DeepInfra Pricing in 2026: Every Model Priced, Verified

DeepInfra prices over a hundred text models per token from $0.02 per million input tokens on Meta-Llama-3.1-8B to $2.85 on Moonshot AI's Kimi K3 as of 2026-10-13, and on the shared gpt-oss-120b checkpoint its $0.05 blended rate is the second cheapest of the 17 providers third-party benchmarking firm Artificial Analysis measures. The trade is speed: DeepInfra's measured 37 tokens per second on that same checkpoint sits far below the custom-silicon hosts, and its privacy posture is a policy document, not an independent attestation.

Key facts

• Price floor: Meta-Llama-3.1-8B-Instruct-Turbo lists at $0.02 input / $0.04 output per million tokens — the cheapest first-party row DeAI verified on any major open-model host (DeepInfra pricing, accessed 2026-10-13).
• Current-frontier rows: DeepSeek V4 Flash (0731) at $0.06 / $0.18 (cached input $0.015), DeepSeek V4 Pro at $1.30 / $2.60 (cached $0.10), Moonshot AI's Kimi K3 at $2.85 / $14.25 (cached $0.285), Z.ai's GLM-5.3 at $0.563 / $2.50 with a 38% promo discount (DeepInfra pricing).
• Cheapest cache: Xiaomi's MiMo-V2.6-Flash bills cached input at $0.003 per million against $0.14 uncached — the lowest cached-input rate DeAI has seen on any rate card (DeepInfra models, accessed 2026-10-13).
• Speed (independent): Artificial Analysis's 72-hour medians measured DeepInfra at 37.1 output tokens/s on gpt-oss-120b (high) — versus Cerebras 1,686.4, SambaNova 707.6, Groq 475.2 — and at $0.05 blended price, second only to CoreWeave ($0.04) among 17 providers (Artificial Analysis, accessed 2026-10-13).
• Rate limits: every account gets 200 concurrent requests per model — a concurrency cap rather than a daily token budget (DeepInfra rate limits).
• Footing: DeepInfra closed a $107 million Series B in May 2026 co-led by 500 Global and Georges Harik with NVIDIA, Samsung Next and Supermicro participating, and says token volume grew 25x since its Series A — a company claim, not an audited figure (DeepInfra Series B announcement).

What DeepInfra is

DeepInfra is a Palo Alto-based inference cloud that hosts more than 150 open-source models behind an OpenAI-compatible chat-completions endpoint at https://api.deepinfra.com/v1, plus an Anthropic Messages-compatible surface, batch processing at 20% off, dedicated GPU deployments, and per-hour B200 rentals. It is vertically integrated by choice: the company says it owns and operates its GPU fleet across eight US data centers rather than renting capacity from hyperscalers, and it co-develops with NVIDIA — early Blackwell deployments and Vera Rubin with the Dynamo inference stack, which the company credits with up to 20x cost-efficiency gains, a vendor claim.

The May 2026 Series B ($107 million, co-led by 500 Global and Georges Harik, with NVIDIA, A.Capital Ventures, Crescent Cove, Felicis, Peak6, Samsung Next, Supermicro, and Upper90 participating) funds global capacity expansion. The company's stated thesis is that agentic workloads — a single task making 50 to 100+ model calls continuously — are the baseline inference workload, and that its own stack is engineered for exactly that pattern.

The full rate card, priced (2026-10-13)

Everything below is DeepInfra's first-party list rate, accessed 2026-10-13. This is a selection of the most-requested rows, not the whole catalog — the full card spans 100+ models including embeddings at $0.005 per million, image generation billed per megapixel, and Whisper transcription.

Open-weight frontier and near-frontier

Open-weight workhorses

Closed-model passthrough rows

DeepInfra also resells Anthropic and Google frontier models from the same API surface — a category none of the other open-model hosts in this series carries:

Two structural notes on this card. First, the spread: input rates span roughly 140x from Llama 3.1 8B to Kimi K3, and output 350x — this is a per-model rate card, not a size-tiered one, so the checkpoint you pick matters more than the provider you pick. Second, the discount flags: DeepInfra actively marks down GLM-5.3 (38% off) and GLM-5.3-Flash (50% off) against what appear to be their own prior rates, and runs a 30% promotion on DeepSeek-V4.1-Flash. Promotional rows are the first to change — date-stamp any comparison.

