Best AI Model Gateways for Reducing Inference Costs 2026

Vercel AI Gateway remains the best managed starting point for keeping gateway overhead low, but its GPT-5.6 Sol discount has expired. A monthly workload with 100 million input tokens and 20 million output tokens now costs $800 at both OpenAI's direct rate and Vercel's OpenAI route, before caching and optional controls. Cost reduction must come from routing, reuse, or operating work the gateway removes.

The permanent lesson is broader than one promotion. An AI model gateway lowers a bill only when routing, response caching, or enforced budgets remove more provider spend than the gateway adds in fees and operating work. Vercel is the default no-ops pick, LiteLLM is the self-hosted pick, Cloudflare is the exact-repeat cache pick, TrueFoundry is the managed semantic-cache pick, Portkey is the $49 governance pick, and OpenRouter is the provider-marketplace pick.

Best AI model gateways at a glance

All plan names, public prices, and limits below were verified against live vendor pages on October 22, 2026. "Free trial" distinguishes a permanent free route from a timed evaluation, because a free prototype can become an expensive production dependency if its real wall is hidden.

Starting price does not reveal the winner by itself. LiteLLM's license can be free while its control plane consumes real engineering time. Cloudflare's core can be free while Unified Billing adds 5%. TrueFoundry can erase whole model calls through semantic reuse, but its production price combines a $25-per-user charge with request overage. The correct comparison is the amount left on the total invoice after the gateway has done its work.

The savings equation before the rankings

The useful budget equation is remaining provider spend + gateway fees + paid controls + operating cost. A gateway earns a place only when that total is lower than direct access for the same number of accepted outputs. Model count, dashboard polish, and failover breadth are benefits, but none is a saving until it removes money or expensive failure from the workflow.

Use one normalized bill to expose the difference. OpenAI currently lists GPT-5.6 Sol at $4 per million input tokens and $20 per million output tokens. A workflow using 100 million input and 20 million output tokens therefore costs $800 before cached-input discounts, tools, or long-context multipliers: $400 of input plus $400 of output.

Vercel's current model page lists its OpenAI route at $4 input and $20 output per million tokens, so the normalized month is now $800, matching direct access. The earlier 50% promotion charged $2 input and $10 output through September 18, 2026, producing the historical $400 month shown below. These comparisons exclude requests above 272,000 input tokens, cache reads and writes, tools, and optional controls.

The model rate is only the first lever. An exact response-cache hit can avoid the provider call entirely. A cheaper-model route can reduce the unit price, but only if its outputs still clear the acceptance bar. A budget cap can stop a runaway agent, which is not a token discount but can be the largest saving of the month. The correct operating metric is cost per accepted result, including retries and review, not cost per token in isolation.

For the underlying provider-price shortlist, use the cheapest AI API comparison. This ranking starts one layer above it and asks whether a gateway can make those provider prices cheaper in practice.

1. Vercel AI Gateway: best overall for managed routing

Vercel AI Gateway is the best overall managed starting point for a team that wants no token markup and can justify the gateway through routing, controls, or reduced operating work.

Its strongest case is the absence of a token surcharge. The live pricing page says both Free and Paid use provider list rates with zero markup. A funded founder moving one high-volume extraction or support workflow can keep the normalized model line at $800 without taking on a self-hosted proxy, but the expired Sol promotion no longer halves that bill.

The Free tier includes $5 of monthly credit, covers an eligible subset of models, applies lower model-specific rate limits, and does not allow bring-your-own-key. The Paid tier uses purchased credits, opens all available models, raises limits, and enables BYOK with no Vercel gateway fee or commitment. Buying credits moves the team to Paid, so the $5 monthly Free credit no longer applies.

BYOK has a billing edge that deserves a runbook entry. When a request using a customer's provider credential fails, Vercel can retry with its system credential for reliability, and that fallback is charged to the team's gateway-credit balance. A team expecting every request to remain on a negotiated provider account should monitor that balance and reconcile fallback traffic rather than treating BYOK as a guarantee that Vercel credits will never move.

The zero-markup headline also excludes paid controls. Custom Reporting costs $0.075 per 1,000 tag, user-ID, or quota-entity writes and $5 per 1,000 reporting queries. A team-wide provider allowlist costs $0.10 per 1,000 successful requests on Pro and Enterprise, while a per-request provider filter has no extra cost. Team-wide zero-data-retention enforcement costs another $0.10 per 1,000 requests, although per-request ZDR is free on Pro and Enterprise.

