Fixed-cost AI means you stop paying a provider per token and run open-weight models, such as Llama, gpt-oss, Gemma, Qwen, DeepSeek or GLM, on GPU capacity you reserve. Your bill then depends on how many GPUs you hold, not on how much people use them. Above a break-even volume, extra prompts and agent loops cost nothing more, and finance gets a number it can forecast.
Key takeaways
- Per-token pricing scales with usage, and agents multiply usage.
- Fixed cost means paying for GPUs or reserved instances; it wins above a break-even volume you can calculate.
- American and Chinese makers publish open-weight models under different licences: Apache 2.0, MIT and custom ones.
Why do per-token AI bills keep growing?
Cost is tokens times a rate, and nothing caps the tokens. A coding agent that reads files, calls tools and retries sends its growing context back to the model at every step, so one request can become dozens of calls (see agentic AI budget explosions). Teams then ration AI to protect the budget.
What does fixed-cost AI mean?
You pay for capacity, not consumption. The worked model below uses illustrative assumptions, not measured results and not any provider's prices. Swap in your own quotes and load tests.
| Assumption | Value |
|---|---|
| Blended API price | $3.00 per million tokens |
| Reserved GPUs | 8 at $2.50 per GPU hour, 730 hours: $14,600 a month |
| Throughput and utilisation | 2,500 tokens per second per GPU, 50% busy |
| Monthly capacity | about 26.3 billion tokens |
| Break-even | 14,600 / 3.00 = about 4.87 billion tokens a month |
| Monthly tokens | Per-token | Fixed | Fixed saves |
|---|---|---|---|
| 2 billion | $6,000 | $14,600 | nothing |
| 8.8 billion (200 developers, 2M tokens a day, 22 days) | $26,400 | $14,600 | 44.7% |
| 26 billion (near capacity) | $78,000 | $14,600 | 81.3% |
Below break-even, per-token is cheaper. Add operating time to your model.
Which open-weight models can you run, and under what licence?
As stated on each maker's own model card on 22 November 2026. Sizes are total parameters, with active parameters for mixture-of-experts models.
| Maker (country) | Model | Size | Licence as stated |
|---|---|---|---|
| Meta (US) | Llama 4 Scout, Maverick | 109B and 400B, 17B active | Llama 4 Community License Agreement |
| OpenAI (US) | gpt-oss-120b, 20b | 117B (5.1B active), 21B | Apache 2.0 |
| Google DeepMind (US) | Gemma 4 | E2B to 31B | Apache 2.0 |
| NVIDIA (US) | Nemotron 3 Super | 120B (12B active) | NVIDIA Nemotron Open Model License |
| Microsoft (US) | Phi-4 | 14B | MIT |
| Mistral AI (France) | Mistral Small 4 | 119B (6.5B active) | Apache 2.0 |
| Alibaba Qwen (China) | Qwen3.8-27B | 27B | Apache 2.0 |
| DeepSeek (China) | DeepSeek-V4 Pro, Flash | 1.6T (49B active), 284B (13B active) | MIT |
| Moonshot AI (China) | Kimi K3.1 Pro | 2.8T (104B active) | Kimi K3.1 Pro License |
| Z.ai (China) | GLM-4.7-Flash | 30B (3B active) | MIT |
Custom licences such as the Kimi K3.1 Pro License add conditions for large "Model as a Service" businesses. Read the full text for your use case.
How do you serve open-weight models for unlimited team use?
- Engines. vLLM manages the KV cache in pages (PagedAttention paper) so more requests fit per GPU. SGLang is a similar server; both offer OpenAI-compatible APIs. llama.cpp runs GGUF files on laptops and small GPUs.
- Quantization. FP8 halves memory against BF16; INT4 and 4-bit GGUF cut further. See choosing a GGUF quant level by memory budget.
- Batching and routing. Continuous batching keeps GPUs busy. Send routine work to a small model and hard reasoning to a large one.
Capacity planning is a load test on your real prompt mix. More in our cheapest LLM inference explainer.
How do you pick a model per task?
Card benchmarks are the maker's own runs. Build 100 to 300 real tasks per route, test each candidate at the quantization you will deploy, score quality and latency, and switch only when it beats your current baseline.
Are open-weight models safe to run?
Treat weights like packages, whichever country they come from.
- Format. Python's documentation says "the pickle module is not secure". Prefer safetensors. Background: our pickle deserialization analysis.
- Provenance and licence. Use the maker's official repository, pin the revision, record checksums and list each model in an AI-BOM. See Hugging Face model supply chain risks.
- Poisoned weights. Research shows hidden behaviour can live in weights (Weight Poisoning Attacks) and persist through safety training (Sleeper Agents). Your own evaluation is the check.