The cache economics are the real story

DeepInfra's cached-input rates are among the steepest discounts in the market, and for agentic workloads — where the same long system prompt and conversation prefix is re-sent on every call — they matter more than headline input prices:

• MiMo-V2.6-Flash: $0.003/M cached vs $0.14 uncached — a 47x discount.
• DeepSeek-V4-Flash-0731: $0.015/M cached vs $0.06 — a 4x discount.
• Kimi-K3: $0.285/M cached vs $2.85 — a 10x discount.
• DeepSeek-V4-Pro: $0.10/M cached vs $1.30 — a 13x discount.

DeepInfra also offers prompt-cache retention — holding a prefix in cache for 5 minutes or 1 hour so reuse across calls skips prefill entirely and bills at the cache-read rate — which converts the discount from opportunistic to contractual for loop-shaped agent work. At Artificial Analysis's default 7:2:1 cache-input-output blend, DeepInfra lands at $0.05 per million tokens on gpt-oss-120b; on genuinely cache-heavy agentic mixes the effective rate drops further.

Third-party context for one shared checkpoint: on gpt-oss-120b DeepInfra lists $0.04 input / $0.17 output per million tokens per Artificial Analysis's provider table, against CoreWeave at $0.03/$0.17 (the only host that undercuts it on that blend), Novita at $0.05/$0.25, and Groq at $0.15/$0.60. This is where DeepInfra's reputation as the budget default comes from — and why our cheapest-LLM-API roundup has repeatedly cited it as the blended-price leader.

Speed: the other side of the trade

The same benchmarking shows what the discount buys you. On gpt-oss-120b (high) with 10,000-token inputs, Artificial Analysis's 72-hour medians measure DeepInfra at 37.1 output tokens per second — roughly 45x slower than Cerebras (1,686.4), 19x slower than SambaNova (707.6) and 13x slower than Groq (475.2) on the identical checkpoint. DeepInfra runs commodity H100/A100 fleets optimized for throughput and cost rather than single-stream latency, and it shows.

There is an escape hatch: a "DeepInfra Turbo" variant of gpt-oss-120b appears in Artificial Analysis's provider list as a distinct, faster measured endpoint at a higher price, and the catalog generally offers Turbo-class variants (Llama 3.3 70B Turbo, Qwen3-Coder Turbo) for the checkpoints where latency matters. For user-facing streaming output, benchmark the custom-silicon hosts — our Cerebras vs Groq and SambaNova Cloud pages carry the measured numbers. For batch, background agents, and anything throughput-tolerant, the per-token discount usually dominates: at 37 tokens/s a 500-token response takes about 13 seconds, which no batch job and few agent pipelines will notice.

Privacy: a policy, not an attestation

DeepInfra's data-privacy documentation makes specific commitments: input is not stored to disk during inference (memory only, deleted on completion), outputs are deleted after delivery, request content is not logged (only request ID, cost, and sampling parameters), and submitted data is not used for training or shared with third parties. The trust center lists SOC 2 Type 2, ISO 27001, and HIPAA compliance, with GDPR marked in progress.

Read the exceptions before relying on any of it: for Google models, Google stores outputs and logs prompts per its own policy; for Anthropic models, Anthropic's retention terms apply; bulk/batch inference may be stored encrypted on disk briefly; and image-generation outputs persist for a short period. Every one of these is a self-published policy statement — DeAI has not found an independent, contractually verifiable attestation behind the no-retention claim, and we treat "zero data retention" marketing labels the same way across all providers. If prompts must stay confidential by construction rather than by policy, the TEE-attested hosts (Chutes and similar) are the different category; our zero-retention explainer covers how to tell the two apart.

Rate limits: generous, and concurrency-shaped

DeepInfra's limit is 200 concurrent requests per model per account — not requests per minute and not a daily token cap. The docs' own arithmetic: at a 1-second average request duration that is ~12,000 requests per minute; at 60-second requests it is 200 per minute. Two models queried simultaneously means 400 concurrent slots. There is no published minimum spend and no daily token ceiling like SambaNova's 20-million-token cap; increases are requested from the dashboard rather than negotiated with sales. For bursty agent traffic, concurrency limits are generally the cheaper constraint to live with — a token-bucket around 200 concurrent requests is straightforward — but auto-scaling lag can still produce transient 429s under load.