Trace drains add $0.05 per 1,000 traces plus $0.50 per GB of egress, with no included Pro allowance. At one million successful requests, a team-wide allowlist is a $100 line, team-wide ZDR is another $100, and one million traces are $50 before egress. The gateway can still be the cheapest option, but the buyer should compare the configured invoice, not the token-markup headline.

The candid wall is the cost of the configured service. The Sol discount has expired. Vercel stays attractive if its zero-markup routing, budgets, failover, and observability replace work the team would otherwise own. If the product runs one stable provider model and needs none of those controls, direct billing can be the simpler choice at the same model rate.

Best for: A small or mid-market team that wants managed routing, spend controls, and no public token markup.
Standout: Zero token markup keeps the normalized GPT-5.6 Sol month at $800, matching direct access before extras.
Pricing: Free includes $5/month on eligible models; Paid uses purchased credits at provider list rates with no markup; optional reporting, policy, and trace meters are separate.
Free trial: An ongoing Free tier, not a timed trial.

The safest adoption path keeps the comparison reversible:

1. 1

   Isolate one accepted workflow

   Choose one job with a countable pass condition, such as a valid structured extraction or an approved support answer. Give it a separate gateway key so its spend and failures do not disappear inside the rest of the product.

2. 2

   Freeze the direct baseline

   Record one representative week of input tokens, output tokens, retries, latency, and accepted results on the direct provider route. Any saving is useful only if acceptance does not fall.

3. 3

   Route a bounded 10%

   Send 10% of that workflow through Vercel with the same model and service tier. Set a key-level budget that matches the sample so a configuration error cannot turn the comparison into an open tab.

4. 4

   Reconcile all four cost lines

   Add model usage, any paid reporting or policy controls, traces, and review or retry cost. Check whether a failed BYOK request consumed system credits.

5. 5

   Make the keep-or-remove decision

   Keep the route only if cost per accepted result beats direct access and the managed controls earn their price. Review the live model rate and optional meters before expanding traffic.

2. LiteLLM: best self-hosted gateway when the platform team already exists

LiteLLM is the best self-hosted gateway for a company that already operates shared infrastructure and wants the gateway license line at $0.

October 22, 2026: LiteLLM's incident report, last updated October 21 at 20:30 UTC, says the incident is contained and investigation continues. Its affected window is October 21, 11:39 to 13:23 UTC for unpinned pip installs that fetched litellm==1.107.0, and 11:46 to 14:20 UTC for Docker tags v1.107.0, latest, main-latest, and main-stable. The credential-stealing artifacts were withdrawn that day; cached copies remain unsafe. If exposed, follow the report's cleanup guidance and rebuild in a clean environment, rather than starting pip in the infected one. Pin to 1.106.2, or a later release only after LiteLLM explicitly confirms it clean, and pin verified-clean image digests. Rotate every provider key held in the proxy database, plus the other reachable credentials listed in the report. LiteLLM says its private Enterprise registry was unaffected.

The Open Source plan is free forever and self-hosted. It exposes 100+ providers through one OpenAI-compatible API and includes virtual keys, users and teams, spend tracking, budgets, rate limits, fallbacks, logging, and Prometheus. For a mid-market CTO whose platform group already runs containers, metrics, secrets, and on-call, that can consolidate provider logic without adding a monthly gateway license.

The word free stops at the license boundary. The company still supplies compute, a datastore where required, deployment pipelines, upgrades, security review, observability, and the person who responds when the gateway fails. LiteLLM is the cost winner only when those jobs are already part of the platform team's marginal workload or when self-hosting is a hard requirement. Running a self-hosted gateway safely includes the time spent pinning releases and image digests, patching, and rotating keys.

The Enterprise plan is a custom annual contract sized by annual gateway request capacity, deployment architecture, and support needs, never by token volume. It adds SSO and SCIM, OIDC or JWT controls, audit logs, secret-manager integration and key rotation, organization administration, multi-region options, formal support, and air-gapped deployment. The public page offers a 30-day Enterprise trial, but it does not publish a dollar quote.

That pricing basis changes the break-even calculation. A team with very high token spend but moderate request capacity can avoid a percentage tax on expensive models. A team with tiny model spend and no platform function can spend more in staff attention than a managed gateway would charge. The flip rule is not request volume alone: it is whether the incremental monthly cost of operating LiteLLM is lower than the hosted platform fees and operational work it replaces.

LiteLLM also fits a regulated company that must keep the data plane inside its own environment. That control can justify the decision even when the pure model bill does not fall. Cost reduction is then a secondary consequence of provider routing, budgets, and central controls, not proof that self-hosting is inherently cheap.

