Run your company's AI with less cost and less risk.
Meridian sits between your teams and AI providers like OpenAI and Anthropic. It caches repeated requests, caps spend with budgets, masks personal data, and keeps your apps running when a provider fails.
Free credits to try it. No fees during the beta.
AI spend this month
$663.25
across 3 projects · 12,840 requests
| Support assistant | $412.3062% of budget |
| Internal search | $186.7537% of budget |
| Sales emails | $64.2043% of budget |
Why Meridian
One control point for all your AI traffic
Teams adopt AI faster than the controls around it. Meridian adds them without slowing anyone down.
Spend less
Repeated and similar questions can be answered from cache, and Meridian shows when a cheaper model would do.
No surprise bills
Budgets per project, enforced before every request, and cost by project and model.
Protect sensitive data
Personal data can be masked before it reaches the provider. Prompts are never logged.
Stay up when a provider fails
Requests can fall back to another model or provider, so one outage does not stop your product.
Cost insight
Find out if a cheaper model would do
Meridian reprices your recent traffic on every model, and flags those that cannot handle your largest requests.
18,200 answered requests used 24,000,000 input and 4,100,000 output tokens. At list prices, your current models come to $89.00; you were billed $79.20, after prompt-cache discounts.
Claude Haiku 4.5 would have cost about $44.50, 50.0% less than your current models.
| Model | Estimated cost | Compared with now |
|---|---|---|
| GPT-4o Miniopenai · your largest request does not fit its context window | $6.06 | 93.2% less |
| Claude Haiku 4.5anthropic | $44.50 | 50.0% less |
| Claude Sonnet 5.5currentanthropic | $89.00 | same |
| GPT-4oopenai · your largest request does not fit its context window | $101.00 | 13.5% more |
| Claude Opus 5.5anthropic | $178.00 | 100.0% more |
Estimates at each model's list price for the same token counts. Other providers' tokenizers count the same text differently, and another model may answer at a different length or quality, so treat these as a guide. Response-cache hits and failed requests are not included.
Requests per day
- Cost per day, model and project, for any period.
- Organizations, projects and members, with five roles from owner to viewer.
- A request log and an audit log of who changed what.
For developers
Change one URL. Keep your SDKs.
Meridian speaks the OpenAI and Anthropic APIs, so existing code, Claude Code and MCP clients work as they are.
from openai import OpenAI
client = OpenAI(
base_url="https://api.sdwares.com/v1", # point at Meridian
api_key="gw_your-project_…",
)
client.chat.completions.create(
model="claude-sonnet-5-5",
messages=[{"role": "user", "content": "Hello"}],
)- OpenAI Chat Completions, the native Anthropic Messages API, embeddings and /v1/models.
- API keys per project with per-minute and per-day limits, rotation with a grace period and instant revocation.
- Cost is reserved before each call, so parallel requests cannot overshoot a budget.
- Model aliases across providers, retries with backoff and a circuit breaker per model.
- Exact-match response cache, plus an optional semantic cache for similar requests; both per project.
- PII redaction per project, including streamed responses.
- Bring your own OpenAI and Anthropic keys, stored encrypted.
- An MCP server for Claude Desktop and other MCP clients.
- 1
Request access
Tell us what you are building. We approve requests by hand during the beta.
- 2
Create a key
Your workspace starts with a project. Create an API key; add your own provider keys when ready.
- 3
Change one URL
Point your client at Meridian. Budgets, limits and cost tracking apply from the first request.
Security and data
Your keys and your data stay yours
Encrypted keys
Your provider keys are encrypted and protected.
No prompt logging
We record usage, not content. Cached answers expire within 24 hours.
Secure sign-in
No passwords to leak.
Isolated by design
Each organization’s data is kept separate.
FAQ
Questions people ask
Which models can I use?+
OpenAI and Anthropic models from our reviewed price list, currently GPT-4o, GPT-4o Mini, Claude Haiku 4.5, Claude Sonnet 5.5 and Claude Opus 5.5. GET /v1/models returns the current list.
What does it cost?+
Nothing during the beta. New workspaces get free credits to try it; after that you use your own provider keys and the providers bill you directly. We will give 30 days’ notice before introducing fees.
Do you store my prompts?+
No. We keep request metadata for usage and billing. The only content we hold is the optional response cache (and, if you turn on the semantic cache, its embeddings), for up to 24 hours.
What happens when a provider is down?+
Requests are retried, a circuit breaker stops traffic to a failing model, and an alias can fall back to another model or provider.
Can I run it in my own cloud?+
Yes. The same code deploys to your Azure subscription with Terraform, so data and keys stay with you. Contact us.
Is there an SLA?+
Not during the beta. The first request after a quiet period can be slower, and features may change.
Closed beta
Request access
Tell us about your project. We will email you when your workspace is ready.
- · Free credits to try it, then your own provider keys
- · No fees during the beta
- · No uptime commitment while in beta (see the Terms)
Want it in your own cloud? support@sdwares.com
By requesting access, you agree to the Terms of Service and acknowledge the Privacy Policy.
Already have access? Sign in · Staff sign-in