Open World Evaluation

Best Flat-Rate LLM API Providers in 2026

updated on Jul 23, 2026

Flat-rate LLM providers sell unlimited model usage for a fixed monthly price instead of billing per token. This model spread because agentic coding sessions can use tens of millions of tokens, so a per-token bill is hard to predict. Very few providers offer a true flat fee; most plans marketed as flat carry a usage quota underneath. Below we compare the providers that actually sell unlimited usage, with the models served, the concurrency limit, and the pricing for each.

Provider	Focus	Models	Concurrency	Price
Giotto.ai	EU-hosted sovereign deployment	Giotto 1 (own model)	1 request at a time	49.90 CHF/mo, announced
Featherless.ai	Open-weight model variety at a fixed cost	30,000+ open-weight models	4 to 8 units per plan	From $25/mo
Awan LLM	Budget flat-fee access	Open-weight, catalogue unlisted	100 to 200 requests/min	$5 to $80/mo

Providers in this table meet two criteria:

A published fixed monthly price.
No token, prompt, or time-window quota on usage.

Flat-fee LLM providers analyzed

Featherless.ai

Featherless serves more than 30,000 open-weight models (DeepSeek, Kimi, GLM, Qwen, Llama, and fine-tunes) through one OpenAI-compatible API with unlimited tokens and requests.¹ Instead of a token meter, it prices concurrency units: a small model costs 1 unit per in-flight request, a 70B-class model costs 4, and requests over your budget return an HTTP 429 error instead of a charge.

The Premium plan includes 4 units, so $25 buys one large-model lane or four small-model lanes. Agent plans add 8 units and a sandbox for autonomous workloads. The company raised a $20M Series A co-led by AMD Ventures and Airbus Ventures on 2026-04-30, the strongest financial backing in this market.

Best use case: Heavy single-stream agentic coding and model experimentation at a hard cost ceiling.

Models: 30,000+ open-weight models; 229B size cap on the $100 tiers, no cap on the $200 tiers.

Concurrency: 4 units (Premium), 8 units (Agent and per Business unit).

Price: Premium $25/mo. Agent $100 or $200/mo. Business $100 or $200 per unit/mo.

Key features:

Unlimited tokens and requests on every plan.
Any catalogued open-weight model, switchable per request.
OpenAI-compatible API.
Documented no-logging policy for prompts and completions.
Concurrency scales linearly on Business plans.

Limitations:

One 70B-class request consumes the entire Premium unit budget.
Over-budget requests are rejected with HTTP 429, so parallel tooling needs its own request queue.
Throughput per lane is not published; the vendor does not state tokens per second.

Giotto.ai

Giotto is a Swiss lab in Lausanne serving its own reasoning model, Giotto 1, which it positions as the strongest model that runs on a single GPU.² The company sells the same system as EU-hosted cloud, private on-premise deployment, or preinstalled hardware, with ISO 27001 certification.

The flat-fee API was announced by CEO Aldo Podestà on 2026-07-09: 49.90 CHF per month, one concurrent request, unlimited usage.³ As of 2026-07-23 the offer is not on giotto.ai, which has no pricing page and routes buyers to a contact form, so all terms below are provisional.

Best use case: EU data residency and sovereignty requirements on a fixed budget.

Models: Giotto 1 only.

Concurrency: 1 request at a time (mono-concurrency).

Price: 49.90 CHF/mo, announcement only; free Giotto Open workspace with unstated API limits.

Key features:

Unlimited usage through a single request lane, per the announcement.
EU hosting in Swiss and EU partner data centers.
Same model available for private deployment on own GPUs.
Free shared workspace includes API access.

Limitations:

Not purchasable yet; no public pricing page as of 2026-07-23.
The announcement claims up to 200x savings versus Fable 5 and 350x versus GPT Pro’s API; these compare price only. Giotto’s own benchmark table positions Giotto 1 against single-GPU models such as Gemma 4 31B and GPT-OSS-120B, not the frontier models in the price claim.
Early-stage company: the product launched 2026-05-18, and the CEO stated in June 2026 that a CHF 5-10M convertible loan was expected to close by July, with a larger seed round planned for the end of the year.

Awan LLM

Awan LLM sells “unlimited tokens” from $5 per month, with per-day request caps that scale by model size class and per-minute rate limits.⁴ The daily caps (30K to 80K requests on the $20 Pro plan) sit far above what a single interactive user generates, which makes the plan flat-fee in practice.

The pricing page does not list which models are served or how current they are. Verify the catalogue against your workload before paying.

Best use case: Lowest-cost flat-fee entry for light and mid open-weight workloads.

Models: Open-weight; catalogue and versions not stated on the pricing page.

Concurrency: Rate-limited at 100 requests/min (Pro) to 200 requests/min (Max).