Best for: Teams with an existing platform function, self-hosting requirements, or enough scale to reject percentage-based gateway fees.
Standout: A $0 open-source license with budgets, virtual keys, fallbacks, logs, and 100+ providers.
Pricing: Open Source is $0 forever; Enterprise is a custom annual quote based on request capacity, architecture, and support.
Free trial: 30 days for Enterprise; Open Source is permanently free to self-host.

3. Cloudflare AI Gateway: best for byte-identical repeat traffic

Cloudflare AI Gateway is the best cost cutter when a meaningful share of requests are safe, byte-identical repeats and the team can use BYOK.

Cloudflare makes its core unusually easy to price. Dashboard analytics, caching, and rate limiting are free on all plans. Free accounts permit 10 gateways and Paid accounts 20. Logging now depends on when the first gateway was created: before September 24, 2026, Legacy Logs retain the old 100,000-record total on Free and 10-million-per-gateway limit on Paid. Newer accounts use Workers Logs: Free includes 200,000 events per day with 3-day retention; Paid includes 20 million per month, then $0.60 per additional million, with 7-day retention.

AI Gateway Logpush is available to all customers and now bills by volume. Internal and external exports each include 25 GB per month, then cost $0.03 and $0.10 per GB respectively; transformations include 1 GB, then cost $0.04 per GB. The old 10-million-request allowance and $0.05-per-million rate apply to Workers Trace Events, not these AI Gateway exports.

Caching is the real saving mechanism, and its limitation is precise. Cloudflare's default cache key includes the provider, endpoint, model, provider authentication header, and full request body. Any difference in messages, tools, or model parameters creates another entry. A cache hit skips the provider call, but two prompts that mean the same thing and use different words do not match unless the team deliberately designs a custom cache key and can defend its safety.

The minimum cache lifetime is 60 seconds, the maximum is one month, and a cacheable request can be at most 25 MB. This is a strong fit for repeated classification jobs, shared static explanations, or deterministic batch prompts. It is a poor default for personalized support, changing source data, or stochastic creative output, where a cheap stale answer can be more damaging than a fresh model call.

The billing route matters as much as the hit rate. When an applicable provider credential is present, BYOK keeps the gateway core free and pays the provider directly. Without one, Cloudflare can fall through to Unified Billing. Unified Billing passes through provider inference prices but adds 5% to purchased credits, so $100 of credits costs $105. It is also limited to 200 requests per 60 seconds per gateway; Cloudflare says that rate limit does not apply to BYOK.

Since September 14, 2026, Cloudflare lets a gateway require provider credentials with byok_only: true or a request opt out of Unified Billing with cf-aig-no-wholesale: true. A third-party-provider request without a usable key then returns HTTP 400 instead of landing on Unified Billing; Workers AI remains allowed under its existing billing mode. This is the fail-closed control cost-sensitive teams should enable after testing both stored-key and request-key paths.

Apply the normalized $800 bill to an explicit 25% exact-repeat assumption. If those repeat requests carry the same token mix and are safe to reuse, BYOK caching leaves $600 of provider spend and saves $200. Unified Billing adds 5% to the remaining $600, producing $630 and a net $170 saving. A lower hit rate can erase the advantage, while a higher safe hit rate makes Cloudflare hard to beat.

DLP scanning is free on all plans, with two predefined profiles for accounts without a Zero Trust subscription. Guardrails are not free: evaluation uses a Workers AI model and is billed at Workers AI token rates. This is another example of the right accounting boundary. The cache may be free while the safety layer adds inference of its own.

Best for: Cloudflare users with deterministic, repeated requests and a team willing to measure cache safety.
Standout: Free exact-response caching that can eliminate the provider call on a hit.
Pricing: Core analytics, caching, and rate limiting are free; Unified Billing adds 5%; logging, Logpush, and guardrails have separate limits or meters.
Free trial: Ongoing free core features and Workers Free log storage, not a timed trial.

4. TrueFoundry AI Gateway: best managed semantic cache

TrueFoundry AI Gateway is the best managed choice when differently worded requests often ask the same safe-to-reuse question.

The distinction is semantic caching. TrueFoundry's cache documentation says exact mode hashes the complete request, while semantic mode embeds the final message and compares its meaning with prior requests. All other request parameters must still match. A full-response cache hit avoids the LLM invocation and carries zero LLM cost, so a support operation answering many variations of "How do I reset my password?" can reuse one approved answer instead of paying for each wording.

