best 14 llm inference providers

on the anchor model novita is $0.135 per million in — 35% above the cheapest price in the category and a seventh of the most expensive — and it undercuts the reference price on deepseek v4 pro at $1.60/$3.20 where three larger vendors all charge $1.74/$3.48. on glm-5.2 it prints the same $1.40/$4.40 as everyone else.

the catalogue is the reason it ranks first rather than second. multiple deepseek generations, multiple glm generations including 5.2, the qwen family and llama — so the model you standardise on today and the one you migrate to in october are both here. deepinfra is cheaper and carries no glm at all, which is a worse trade than $0.035 per million.

two extras that matter operationally: a stated 50% batch inference discount, and cache reads at 10-20% of the standard input rate. two models are served free outright. the gap is the usual one — no published rate limits and no speed figures, so capacity planning is guesswork until you're running.

$0.10 per million input tokens on llama 3.3 70b is the lowest verified figure in this category, and it is not close: together charges $1.04 for the identical model. deepseek v4 pro at $1.30/$2.60 also undercuts the $1.74/$3.48 that fireworks, baseten and together all charge, so the discounting extends to the frontier rather than stopping at legacy models.

the service-tier system is the cleverest pricing here and the most easily missed. the same model bills at 1x on standard, 1.5x on priority, and 0.8x on flex if you can tolerate latency — a batch discount in everything but name, applied per request rather than per job.

the catalogue gap is real and specific: we found no glm model anywhere on deepinfra's pricing page. given glm-5.2's standing among current open models that is a significant hole, and it is the only reason this isn't first. dedicated gpu deployments run $0.89-$4.89/hr for teams that outgrow serverless.

the pricing docs are unusually honest about complexity instead of hiding it: every model has an explicit standard and priority rate, cached input is priced separately and steeply discounted — deepseek v4 pro's cached input is $0.145 against $1.74 standard — and models without a named row fall into published parameter-count tiers rather than a quote form.

deepseek v4 flash is the find. at $0.14 in and $0.28 out it costs a twelfth of v4 pro, and for the large share of workloads that don't need the pro tier that is the cheapest frontier-adjacent capability on this page.

it charges the reference $1.40/$4.40 on glm-5.2 like everyone else, adds 50% off batch inference, and rents on-demand gpus at $7/hr for h100 and h200 through $12/hr for b300 if you outgrow serverless. the one gap: no named llama flagship row, which likely means llama lands in the generic over-16b tier at $0.90 — pricier than the specialists.

the differentiator is deployment range rather than price. together sells serverless tokens, provisioned throughput units with guaranteed capacity, single-tenant dedicated inference instances at $5.49-$8.99/hr, and reserved gpu clusters with volume discounts up to 32%. nobody else here covers that whole span, and the frontier open models land on it fast.

on price it is the most expensive place to run the anchor model: $1.04 per million input for llama 3.3 70b, against $0.10 at deepinfra. it charges the reference rate on glm-5.2 and deepseek v4 pro like the rest of the field, and qwen3.6-plus at $0.50/$3.00 is genuinely competitive — so the premium is concentrated on llama rather than universal.

watch two things. cached input is dramatically cheaper across deepseek, glm and qwen — $0.20 against $1.74 on v4 pro — so your effective rate depends on hit rate more than on the sticker. and there is no batch api discount; savings come from committing to capacity instead, which is a different financial decision.

groq publishes a tokens-per-second figure per model directly on its pricing page: 394 tps for llama 3.3 70b, 500 for qwen 3.6 27b and gpt-oss 120b, 840 for llama 3.1 8b, 1,000 for gpt-oss 20b. every other vendor in this category either markets a vague multiple or says nothing. treating speed as a published spec rather than a claim is the right instinct and nobody else has it.

the price is mid-field and reasonable — $0.59/$0.79 on the anchor model, roughly six times deepinfra but well under together — with a 50% batch discount, 50% off cached input, and drop-in openai sdk compatibility.

the catalogue is the constraint. five models on the public pricing page, no deepseek and no glm anywhere, and minimax reserved for enterprise. the free tier is real but tight: 30 requests a minute and 1,000 a day on the 70b model, which is a demo allowance rather than a runway.

openrouter's own faq states it passes through the underlying provider's pricing without markup on standard inference. for an aggregator that is a genuinely good deal, and it removes the usual objection — our ai image models ranking documents aggregators charging 12-25% over going direct.

what you buy is optionality. one openai-compatible endpoint reaches most of the vendors ranked above, so switching provider or model is configuration rather than engineering, and a free tier lets you try models at 50 requests a day, rising to 1,000 once you've bought $10 of credit.

two caveats. the price of any given call is whatever the routed provider charges, so there is no openrouter price list to plan against — you are still doing the comparison this page exists to help with. and bring-your-own-key is free only to a million requests a month, after which it costs 5% of what openrouter would have charged.

baseten is the only entry here leading with soc 2 type ii and hipaa, and one of the few offering self-hosted and hybrid deployment alongside its cloud. for teams whose blocker is a security review rather than a price, that combination decides it.

published pricing sits exactly on the market reference — glm-5.2 at $1.40/$4.40, deepseek v4 at $1.74/$3.48 — with glm 4.7 at $0.60/$2.20 as a cheaper step down and a 'fast' glm-5.2 variant at $2.10/$6.60 for latency-sensitive work. billing is compute-time only, with no idle charge.

