AI gateway fees compared (October 2026)
On pay-as-you-go AI gateways, $100 of model usage costs $100 to $115 in October 2026, because the fee on credits ranges from 0% to 15%, and only some of them stop spending at a hard limit.
| Gateway | Platform fee | Card / crypto fee | BYOK fee | Minimum top-up | Credit expiry | Subscription | Budget cap | Claude Code | Codex | EU routing | Source, checked Oct 13, 2026 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| ClinePass (Cline) | 0%Cline provider, “at cost” | Not publishedA processing fee “may apply” | 0%In the Cline extension | Not published | Not published | $9.99/moClinePass, open models only | Not published5-hour, weekly and monthly quotas | NoChat Completions only | No | Not published | Pricing, ClinePass, API |
| Cloudflare AI Gateway | 5%On Unified Billing credits | 0%Inside the 5%; crypto is a US-only beta | 0% | Not published | Not published | No | Per gateway: HTTP 429“Eventually consistent”: bursts can exceed | Yes | PartialCodex setup takes OpenAI models only | Not published | Unified Billing, Spend limits, Codex |
| CodeRouter | 15%On plan overage and Direct pay-as-you-go | Not published | Not publishedA “platform fee” is charged | $10Direct | Not published | $19 to $299/moSold in tokens | Soft overage capReroutes to cheaper models; Direct: HTTP 429 at $0 | Yes | Not published | Enterprise only | Pricing, Direct |
| DevPass (LLM Gateway) | n/aPlan includes 3x its price at list, 2x from Oct 15 | Not published | Not published | $10Overflow top-ups, at +5% | MonthlyAllowance resets, no rollover | $29 to $179/moInteractive coding only | Plan: HTTP 402From Oct 15, also a 24-hour cap of 8% to 10% | Yes | Yes | NoOnly a “No AI training” filter | Pricing, Terms |
| Kilo Gateway | 0% | 5%Processing fee on every purchase | 0% | Not publishedOrganizations: $100 auto top-up | 1 yearPer the terms; team docs say never | $19 to $199/moKilo Pass: credits plus a bonus | Per user, dailyOrganizations only; no per-key caps documented | Not publishedDocs list Chat Completions only | Not published | Enterprise only | Pricing, Billing docs, Terms |
| LiteLLM | 0%Free to self-host (MIT) | n/aNo credits | 0%Always your own keys | n/a | n/a | EnterpriseLicense price not published | Per key: HTTP 400/401max_budget; needs a database | YesAny provider | YesAny model | Self-hostedWherever you run it | Site, Enterprise, Budgets |
| LLM Gateway | 5%On credit purchases | 1.5%Non-US cards only | 0% | None“No minimums” | Not published | NoDevPass is separate | Per key: HTTP 401 | YesAny model | Yes | Per requestRegion pin; full policy on Enterprise | Pricing, Billing docs |
| NanoGPT | 0%Auto-routed; 5% if you pin a provider | 0% | 5% | $1$0.10 in crypto | Not published | $12/moPro: 60M input tokens a week, personal use | Per key: HTTP 429Daily USD cap, approximate | YesAny model | YesAny model | Per requestPin an EU provider route, +5% | Pricing, Limits |
| OpenCode Zen (Console) | 0% | 4.6% + $0.31Models page says 4.4% + $0.30 | 0% | Not publishedPresets from $10 | Not published | $10 or $40/moGo and Go Plus: open models, one member | Per key: HTTP 429Also workspace and member; running requests finish | PartialClaude and Qwen models | PartialGPT, Grok and Muse Spark models | NoPages disagree: US, or US and EU | Billing, Models, Budgets |
| OpenRouter | 5.5%Min $0.80; 8% on Business | 0%Card inside the 5.5%; crypto pays 5% | 0%Up to $25,000/mo, then 5% | $5 | 365 daysOpenRouter “may” expire unused credits | No | Per key: HTTP 402/403Every plan; the docs give both codes | Yes | YesAny model | Business and up | Pricing, FAQ, Terms |
| Opper | 3%Gateway plan; 5.5% on Control Plane | 0%Inside the fee; no crypto listed | 0%Gateway plan; Control Plane charges a fee | None | Not published | No | Org, project: HTTP 402Control Plane only; the crossing call completes | Yes | Yes | Every planHosted in Stockholm; EU routes per call | Pricing, Spend, EU |
| Portkey (Prisma AIRS) | n/aYour own keys only | n/aNo credits | 0% | n/a | n/a | $49/moProduction plan for logs, no tokens | Per key: HTTP 412Enterprise and select Pro; blocked at once | YesAny model | YesAny model | Enterprise onlyHybrid or VPC | Pricing, Budgets |
| Requesty | 5%On model cost | 0%None published; no crypto | 0% | Not published | Not published | No | Per key, monthlyNew requests blocked; code not published | Yes | YesNot only OpenAI models | Every planFrankfurt endpoint | Pricing, Spend limits, EU |
| Standard Compute | 0%Budget is 5% to 14% above plan price | 0% | n/aNot offered | $23Pay as you go step at +15%, Economy and up | MonthlyBudget renews, no rollover | $19 to $2,499/moFixed compute budget | Per key: HTTP 402New requests stop at 100%; running ones finish | Yes | Yes | Every planOpt-in Europe or US rule | Budgets, Regions |
| Together AI | 0%Its own per-token prices | 0%Card or ACH; no crypto | Not published | $15First purchase, then $5 | Never | No | NoneStops at a $0 balance, not exactly | PartialOnly through Together Link (beta) | PartialOnly through Together Link (beta) | Dedicated onlyNot on serverless | Credits, Spend controls, Together Link |
| Vercel AI Gateway | 0% | Not publishedA processing fee is passed on | 0%Needs purchased credits | Not published | 1 year | No | Per key: HTTP 402Soft: the request that crosses completes | Yes | YesAny model | Per requestinferenceRegion; gateway hop not pinned | Pricing, Budgets, Regions |
The same $100 of model usage, through each gateway
One month of agent traffic that costs exactly $100 at the providers' list prices, bought in one purchase with a US card, before tax.