Price: Lite free, Core $5/mo, Plus $10/mo, Pro $20/mo, Max $80/mo.

Key features:

Unlimited tokens on every paid tier.
Free tier available for small-scale testing.
Max tier removes daily request caps.
Cheapest paid entry in this market at $5/mo.

Limitations:

Served models and their versions are not documented; quality is unverified.
Request caps by model size class replace the token meter.

Flat-priced plans with usage quotas

Most plans in the flat-fee conversation are subscriptions with a refreshing quota: the price is fixed, and a prompt or token budget stops you instead of billing you. What happens at the quota edge differs by vendor and matters as much as the quota size. We cover these plans tier by tier in our LLM pricing guide.

Z.ai GLM Coding Plan

Meter: ~80 to 1,600 prompts per 5-hour window plus ~400 to 8,000 per week by tier; each prompt triggers an estimated 15 to 20 model calls; 3x quota deduction at peak hours.
At the edge: hard block until the window refreshes, no fallback to pay-as-you-go.
Overview: $18 to $160/mo, serves GLM-5.2 through an Anthropic-compatible endpoint in 20+ coding tools.⁵

Kimi Code

Meter: a credit pool shared with all Kimi membership features, refreshed weekly with no rollover, plus a rolling 5-hour rate window; quota volumes not published.
At the edge: metered top-up via Extra Usage instead of a hard block.
Overview: $19 to $199/mo for K2.7 Code with 2 to 8 concurrent agent tasks; HighSpeed mode consumes 3x quota.⁶

MiniMax Coding Plan

Meter: ~1.7B to 12.5B tokens per month inside 5-hour and weekly windows; unused quota does not carry over.
Overview: $20 to $120/mo; MiniMax recommends pay-as-you-go for production use.⁷

Cerebras Code

Meter: 24M to 120M tokens per day.
Overview: $50 or $200/mo for GLM 4.7 on Cerebras hardware; every tier still showed sold out on 2026-07-23.⁸

NanoGPT Pro

Meter: 60M input tokens per week; large models deduct at 2x.
Overview: $12/mo across an included open-source model set, web and API.⁹

Synthetic

Meter: 500 requests per 5 hours, 1 concurrent request per model.
Overview: $30/mo for seven always-on open-weight models, UI plus API.¹⁰

Chutes

Meter: included usage capped at 5x the plan price, then billing continues per token.
Overview: $10 or $20/mo prepaid bundles on decentralized compute; a discount program rather than a flat fee.¹¹

Claude and ChatGPT subscriptions

Meter: 5-hour sessions plus weekly caps (Claude); usage tiers with limits (ChatGPT).
Overview: $20 to $200+/mo; the only flat-priced route to frontier closed models, sold as seats, not APIs.¹²

How flat-fee pricing works

1. Concurrency lanes replace the token meter

A flat-fee provider caps how many requests you can run at once instead of how many tokens you consume. One request lane cannot draw more than one GPU stream’s output, so the provider’s worst-case serving cost is fixed. The arithmetic also caps you: a lane streaming 50 tokens per second around the clock produces at most about 130M output tokens in a 30-day month, and real usage sits well below that because the lane idles between calls. Tokens per second therefore becomes a purchasing criterion, and none of the three providers publishes it.

2. Open-weight economics

Every true flat-fee provider serves either open-weight models on commodity GPUs (Featherless, Awan) or its own model on its own infrastructure (Giotto). No one sells flat unlimited access to a frontier closed model, because a reseller cannot cap the serving cost of a model it licenses per token. Anthropic and OpenAI sell flat prices only for their own models, bundled into quota-limited chat subscriptions.

3. Break-even against pay-per-token

A flat plan wins when metered spend for the same model would exceed the plan price, and the break-even point depends heavily on which metered host you compare against. Llama 3.3 70B costs $0.10 per million input tokens and $0.32 per million output on DeepInfra, against a flat $1.04 per million on Together AI.¹³ Assuming a 70:30 input-output mix, Featherless Premium at $25 breaks even at roughly 150M tokens per month against DeepInfra, but at only about 24M tokens against Together AI. Heavy single-stream agentic use clears the lower bar easily and can clear the higher one.

The reverse case matters just as much. A workload of 5M tokens a month spread over twenty parallel short sessions costs under $1 metered on DeepInfra, while serving that concurrency on flat-fee lanes would cost $100 or more. Flat fee suits high-volume, low-concurrency work; pay-per-token wins on low-volume or highly parallel traffic. Figures are illustrations at the access dates cited; run last month’s token counts through the same arithmetic.

4. Sustainability of unlimited pricing

Unlimited plans attract the heaviest users, and the market already shows the strain. Cerebras Code was sold out at every tier on 2026-07-23, ten days after we first checked. NanoGPT raised its subscription from $8 to $12 within a year. Z.ai’s 30% discount is time-boxed.

