DeepInfra

Live·11 models

The cheapest way to run large open-weight models at volume.

Lowest priceLarge open modelsEmbeddings

11 models on DeepInfra

DeepInfra's own rate for each of these. Where another provider serves the same model it may charge differently. The model page shows every route side by side.

Model	Price	Context
BGE-M3baai/bge-m3	$0.010 inFree out	8K
Llama 3.1 8B Instructmeta/llama-3.1-8b-instruct	$0.020 in$0.040 out	131K
Qwen3 32Bqwen/qwen3-32b	$0.080 in$0.280 out	131K
Llama 3.3 70B Instructmeta/llama-3.3-70b-instruct	$0.100 in$0.320 out	131K
DeepSeek V3.2deepseek/deepseek-v3.2	$0.260 in$0.380 out	164K
Qwen3 235B A22B Instructqwen/qwen3-235b-a22b-instruct	$0.090 in$0.550 out	262K
DeepSeek V3.1deepseek/deepseek-v3.1	$0.250 in$0.950 out	164K
DeepSeek R1 (0528)deepseek/deepseek-r1	$0.500 in$2.15 out	64K
DeepSeek V4 Prodeepseek/deepseek-v4-pro	$1.30 in$2.60 out	512K
Claude Haiku 4.5anthropic/claude-haiku-4-5	$1.00 in$5.00 out	200K
Claude Sonnet 5anthropic/claude-sonnet-5	$2.00 in$10.00 out	1M

Availability, from our vantage point

What DeepInfra looked like from Multigrid's servers, not DeepInfra's own uptime figure, which only they can publish.

Probes 99.89%Real traffic no traffic yet

The ring is reachability probes, measured across 100% of the last 30 days a share of what we watched, not of the month. We probe when someone loads this page, so the gaps are ours.

Succeeded
Failed on a degraded day
Failed on a bad day (< 50%)
Nothing measured yet

The two lanes are never averaged. Probes ask whether DeepInfra is answering at all; traffic is the share of real requests we routed there that completed. A provider can pass every probe while completing nothing, which is why they are stacked rather than blended, and why a green top lane over a red bottom one is the pattern worth looking for.

Coverage is ours, not theirs: we probe when someone loads the status page, at most once every five minutes, so an empty slot means we were not looking rather than that anything was wrong. For an authoritative account of an incident, read DeepInfra’s own status page. They are the only ones who can write it.

None of this is latency. What happened to your own requests (timing, errors and cost per provider) is on your analytics page, measured from traffic you actually sent.

How this provider is reached

Wire format: openai
Base URL: https://api.deepinfra.com/v1/openai
Route order: Ranked by price, per model: where another provider serves the same model more cheaply, that one is tried first. Your own key outranks every platform route.
Billing: We hold a platform key, so usage can be charged to your Multigrid balance. Your own key, if you add one, is still preferred over ours.

Data handling and compliance

Multigrid does not publish DeepInfra’s retention window, training terms or certifications. We hold no audit report for them, have no contract obliging them to anything beyond their ordinary API terms, and could not tell you if any of it changed tomorrow, so anything stated here would be hearsay in a document your own auditor may read.

Their own documentation is authoritative and stays current without us. What we can tell you is what Multigrid does with the request on our side, which is on the Trust Center.

DeepInfra docs Their status page Privacy policy Terms

Headquartered in US. That is the company’s jurisdiction, not where your request is processed — which is not something this gateway can pin or promise. See where your requests go.

Use in a policy

Prefer, or exclude, DeepInfra on any request. The slug is what the gateway matches on.

"provider": {
  "order": ["deepinfra"],
  "allow_fallbacks": true
}

Provider preferences docs

Reach DeepInfra on one balance

We hold a platform key for DeepInfra, so everything above is reachable on Multigrid credit: one balance across every vendor, at the prices listed. Signing up takes no card and commits you to nothing. If you already have a DeepInfra contract, bring that key instead and we take no percentage on the traffic.

Create an account Bring your own key