The cache is automatically isolated by user or virtual account, and optional namespaces can separate tenants or environments. That matters because a high hit rate is worthless if one customer's cached answer reaches another. A self-hosted semantic cache requires Redis and an embedding model; the SaaS route manages that infrastructure.

The public pricing page now has three tiers. Developer costs $0 for up to 3 users and includes 10,000 requests per user each month. Pro costs $25 per user per month, permits unlimited users, includes 20,000 requests per user, and charges $20 for each additional 100,000 requests. Enterprise is custom and adds private-deployment and SLA options.

The gateway page offers the first 30,000 requests each month across up to 3 seats with no card and access to 1,600+ models. That is a permanent Developer route, and it is enough to measure a realistic cache-hit rate before moving to Pro. It is not permission to infer a production saving from a synthetic FAQ set.

The current calculator makes the production hurdle concrete. At 7 users and 1 million monthly requests or tool calls, Pro costs $355: $175 for seats plus $180 for request overage. Against the normalized $800 provider bill, that separate traffic scenario creates a 44.375% inference-only hurdle. Pro can start at $25 for one user, so every buyer should rerun the calculation with its own seats and request volume rather than treating $355 as a flat tier price.

The real wall is false reuse. A semantic threshold that is too loose may return a prior answer that sounds relevant but violates the current user's context or policy. The safe test scores cached and fresh answers against the same acceptance set, segments by tenant, and refuses caching for questions with live account, legal, medical, or changing inventory context.

Best for: Support, knowledge, and FAQ workloads with repeated intent but varied wording.
Standout: Managed semantic full-response caching with automatic account isolation and optional namespaces.
Pricing: Developer $0 for up to 3 users; Pro $25/user/month plus $20 per additional 100,000 requests after the per-user allowance; Enterprise custom.
Free trial: The permanent Developer tier provides up to 30,000 requests/month at 3 users; Pro offers a trial without a published duration, and an instant sandbox requires no card.

5. Portkey: best managed governance step-up at $49

Portkey is the best low-entry managed option for a team that has outgrown a prototype and wants a $49 production entry with block-based overage instead of per-user-plus-overage pricing.

The plan ladder begins with Open Source, which is self-hosted with no request limit and includes a universal API, retries and timeouts, routing, guardrails, automatic fallbacks, a basic dashboard, load balancing, and community support. The page does not state a license price for this route, so it should be evaluated like LiteLLM: control is high, but infrastructure and operator cost remain with the buyer.

The hosted Developer tier is Free Forever and includes 10,000 recorded logs a month. Requests continue beyond that limit; additional logs are not recorded. It remains positioned for prototypes and enterprise proofs of concept rather than production. Production costs $49 a month and includes 100,000 recorded logs, with $9 overage per additional 100,000 requests; the comparison grid lists that schedule up to 3 million requests. It includes simple and semantic caching, role-based access, service-account keys, and production support.

At one million monthly recorded requests, Production costs $130 before model usage: $49 for the included 100,000 plus nine $9 overage blocks for the remaining 900,000. That is the useful comparison point. Portkey must recover more than $130 through caching, routing, budget enforcement, or operating work to beat a no-fee gateway on the normalized workload.

Enterprise uses custom pricing and request capacity. It adds SSO, granular budgets and rate limits, private-cloud or VPC options, and expanded compliance controls. A regulated organization may buy that tier for governance even when the provider invoice stays flat.

There is a metering wrinkle worth putting in a procurement note. The live pricing cards describe recorded logs, while the comparison grid still labels allowances and overages as requests. The $130 example assumes every request is recorded; confirm the billable unit in the order form. The page also says Portkey is now Prisma AIRS AI Gateway.

Portkey ranks below TrueFoundry for semantic-cache depth because its public materials do not make the savings mechanism as easy to audit. It ranks above OpenRouter for a team that wants a predictable hosted base fee instead of a percentage on every shared-credit purchase. The $49 entry point makes a controlled production trial financially legible.

Best for: A growing AI product that needs hosted caching, roles, keys, routing, and support on a three-digit monthly control-plane budget.
Standout: Production starts at $49 and reaches one million monthly recorded requests for $130 before model usage.
Pricing: Open Source self-hosted; Developer Free Forever; Production $49/month plus $9 per additional 100,000 requests; Enterprise custom.
Free trial: The ongoing Developer tier records 10,000 logs/month but is not positioned for production; traffic continues after logging stops.