The precedents run bigger than this niche. OpenAI’s CEO said in January 2025 that the $200/mo ChatGPT Pro plan was losing money because subscribers used it more than projected.¹⁴ GitHub replaced Copilot’s premium request units with usage-based AI Credits on 2026-06-01 while keeping subscription prices unchanged.¹⁵ Cursor swapped its 500-request allocation for a $20 usage pool in June 2025 and issued refunds after the backlash.¹⁶ When the largest sellers of flat-priced AI usage retreat toward meters, treat a small provider’s unlimited plan as an introductory price, and break-even math as valid for months rather than years.

Get our team to automate one of your business processes with AI agents, free of charge.

Automate a process

Why would you need a flat-fee LLM plan in 2026?

1. Heavy agentic coding

A single coding agent session makes hundreds of sequential model calls and can burn tens of millions of tokens. One lane serves this pattern continuously with no bill variance. Featherless fits; Giotto will once purchasable.

2. EU data residency

Giotto hosts in Swiss and EU data centers, serves its own model, and offers the same system for private deployment, which suits regulated teams that cannot send data to US-hosted APIs.

3. Model experimentation

Featherless exposes 30,000+ open-weight models behind one API and one fee, so comparing models costs nothing beyond the subscription.

4. Cost-capped side projects

A hobby project with spiky usage risks surprise bills on a meter. Awan LLM at $5 to $20 per month caps the downside at the subscription price.

5. Batch and background processing

Overnight summarization, classification, or data-cleaning jobs tolerate queueing, so they can saturate a lane’s idle hours at zero marginal cost. Any of the three providers fits if the model quality suffices.

How to choose the right flat-fee LLM plan?

The workload shape decides more than the price. Sequential work (one agent, one chat, batch jobs) runs well on a single lane. Parallel agent fleets serialize behind a lane and need more concurrency: Featherless prices extra units explicitly and rejects over-budget requests with HTTP 429, while a mono-concurrency plan like Giotto’s announced API cannot be widened at all.

Model quality sets the floor. A flat plan only saves money if the served model can do your work, so check independent benchmark scores for the specific model before comparing prices. Price multiples quoted against frontier models compare different quality classes.

Monthly volume and the metered alternative decide flat versus metered together. Against a low-cost host the crossover for a $25 plan sits near 150M tokens per month; against a premium host it drops to about 24M. Below your crossover, pay-as-you-go is cheaper; far above it, a flat plan returns multiples of its price.

Limitations worth checking before subscribing:

Tokens-per-second and latency figures, which no provider currently publishes.
What one large-model request costs in concurrency units.
Behavior at the limit: rejected requests (Featherless), queueing, or metered top-up.
Whether the served model catalogue is documented (Awan does not list it).
Commercial and production-use terms.

Provider durability is a real criterion in this market. Featherless carries a $20M Series A; Giotto was still closing a CHF 5-10M convertible loan as of June 2026. Prefer providers whose throttle gives them sustainable unit economics, and avoid standardizing a team on a plan that may be repriced or sold out within a year.

See more of our benchmarks and data-driven insights in Google Search.

Add as preferred source

FAQs

As of July 2026: Featherless.ai (from $25/mo, concurrency-capped), Awan LLM ($5 to $80/mo, request-rate-capped), and Giotto.ai’s announced 49.90 CHF/mo API, not yet publicly purchasable. Every other flat-priced plan we checked meters usage through refreshing quotas.

By capping concurrent requests. One lane cannot consume more than one GPU stream’s output, so the provider’s worst-case cost is fixed, and the served models are open-weight or vendor-owned, so no per-token license cost applies.

Generally no. Throughput tops out at the purchased lanes, no provider in the main table publishes an SLA for these plans, and quota-based alternatives explicitly point production users to pay-as-you-go.

Cite this research

Pick the format that matches where you're publishing. Pasting the link version into your CMS preserves the backlink.

Berk Kalelioğlu (2026) - "Best Flat-Rate LLM API Providers in 2026". Published online at AIMultiple.com. Retrieved July 23, 2026, from: https://aimultiple.com/flat-rate-llm-api [Online Resource]

Kalelioğlu, B. (2026, July 23). Best Flat-Rate LLM API Providers in 2026. AIMultiple. https://aimultiple.com/flat-rate-llm-api

@misc{kalelioglu2026,
  author = {Kalelioğlu, Berk},
  title  = {{Best Flat-Rate LLM API Providers in 2026}},
  year   = {2026},
  month  = jul,
  howpublished    = {\url{https://aimultiple.com/flat-rate-llm-api}},
  note   = {AIMultiple. Retrieved July 23, 2026}
}