the frustration is that only nine models carry a public per-token rate. llama 4, llama 3.3, qwen3.6, kimi k2.7 and nemotron all appear in the model library with no price beside them, which in practice means dedicated deployment and a conversation. that's a defensible enterprise motion and a bad experience if you just wanted to know what llama costs.

the wafer-scale hardware is not marketing: independent leaderboards have put cerebras far ahead of gpu-based inference on large models, and no competitor here claims comparable throughput. if generation speed is the product constraint, this is the shortlist.

what we could verify is thin. $5 of free credits, a $10 self-serve minimum, and two subscription plans — code pro at $50/month for up to 24 million tokens a day, code max at $200/month for up to 120 million. those are strong allowances if your usage fits the daily cap shape.

what we could not verify is the per-model, per-token price, across three attempts at the vendor's own pricing page. the site's only speed claim in readable text is '20x faster than openai and anthropic', with no model attached. a vendor selling on performance that publishes neither a numeric benchmark nor a token rate in machine-readable form is asking buyers to take both on faith. one preview model also carries a deprecation date of 17 august 2026.

the llama price is competitive — $0.60/$1.20 on the anchor model, roughly matching groq — and the rdu architecture is a genuine third approach alongside gpus and groq's lpu. openai-sdk compatibility is confirmed in its docs.

the catalogue is where it falls behind. six models, no glm, no qwen, and deepseek stops at v3.1 and v3.2 while most of this list has moved to v4 — both v3 variants priced at $3.00/$4.50, which is more than fireworks charges for v4 pro. a cached-input rate on minimax at $0.06 against $0.60 standard is the one striking discount.

the documentation gaps compound it. no tokens-per-second figures anywhere, despite speed being the company's entire hardware thesis, and the rate-limits page returned a 404 when we looked. for a vendor asking developers to adopt unfamiliar silicon, that is the wrong pair of things to leave blank.

sixty-plus open-source models is real breadth, and the named ones — deepseek v4 pro, glm-5.1, qwen3 variants, llama — are the current generation rather than a stale list. the api is openai-compatible, and a european provider is a genuine consideration for teams with data-residency constraints.

the pricing is inaccessible. the per-model table lives at a javascript-rendered dashboard url, and no rates could be read from it. we will not publish figures for a vendor whose own price list we couldn't open, so every rate here is unverified.

it is also mid-rebrand, from 'nebius ai studio' to 'nebius token factory', with both names live on its own properties. that is a small thing, but combined with unreadable pricing and no published rate limits it makes evaluation harder than it should be for a provider whose catalogue deserves a look.

llama 4 scout at $0.08/$0.15 and maverick at $0.35/$0.40 are the cheapest rates we verified for those models anywhere in this category, and deepseek v3.2 at $0.14/$0.28 is likewise strong.

the plan layer is unusual and needs reading twice. on top of per-token usage sit tiers metered by 'gateway requests': pay-as-you-go includes a million a month at 30 requests per minute, growth costs $250/month for fifty million at 250 per minute. that request ceiling, not the token price, is what will actually constrain a busy application.

verification thinned out fast. the main pricing table didn't render as static text, so the figures above came from the vendor's own comparison post rather than its price list, and we found no published rate for glm-5.2, deepseek v4 or any qwen model despite glm-5.2 being named elsewhere on the site. total catalogue size is unpublished.

the model is the point: a flat monthly fee for unlimited tokens, gated by concurrency rather than volume. $25 buys four concurrent units against the catalogue; $100 covers models up to 229b with eight units and 256k context; $200 unlocks anything including deepseek, kimi and glm. for constant high-volume traffic on smaller models, that inverts the usual arithmetic in your favour.

the catalogue breadth is the other pitch, and it is where the credibility wobbles. featherless states '47,300+ models' on one page, '20,000+' on another, and '6,700+' in its own blog. we cannot tell you which is true, and a vendor whose headline number varies sevenfold across its own properties is asking to be taken on trust.

two structural limits: the flagship open models everyone else prices at $1.40 per million sit behind the $200/month tier here, and there is no per-token option for like-for-like comparison unless you use the separate per-request product. openai compatibility isn't confirmed on the pages we could read.

modal sells raw serverless gpu by the second: h100 at $0.001097, a100 80gb at $0.000694, b300 at $0.001972, down to t4 at $0.000164, billed only while running with no idle charge. the starter plan includes $30 of monthly credits and ten concurrent gpus.

there is no hosted open-weight model api here and no per-token rate to compare, because the serving layer is yours — you deploy vllm or sglang and operate it. that is genuine freedom for a custom fine-tune or an unusual checkpoint nobody else hosts, and pure overhead if you just wanted glm-5.2 behind an endpoint.

the honest comparison: at $0.001097 a second an h100 costs about $3.95 an hour. whether that beats $1.40 per million tokens depends entirely on your utilisation, and for bursty traffic it usually doesn't. run the arithmetic against your own volume before treating this as the cheap option.

replicate's own pricing page says most models are billed by the time they take to run, not by token. that suits image models, audio models and obscure community checkpoints — which is what it is genuinely good at — and makes it structurally incomparable to everything above it here.

only two entries carry per-token rates, and one of them, claude 3.7 sonnet, isn't an open-weight model at all. the other is deepseek r1 at $3.75 in and $10.00 out — a generation behind the v4 that most of this list serves, at several times the price. the output rate is printed as $0.01 per thousand tokens, a different unit convention from every other vendor here, which is an easy way to misread it by a factor of a thousand.

we tried to price a current llama flagship on it and the model page returned a 404 within our budget. if you want llama, glm, qwen or deepseek v4 behind an api, everything above this entry does it more cheaply and more legibly.