6. OpenRouter: best marketplace for price-sorted provider routing

OpenRouter is the best gateway when the job is choosing among many providers and models by price, not minimizing the gateway line on an unchanged route.

The Free plan offers 25+ free models through 4 free providers, with 50 requests a day and community support. Standard, its entry pay-as-you-go plan, opens 500+ models and 80+ providers with no minimum spend and a 5.5% credit-purchase fee. Business is also pay-as-you-go and charges 8% on credit purchases. The FAQ lists a $0.80 minimum for the 5.5% fee; confirm payment-method charges at checkout.

Enterprise keeps the 500+ model and 80+ provider catalog, adds invoicing, managed controls, optional dedicated limits, a contractual SLA, and a shared Slack support channel. Pricing uses custom volume commitments and can include fee discounts. It is a procurement tier, not a public per-seat subscription.

BYOK changes the decision. Standard and Business include $25,000 a month of list-price BYOK inference without an OpenRouter fee, then charge 5% above it. Enterprise now has a custom allowance, with 5% above the contracted threshold. A company that already owns direct provider accounts can therefore use OpenRouter's control layer without paying the shared-credit percentage while it remains inside the allowance.

The cost feature is provider selection. OpenRouter can sort eligible providers by price, enforce a maximum price, and fall back when a route fails. Its response cache returns exact matching requests without billable usage and works across models, but it is not a semantic cache. Account-level zero-data-retention enforcement disables response caching. Price sorting saves money only if provider interchangeability holds for the workload's quality, data policy, latency, and availability requirements.

On an unchanged $800 provider bill, Standard shared credits add $44 and produce an $844 total. OpenRouter must therefore recover more than $44 through a cheaper eligible provider, cache hits, lower failure cost, or simpler operations to beat direct billing. If one provider already serves the workload reliably at the contracted rate, the marketplace is optional overhead.

The detailed fee floors, BYOK rules, funding risks, and routing comparison live in the OpenRouter pricing breakdown. The practical verdict here is narrower: OpenRouter is a savings tool when the team actively uses its market, not when it places the same request behind one more URL.

Best for: Teams that genuinely move among providers or model families and can measure the saving from that choice.
Standout: Price-sorted routing across 500+ models and 80+ providers, with large no-fee BYOK allowances.
Pricing: Free $0; Standard model prices plus 5.5% on credit purchases; Business 8%; Enterprise custom with fee discounts available.
Free trial: An ongoing Free plan with 25+ free models and 50 requests/day.

Who should pick what

Vercel is the default for a team that wants managed routing and no public token markup. With the model discount expired, the choice flips away from Vercel when its controls no longer save more than direct access, or when a team-wide policy and trace configuration makes the add-on bill material.

LiteLLM is the call when self-hosting is required or an existing platform team can absorb the gateway, including release pinning, patching, and key rotation, at low marginal cost. The choice flips to managed when LiteLLM creates a new operational owner, a new pager, or an Enterprise quote larger than the hosted alternative's complete bill.

Cloudflare wins when request bodies repeat exactly, the cached response is safe to reuse, and BYOK fits. The choice flips to TrueFoundry when varied wording carries the same intent and a measured semantic cache can clear the quality bar. It flips back to no cache when freshness, personalization, or tenant context makes reuse unsafe.

TrueFoundry wins managed semantic reuse only after a real evaluation proves enough accepted cache hits to cover its tier. Portkey wins when the team needs a lower-cost hosted production control plane and can live within its request economics. OpenRouter wins when provider price sorting or BYOK consolidation saves more than its fee.

The strongest default may still be no gateway. A stable, high-volume workload on one provider with good native budgets, low failure cost, and little repeated content has no obvious gateway saving. Add a control plane only when a named routing, cache, governance, or operating benefit exceeds its full cost.

How these were picked

The ranking weights the economics a buyer can verify: 35% for the public gateway cost and fee transparency, 30% for a credible mechanism that removes model spend, 20% for the production wall, and 15% for operating burden and migration friction.

Pricing pages, plan comparisons, feature documentation, cache behavior, request limits, and the GPT-5.6 Sol rates were live-verified on October 22, 2026. The existing pricing screenshots are September 17 snapshots; use the linked live pages for current terms. The tools were not exercised with production traffic in this run, so no latency, uptime, cache-hit, or output-quality result is presented as a test.

Six gateways made the final cut because each has a distinct, auditable cost job. Vercel removes the managed token markup. LiteLLM removes the license line through self-hosting. Cloudflare removes exact repeat calls. TrueFoundry targets semantic repeats. Portkey sells a legible managed step-up. OpenRouter exposes provider price competition.