| Gateway | How the fee works | You pay |
|---|---|---|
| DevPass Lite | From Oct 15, $29 includes $58; $42 overflow at +5%. Interactive coding only, with a 24-hour cap | $73.10 |
| NanoGPT | No fee on auto-routed models | $100.00 |
| Together AI | Its own prices, no fee; open models only | $100.00 |
| Vercel AI Gateway | No markup; card processing fee not published | $100 + card fee |
| LiteLLM | Your own provider keys; you run the server | $100 + hosting |
| Opper | 3% on the Gateway plan ($103.09 if taken from the purchase; not stated) | $103.00 |
| OpenCode Zen | 4.6% + $0.31 per purchase | $104.91 |
| Requesty | 5% on model cost | $105.00 |
| Cloudflare AI Gateway | 5% on Unified Billing credits | $105.00 |
| Kilo Gateway | 5% processing fee | $105.00 |
| LLM Gateway | 5% on credits; non-US cards add 1.5% | $105.00 |
| OpenRouter | 5.5% on credits (8% on Business) | $105.50 |
| Standard Compute | Standard plan $89 covers $95; one $23 Pay as you go step adds $20 | $112.00 |
| CodeRouter Direct | List price +15% | $115.00 |
| Portkey | Your own provider keys, plus the $49 Production plan for logs | $149.00 |
ClinePass is left out because it publishes no dollar quota.
Fees vs hard stops
The fee percentage is the small number. A 5% fee on an agent that loops all night is 5% of an open-ended bill, and several caps are not hard stops: Vercel and Opper let the request that crosses the limit finish, Cloudflare's limits are eventually consistent, and CodeRouter's overage cap moves you to cheaper models instead of stopping.
Standard Compute is the other shape: a plan buys more usage than it costs (the $89 Standard plan covers $95, and budgets run 5% to 14% above plan price), so a month that fits its plan costs less than list price. A month that lands between two plans, like the $100 example above, costs more. The bigger saving is routing: the same agent work sent to efficient models for routine turns uses a fraction of the budget a single frontier model would.
How we checked
Every cell was read on October 13, 2026 on the vendor's own pricing page, documentation, FAQ, terms or llms.txt, never on a third-party review. Each row links to the pages we used. Where two of a vendor's own pages disagree, the cell says so: Kilo's terms and team docs on credit expiry, OpenCode's card fee and hosting region, and OpenRouter's status code for a spent key.
Standard Compute's row comes from the plan and budget code that the gateway enforces and from our public docs. To correct a row, email contact@standardcompute.com with a link to the vendor's page and we will update the row and its date.
Where Standard Compute fits
Standard Compute is a monthly plan from $19 with a fixed compute budget, metered at provider rates with no card fee or credit purchase on top. Each key stops with HTTP 402 when its budget is used, and Pay as you go at +15% is there only if you turn it on. Claude Code connects on /v1/messages, Codex on /v1/responses, and Europe or US routing rules work on every plan.
FAQ
Which AI gateway has the lowest fees?
On tokens bought through the gateway, NanoGPT's auto-routed models and Vercel AI Gateway add no fee (Vercel passes on an unpublished card processing fee), and Together sells its own models with no fee. Among gateways that charge on credits, Opper is lowest at 3%, Requesty, Cloudflare, Kilo and LLM Gateway charge 5%, OpenRouter 5.5%, and CodeRouter's pay-as-you-go adds 15%.
What fee does OpenRouter charge?
5.5% when you buy credits, with a $0.80 minimum, or 8% on the Business plan. Crypto purchases pay 5%. Your own provider keys are free up to $25,000 of list-price usage a month, then 5%. The minimum purchase is $5, and OpenRouter may expire unused credits 365 days after purchase.
Do AI gateway credits expire?
OpenRouter may expire credits after 365 days and Vercel's expire after one year. Together's do not expire. Kilo's terms say one year while its team docs say never. Requesty, Cloudflare, Opper, NanoGPT, OpenCode Zen, LLM Gateway and CodeRouter do not publish an expiry rule.
Which AI gateways stop spending at a hard limit?
Requesty blocks new requests once a key's monthly cap is reached, LiteLLM rejects requests past a key's max_budget, Portkey returns 412 on Enterprise and select Pro plans, and Standard Compute returns 402 to new requests while running ones finish. Vercel and Opper let the request that crosses the limit complete, Cloudflare's limits are eventually consistent, and CodeRouter's overage cap reroutes to cheaper models instead of stopping.
Which AI gateways work natively with Claude Code and Codex?
OpenRouter, Requesty, Vercel AI Gateway, LiteLLM, Portkey, NanoGPT, Opper, LLM Gateway, DevPass and Standard Compute serve both the Anthropic Messages API and the OpenAI Responses API. OpenCode Zen splits them by model family, Cloudflare's Codex setup takes OpenAI models only, CodeRouter documents Messages but not Responses, Together supports both only through Together Link, and Kilo Gateway and Cline document Chat Completions only.