Modal ranks #13 of 14 in our llm inference providers testing. gpu seconds, not model tokens — the right tool for a different job than this page is about..
62/100
excellent serverless gpu with no idle billing, ranked low only because it isn't the product this category compares: there is no token price, because you build the endpoint.
why people look for an alternative
−no hosted model api and no per-token pricing at all
−you build and operate the serving layer
−openai-compatible endpoints are yours to provide
−only economic at high, steady utilisation
stay with Modal if per-second billing with no idle charges is the thing you care about most — nothing below beats it on that.
#4 in llm inference providers · every deployment shape from serverless to your own gpu cluster — at the highest llama price on this list.
86/100
verdictthe widest ladder in the category — pay-per-token, provisioned throughput, dedicated instances, whole clusters — and you pay for the ladder on every token.
Together AI vs Modal
Modal
Together AI
price
$0.001097/s
$1.04 / 1m in
free tier
yes
no
billing
gpu per-second
per-token to dedicated
llama 70b input
n/a — you host it
$1.04/1m
deepseek v4
n/a — you host it
$1.74/1m in
glm-5.2
n/a — you host it
$1.40/1m in
batch discount
n/a
none; commit instead
switch forteams who expect to graduate from serverless to reserved throughput or dedicated hardware without changing vendor.
pros
+serverless through ptu, dedicated instances and full clusters
+same-day access to new frontier open releases
+steep cached-input rates across deepseek, glm and qwen
+volume discounts up to 32% on commitments
cons
−10x deepinfra's price on the anchor llama model
−no batch api discount — savings require capacity commitments
−no published rate limits; dynamic and account-level
#5 in llm inference providers · the only vendor here that prints tokens-per-second next to the price — on a catalogue of five models.
85/100
verdictuniquely transparent about throughput and priced fairly for it — but a five-model catalogue with no deepseek and no glm rules it out for a lot of work.
Groq vs Modal
Modal
Groq
price
$0.001097/s
$0.59 / 1m in
free tier
yes
yes
billing
gpu per-second
per-token
llama 70b input
n/a — you host it
$0.59/1m
deepseek v4
n/a — you host it
not served
glm-5.2
n/a — you host it
not served
batch discount
n/a
50%
switch forlatency-sensitive products that can live on llama or gpt-oss and want the speed number in writing.
pros
+per-model tokens-per-second published on the pricing page
+50% off both batch and cached input
+openai sdk drop-in compatibility
+genuinely low latency at a mid-field price
cons
−only five models on the public pricing page
−no deepseek and no glm models at all
−free tier capped at 1,000 requests a day on the 70b
#6 in llm inference providers · one endpoint in front of the whole market, at genuinely no markup — which makes the routing itself the product.
84/100
verdictcharges nothing to sit in the middle, which is rare enough to be worth using — you just can't look up 'the openrouter price' for anything, because there isn't one.
OpenRouter vs Modal
Modal
OpenRouter
price
$0.001097/s
the routed provider's price
free tier
yes
yes
billing
gpu per-second
pass-through
llama 70b input
n/a — you host it
routed provider's rate
deepseek v4
n/a — you host it
via routed provider
glm-5.2
n/a — you host it
via routed provider
batch discount
n/a
provider's own
switch foranyone still choosing, still comparing, or wanting to switch providers without shipping a code change.
pros
+no markup on standard inference, per its own faq
+one openai-compatible endpoint across most of this list
+switching provider or model needs no code change
+free tier for evaluation, 1,000 requests/day after $10 credit
cons
−no price list of its own — cost depends entirely on routing
−byok costs 5% beyond 1m requests a month
−free models explicitly not recommended for production
#8 in llm inference providers · the fastest inference hardware built, attached to a pricing page that wouldn't tell us what anything costs.
76/100
verdictthe technology is genuinely in a class of its own and the subscription tiers are good value — but we could not verify a single per-token rate, and that costs it eight places.
Cerebras vs Modal
Modal
Cerebras
price
$0.001097/s
unpublished
free tier
yes
yes
billing
gpu per-second
per-token + subscriptions
llama 70b input
n/a — you host it
unpublished
deepseek v4
n/a — you host it
unpublished
glm-5.2
n/a — you host it
unpublished
batch discount
n/a
unpublished
switch fordevelopers who want extreme throughput and can work within a daily-token subscription rather than a per-token budget.
pros
+fastest inference hardware in the category by a wide margin
+generous daily token allowances on flat subscriptions
+$5 free credits and a low $10 self-serve minimum
cons
−per-model token prices unreadable across three attempts
−speed marketed as '20x' with no model or figure attached
−no published numeric rate limits
−a preview model is already scheduled for deprecation
#10 in llm inference providers · sixty-plus current open models behind a price table that won't render, during a rename.
70/100
verdictthe catalogue looks right and the european base may matter for your data rules — but you cannot compare it on price without signing up, and it's mid-rebrand.
Nebius AI Studio vs Modal
Modal
Nebius AI Studio
price
$0.001097/s
unpublished
free tier
yes
no
billing
gpu per-second
per-token
llama 70b input
n/a — you host it
unpublished
deepseek v4
n/a — you host it
offered, price unreadable
glm-5.2
n/a — you host it
glm-5.1 listed
batch discount
n/a
unpublished
switch forbuyers who will open the dashboard themselves and want breadth from a european provider.
every tool on this page went through the same test as Modal — same tasks, same order, scored the same way. the comparison tables are the figures from that testing, not vendor spec sheets.