A feature was not counted as a saving merely because a vendor describes it as optimization. The cost math uses a stated workload, a named price, and an explicit cache assumption. Any company applying the rankings should replace those inputs with its own model mix, accepted-output rate, and operator cost before signing an annual agreement.

The ones to avoid

Kong AI Gateway for a greenfield cost-first deployment

Kong AI Gateway is not the first choice when reducing inference cost is the only goal. Kong's current pricing gives Konnect a 30-day trial, then makes paid AI Gateway plugins an add-on on Plus and charges $100 a month for each unique model proxied by AI plugins, with a five-model limit. Five models create a $500 monthly proxy line before inference usage and other platform costs.

That does not make Kong a bad platform. A company already standardizing APIs, security, and service governance on Kong may find the incremental AI controls cheaper than adding another vendor. The advice is narrower: do not introduce a broad API platform into a greenfield AI stack solely to chase token savings when zero-markup and free-core gateways exist.

A homegrown proxy with no owner

A thin internal proxy looks free until it owns retries, provider changes, budgets, logs, secrets, and incidents. Build one only when the workload is stable enough to keep the surface small and a named team already owns the service. Otherwise the proxy hides its cost in engineering time while providing fewer controls than the managed options.

Any gateway on a one-provider workload with no reusable responses

Routing breadth has no value when traffic never moves. Exact and semantic caches have no value when every request needs a fresh answer. If native provider budgets and logs are sufficient, the direct API has the fewest billing layers and the smallest failure surface.

A permanent plan built on the September discount

Vercel's Sol promotion ended September 18, 2026. Keeping its former $400 sample month in a current budget understates the model line by $400. The architecture must work at today's $800 sample cost before any new discount is counted.

The Monday move: put one workflow on a cost-per-accepted-result sheet

Choose one production workflow that generated a meaningful model bill last week and can be judged with a binary acceptance rule. Export its input tokens, output tokens, provider spend, retries, and accepted results. Keep personal, confidential, and customer data out of any evaluation set that is not approved for the chosen route.

Route 10% of next week's traffic through the gateway whose mechanism matches the workload. Use Vercel for managed routing without token markup, Cloudflare for proven exact repeats, TrueFoundry for varied but equivalent questions, LiteLLM for an existing self-hosted platform, Portkey for a low-entry managed control layer, or OpenRouter for genuine provider price competition.

Set a hard budget on the isolated key and define a stop condition before the first request. Stop if accepted-result rate falls, cache safety fails, a provider route violates policy, or the total gateway line exceeds the savings. At the end of the week, divide model spend, platform fees, paid controls, retries, and review cost by accepted results.

The Monday output is one decision finance and engineering can share: keep, switch, or remove the gateway. If Vercel wins, record the current model rate and paid controls in the budget. If a cache wins, ship the cache only for the request classes that passed. If none wins, keep direct access and avoid paying for optional complexity.

Frequently asked questions

How much does Cloudflare AI Gateway cost?

Cloudflare's core AI Gateway analytics, caching, and rate limiting are free on all plans. Accounts whose first gateway predates September 24, 2026 keep Legacy Logs limits: 100,000 total on Free or 10 million per gateway on Paid. Newer accounts use Workers Logs allowances and retention. Unified Billing adds 5% to purchased credits, while provider prices pass through without markup. Logpush and guardrail inference can create separate charges.

Which AI gateway is the best?

Vercel AI Gateway remains the best managed starting point on October 22, 2026 for zero token markup, but its Sol discount has expired and savings must come from routing, controls, or operating work. LiteLLM fits teams that already operate self-hosted infrastructure and own pinning, patching, and key rotation; follow the incident guidance above before using affected deployments. Cloudflare fits exact-repeat caching, TrueFoundry managed semantic caching, Portkey a $49 production entry, and OpenRouter provider price sorting.

Is there a free plan for Cloudflare AI Gateway?

Yes. Core analytics, caching, and rate limiting are free. Free accounts that first created a gateway before September 24, 2026 keep 100,000 Legacy Logs across gateways; newer Free accounts use Workers Logs with 200,000 events/day and 3-day retention. The free account limit is 10 gateways. Provider inference still costs money, and optional Unified Billing, guardrails, or paid logging features can add charges.

Get the AI Tools Map for Business Owners

Turn this gateway decision into a complete adoption stack with a cost trigger and quality floor for every tool. Subscribe to get the AI Tools Map free.