- Prompt injection (OWASP LLM01) still applies when you self-host.
- Data. Prompts and outputs stay on infrastructure you control.
Where does Safeguard fit?
Safeguard can run open-weight models, American and Chinese, inside your environment, so AI cost becomes a fixed capacity cost instead of a per-token bill. Companies we have worked with have cut AI cost by up to about 90% and now use AI across teams without usage limits.
AI modes are configurable per tenant: Full AI with Safeguard's own security models, Zero AI with no LLM in the loop while every scanner still works, bring your own model, and local and on-prem models. Customer code is not used to train shared models. Compare approaches in bring your own model vs bundled model pricing.
FAQ
What is fixed-cost AI?
Running models on capacity you pay for by the hour or month, so cost does not change with tokens used.
When is self-hosting cheaper than an API?
Above break-even: fixed monthly cost divided by your per-token price.
Are open-weight models free for commercial use?
Many are, under Apache 2.0 or MIT. Custom licences add conditions, so read each one.
Are Chinese models riskier than American ones?
Apply the same checks to all: official source, safetensors, pinned revision, licence review and your own tests.
Does self-hosting stop prompt injection?
No. Input handling and tool permissions still need controls.
Can Safeguard run with no AI?
Yes. Zero AI mode runs the platform with no LLM in the loop.
The serving stack you deploy ships open source dependencies too, so the usual supply chain hygiene applies.
Fix it without waiting for upstream
Whatever the vulnerability, most of the risk sits in two gaps. The first is the time between a fix being published and that fix actually running in your production builds. The second, and worse, is the time before any fix exists at all, while a maintainer works out a patch and a release schedule. Waiting through both gaps is a habit, not a requirement. Safeguard gives teams three ways to close them.
1. Autonomous remediation: pull requests, not tickets
When a fixed version exists, Griffin, Safeguard's security model, opens a policy gated pull request that moves the affected dependency to the safe version, and your own CI runs your tests on it before anyone merges. You can run it on one repository or in bulk across every project that carries the vulnerable version. Remediation categories include dependencies, licenses, security posture and findings, and each one is set separately. Reachability analysis ranks the vulnerable code paths your application actually calls first, so the first pull requests you merge are the ones that matter most. Auto-merge is a separate switch and stays off by default, so your team decides what merges on its own and what still needs a human approval.
2. Gold Open Source: the safe version in seconds, free
gold.safeguard.sh is free and needs no login. Search any CVE or package to see affected and fixed versions, CVSS, EPSS, CISA KEV status and a recommended safe version, or use the "Am I affected?" checker. The same data is available through a free read API, README status badges and RSS threat feeds.
3. Gold Registry: start from components that are already fixed
Safeguard Gold rebuilds patched versions of open-source packages and images from source, signs each build, attaches an SBOM, SLSA provenance and VEX, and serves them from registry.safeguard.sh to your organization. Every Gold artifact is held continuously to zero exploitable critical or high CVEs and a completed license check, with signed build provenance (signed with Safeguard's key and verifiable with cosign, SLSA Build Level 3 as the design target) and a public evidence page you can hand to an auditor. An artifact loses its Gold badge if it regresses. Teams can point their own package directory at the registry so hardened equivalents resolve without changing how developers install anything.
When upstream has no fix yet, or will not ship one, Safeguard first works to get the fix accepted upstream. If the issue is severe or upstream will not move, Safeguard builds a fork, fixes the vulnerability there and publishes a hardened Gold build, so you are not stuck on a maintainer's release schedule. You can also request a hardened build of a specific package. Ask which packages it covers before you plan around it.
| Your situation | What to use |
|---|---|
| A fixed version exists | Autonomous remediation opens the upgrade pull request; Gold Open Source confirms the safe version |
| You need to know today whether you are affected | Gold Open Source lookup and "Am I affected?" checker, free and without login |
| You want vulnerable versions to stop shipping | Safeguard policies and gates in CI |
| Upstream has no fix yet, or will not ship one | A custom Gold request; fork and patch |
| You want to start clean on new projects | Pull hardened components and images from Gold Registry |
One honest limit: when the vulnerable code lives in a vendor appliance, a hosted service or an operating system runtime you do not build yourself, the fix still has to come from that vendor or distribution. The steps above apply to the open source dependencies and container images you build and ship, which is where most teams carry the largest and least tracked share of their exposure.
See how Safeguard would help
This demo sets up the runner on your own machine, approves it with its device code, then scans from the terminal, the same flow you would use on the hosts that serve your models.
Click through it at your own pace, or open it full screen. Get started at safeguard.sh/download and the CLI docs.
Sources
Model cards are linked in the table. Also: vLLM quantization, Python pickle.
Last reviewed: 22 November 2026.