll find a time.
Quality
Benchmark scores. Higher is better.
Ours, reasoning onDefault settingsMax effort
| Model | Intelligence Index10 evaluations in one score | AA-LCRReasoning over long documents | AutomationBench-AAAgent tasks with SaaS tools | Terminal-Bench 4.0Agent tasks in a terminal | List price per 1M, in / out |
|---|---|---|---|---|---|
| GLM-5.3 FlashOn TrustedRails | 42 | 80.0% | 60.4% | 32.8% | $0.10 / $0.10 |
| DeepSeek V4 FlashOn TrustedRails | 34 | 79.7% | 54.0% | 12.1% | $0.10 / $0.10 |
| Claude Haiku 5.5 | 3443 | 82.7% | 35.4% | 32.8% | $0.10 / $0.50 |
| GPT-6 Luna | 3038 | 83.3% | 53.2% | 12.6% | $0.10 / $0.50 |
| GPT-6.1 Sol | 4852 | 83.0% | 64.9% | 56.1% | $2 / $10 |
| Claude Sonnet 5.5 | 4756 | 82.7% | 71.8% | 63.6% | $2 / $10 |
| Claude Opus 5.5 | 5158 | 84.7% | 69.5% | 59.6% | $4 / $20 |
Which of our modelsStart with GLM-5.3 Flash. Pick DeepSeek V4 Flash when the answer is in the context you send, like extraction or a summary of a document. Answering from memory, it makes things up more often.
Default settings and max effortClosed models show two scores. At max effort, Haiku 5.5 and GPT-6 Luna think longer, and thinking is billed as output: $0.50 per 1M there, $0.10 with us. Our models were scored with reasoning on.
Source: Artificial Analysis.
Send 5–10% of one workload through us and compare the answers with what you get today.
How a test worksWe checked it ourselves
We ran MMLU-Pro through our own API. GLM-5.3 Flash scored 85.3% and DeepSeek V4 Flash 84.5%, within the margin of the published results. The samples are small, so a gap of a few points cannot be ruled out.
Read the Quality CheckFor larger teams
How to start
Nobody moves a whole workload on day one. A test usually goes like this.
The first 10 teams that test with us get free test tokens and a direct line to our CTO. What they tell us helps decide what we build next.
Book a callTell us what you run, how much, and what you need from a provider. We tell you straight whether a test makes sense.
We add test tokens to your account. No card needed.
Keep the OpenAI or Anthropic SDK you use today. Point a slice of one workload at TrustedRails by changing the base URL, the key and the model name. Your retries and router stay as they are.
import osfrom openai import OpenAIclient = OpenAI( base_url="https://api.your-provider.com/v1", base_url="https://proxy.trustedrails.com/v1", api_key=os.environ["PROVIDER_API_KEY"], api_key=os.environ["TRUSTEDRAILS_API_KEY"],)reply = client.chat.completions.create( model="deepseek-v4-flash", model="deepseek-ai/DeepSeek-V4-Flash-0731", messages=messages,)Look at price, latency and the answers themselves, next to your current provider.
Shift more traffic where the results hold, and keep the hard tasks on frontier models. Moving back is the same three-line change.
Anything that takes a base URL and an API key will connect.
Roadmap
If a missing piece blocks your switch, tell us on the call.
Models
Open-weight releases, served in FP8. Call each one by its full model ID.
Model
Start here. Agent steps and tool calls, classification, and tasks that need some reasoning. Thinking is on by default.
Model
Extraction, summaries and answers over a document you send. Without context it makes things up more often, so don’t rely on it to answer from memory. Thinking is off by default; you turn it on per request.
Model
Always thinks before it answers. Its thinking can’t be turned off.
Embeddings for search and RAG, in many languages.
Not here yetImage input, JSON mode and a batch API. GLM-5.3 Flash reads images in its original release, but here it runs text only.
Streaming and tool calling work on all three chat models, and reasoning tokens are billed as output at the same price. If prices change, you hear it from us first. Model reference
Straight answers to what security teams ask first.
DeepSeek V4 Flash and GLM-5.3 Flash are open-weight models. They run on GPUs in partner data centers, not on DeepSeek’s or Zhipu’s API.
Our API doesn’t store prompts or completions. We keep token counts for billing.
You call the published open-weight releases under their own model IDs, served in FP8.
Use it for work without personal or regulated data. Ask your security questions in writing, and we answer in writing.
We serve the open-weight releases published under these model IDs, in FP8. If you call DeepSeek’s own API today, you get V4.1 Flash there, a newer version than our V4 Flash. Keep long multi-step reasoning on Claude Opus or GPT. Classification, extraction, summaries and support replies are the work these models are for. Run your evals on a sample of your own calls. Ask for test tokens.
Z.ai lists GLM-5.3-Flash at $0.15 per 1M input and $0.50 per 1M output. DeepSeek’s own API no longer serves V4 Flash. It serves V4.1 Flash at $0.15 input and $0.60 output off-peak, and twice that at peak hours. We bill $0.10 for input and for output, with one key for both models and a spend cap on each key. DeepSeek’s cache hits cost less than ours, so if most of your prompt repeats, compare at your cache rate.
Input costs the same, $0.10 per 1M. Output costs $0.50 per 1M there and $0.10 here, so the gap grows with long answers and reasoning. Claude Haiku 5.5 also costs more for prompts over 100K tokens. On the benchmark index, GLM-5.3 Flash scores level with Haiku 5.5 and above GPT-6 Luna. Both come with a DPA and SOC 2. We don’t have a DPA yet, so if you need one today, they are the safer choice.
In code, three lines: the base URL, the key and the model name, in the OpenAI or Anthropic SDK you already use. The real work is checking quality, so send 5–10% of one workload first and keep the rest where it is. Prompts written for Claude sometimes need small edits.
Image input, JSON mode and structured outputs, the batch API, the Images API, the Audio API, the Responses API, the Assistants API and fine-tuning. The models are text only, and output is capped at 16K tokens per request. JSON mode and image input are on the roadmap.
On GPUs in partner data centers, not on DeepSeek’s or Zhipu’s API. We don’t store prompts or completions. There is no DPA yet, so keep personal and regulated data out.
Errors come back in the standard OpenAI format, so the retries in your client keep working. Output is capped at 16K tokens per request. There is no fallback to another provider yet, so keep a second provider in your router as a fallback.
It takes 30 minutes on Google Meet. We go through the models you run, your monthly volume and what you need from a provider. If a test makes sense, we set up free test tokens and agree which 5–10% of traffic to send first. Pick a time.
TrustedRails is operated by Woodenshark LLC, a Delaware company named in the Terms of Service. We’re a small team: write to info@trustedrails.com and the people who build the API answer.
Yes. Bring a client and get 20% of what they pay, every month, for as long as they pay. We agree the terms in writing before your first client goes live. Ask for partner terms.
Start small: create a key, set a daily cap and send 5–10% of one workload through it. Compare a day of results.
Prefer to test first? Ask for test tokens.