How to choose

• Cheapest per token on a current frontier model, period: DeepInfra's DeepSeek V4 Flash row at $0.06/$0.18 (or GLM-5.3-Flash at $0.075/$0.25 during the promo) is hard to beat among first-party hosts.
• Cache-heavy agentic workloads: the MiMo-V2.6 rows' sub-cent cached rates and the 1-hour cache-retention option are the differentiator; rerun your own blend before concluding.
• Need closed models in the same account as open ones: DeepInfra is the only host in this series carrying Claude and Gemini rows alongside open weights — useful for eval harnesses that A/B across both.
• User-facing streaming latency: the fast-silicon hosts win on measured speed; pay the premium only if someone is watching the stream.
• Confidentiality by construction: TEE-attested providers, not DeepInfra's policy page.

Switching to DeepInfra is a base-URL swap from any OpenAI-shaped client:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepinfra.com/v1/openai",
    api_key=os.environ["DEEPINFRA_API_KEY"],
)

resp = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V4-Flash-0731",  # exact slug varies by provider
    messages=[{"role": "user", "content": "ping"}],
)

Model slugs are namespaced here (deepseek-ai/DeepSeek-V4-Flash-0731, moonshotai/Kimi-K3) and are not portable across providers; sampling defaults and quantization (fp8 is standard on the big MoE rows, fp4 on some GLM rows) differ per endpoint and show up in eval scores, not in the price table. A cheaper model that fails your evals is not cheaper after retries.

Current state (October 2026)

The budget tier of open-model inference remains a commodity-GPU price war, and DeepInfra is currently the price-setter DeAI can verify: lowest first-party floor on most shared checkpoints, near-lowest blended price measured by third parties, and the market's steepest cache discounts. Its vulnerabilities are the mirror image — measured single-stream speed far behind custom silicon, a privacy posture built on its own documentation rather than attestation, and a promotional pricing style (sitewide "30–50% off" banners) that makes its rate card the most churn-prone of any host we track. DeAI's Price Index re-shoots this spread weekly precisely because of that churn. For the same menu from other incumbents' angles, see our Groq alternatives, Together AI alternatives, and Fireworks AI alternatives pages. All prices here are dated 2026-10-13 and will move.

FAQ

How much does DeepInfra cost per million tokens?

As of 2026-10-13, DeepInfra's first-party rate card spans $0.02 input / $0.04 output per million tokens on Meta-Llama-3.1-8B-Instruct-Turbo up to $2.85/$14.25 on Kimi K3. Current-frontier rows: DeepSeek V4 Flash $0.06/$0.18, DeepSeek V4 Pro $1.30/$2.60, Kimi K2.6 $0.75/$3.50, GLM-5.3 $0.563/$2.50. Cached-input rates run as low as $0.003 per million on MiMo-V2.6-Flash.

Is DeepInfra cheaper than Together AI or Fireworks AI?

On the checkpoints both serve, DeepInfra's list rates sit below Together and Fireworks: DeepSeek V4 Pro is $1.30/$2.60 per million tokens on DeepInfra versus $1.74/$3.48 reported for Fireworks and $2.10/$4.40 for Together as of mid-2026. Third-party blended rankings put DeepInfra at $0.05 per million tokens on gpt-oss-120b — second cheapest of 17 measured providers. Rates move weekly; compare on the day you buy.

What are DeepInfra's rate limits?

DeepInfra limits every account to 200 concurrent requests per model — a concurrency cap, not a per-minute or daily token cap. At a 1-second average request that is roughly 12,000 requests per minute; at 60-second requests it is 200. Increases are requested from the dashboard. There is no published minimum spend.

Does DeepInfra train on your data?

DeepInfra's data-privacy documentation states that input is not stored to disk during inference, outputs are deleted after delivery, prompts are not logged beyond metadata, and data is not used for training. That is a self-published policy statement, not an independent verification; it also carves out Google and Anthropic models, whose own policies apply.

How fast is DeepInfra?

On gpt-oss-120b (high), Artificial Analysis's 72-hour medians measured DeepInfra at about 37 output tokens per second — far below Cerebras (1,686), SambaNova (708) and Groq (475). Its DeepInfra Turbo variant trades a higher price for faster measured speed. For throughput-tolerant workloads the per-token discount usually wins; for streaming user-facing output, benchmark the fast-silicon hosts.

Does DeepInfra serve closed models like Claude and Gemini?

Yes. DeepInfra's 2026-10-13 rate card lists Anthropic's Claude family (Claude Sonnet 5 at $3.00/$15.00 per million tokens, Claude Opus 5 at $5.00/$25.00, Claude Fable 5 at $10.00/$50.00) and Google's Gemini family (Gemini 3.5 Flash at $1.50/$9.00), alongside 100+ open-weight models — a passthrough layer on top of its own hosted open-model fleet.