Vercel AI Gateway

AI SDK TypeScript

Anthropic Messages API

Bring your own provider API key

Your provider bills model usage without a Gateway BYOK markup. System fallback uses Gateway credits; your provider-key spending is outside Gateway budgets.

Automatic prompt caching

Session cache affinity

Use a stable opaque x-session-affinity ID for provider cache locality; it does not change provider routing.

Explicit and implicit caching

Auto inserts cache markers for Anthropic direct/Vertex/Bedrock, MiniMax and Alibaba. OpenAI/Google/DeepSeek use implicit caching. Unset caching passes requests through. Anthropic 5-minute writes cost 1.25× base input and reads 0.1×; Responses 1h writes cost 2× with auto, reads remain 0.1×.

Model fallbacks

OpenResponses API

Optional features bill when enabled

Custom Reporting counts each unique tag/user ID/quota entity ID write within the request scope once. gateway.cost includes inference only, excluding reporting and ZDR surcharges. Trace Drains is billed through the Vercel plan; Pro has no trace/egress allowance and bills from the first trace/byte.

Call from local, CI and other clouds

ZDR and no-training filters

Team and request ZDR combine with OR; request false cannot bypass the team setting. No-training filtering is free for all users. Your BYOK credentials bypass filtering, while system fallback follows the filter.

Try Glyph Cluster free

Glyph Cluster anonymous early access currently has zero input/output pricing. The provider may retain and train on content; personal data is prohibited.

Pay as you go teams only; purchase Gateway Credits first and use stealth/glyph-cluster.

Jev free promotion

The historical free promotion has ended; current model pricing applies.

Historical Jev promotion.

Laguna S 2.1 Free

Laguna S 2.1 Free currently has zero input/output prices and is deprecated at 2026-10-31T00:00:00Z.

Free or Pay as you go teams with a valid payment method; use poolside/laguna-s-2.1-free.

Try Laya free

Laya and Laya Free via Boundless are free through 2026-10-31. Standard becomes paid afterward; -free stops serving.

Free or Pay as you go teams with a valid payment method; use the Laya or Laya Free route through Boundless.

Ends: 2026-10-31

Ling Sante free promotion

The historical free promotion has ended. Current input/output/cache-read prices are $0.075/$0.22/$0.015 per million tokens.

Historical Ling Sante promotion.

Try Nemotron free

Nemotron 3.5 Lightning Free is a non-production trial. All prompts and outputs are logged to improve models.

Free or Pay as you go teams with a valid payment method; use nvidia/nemotron-3.5-lightning-free.

Team provider allowlist

Requires Vercel Pro or Enterprise.

Provider order is a preference; only is an allowlist

The only filter runs before ordering. Providers omitted from order may still participate. Sort by cost, ttft or tps; health affects the order.

AI SDK Python beta

Gateway request rate limits

Per-model limits; fixed RPM/TPM values are not published.

Global, US or EU inference

Default global; zone requires us/eu. Unsatisfied regional constraints return 400 without cross-region fallback. Provider inference and storage follow the selected region; Gateway TLS and processing may occur in any Vercel region.

Per-request provider only has no surcharge

Per-request ZDR has no surcharge

Requires Vercel Pro or Enterprise.

Provider ordering, filtering and failover

Trace Drains

Requires Vercel Pro or Enterprise.

Disallow prompt training filter

Zero data retention filter

Requires Vercel Pro or Enterprise.

AI SDK TypeScript

Anthropic Messages API

Bring your own provider API key

Your provider bills model usage without a Gateway BYOK markup. System fallback uses Gateway credits; your provider-key spending is outside Gateway budgets.

Team has purchased Gateway Credits.

Automatic prompt caching

Session cache affinity

Use a stable opaque x-session-affinity ID for provider cache locality; it does not change provider routing.

Explicit and implicit caching

Auto inserts cache markers for Anthropic direct/Vertex/Bedrock, MiniMax and Alibaba. OpenAI/Google/DeepSeek use implicit caching. Unset caching passes requests through. Anthropic 5-minute writes cost 1.25× base input and reads 0.1×; Responses 1h writes cost 2× with auto, reads remain 0.1×.

Model fallbacks

OpenResponses API

Optional features bill when enabled

Custom Reporting counts each unique tag/user ID/quota entity ID write within the request scope once. gateway.cost includes inference only, excluding reporting and ZDR surcharges. Trace Drains is billed through the Vercel plan; Pro has no trace/egress allowance and bills from the first trace/byte.

Call from local, CI and other clouds

ZDR and no-training filters

Team and request ZDR combine with OR; request false cannot bypass the team setting. No-training filtering is free for all users. Your BYOK credentials bypass filtering, while system fallback follows the filter.

Try Glyph Cluster free

Glyph Cluster anonymous early access currently has zero input/output pricing. The provider may retain and train on content; personal data is prohibited.

Pay as you go teams only; purchase Gateway Credits first and use stealth/glyph-cluster.

Jev free promotion

The historical free promotion has ended; current model pricing applies.

Historical Jev promotion.

Laguna S 2.1 Free

Laguna S 2.1 Free currently has zero input/output prices and is deprecated at 2026-10-31T00:00:00Z.

Free or Pay as you go teams with a valid payment method; use poolside/laguna-s-2.1-free.

Try Laya free

Laya and Laya Free via Boundless are free through 2026-10-31. Standard becomes paid afterward; -free stops serving.

Free or Pay as you go teams with a valid payment method; use the Laya or Laya Free route through Boundless.

Ends: 2026-10-31

Ling Sante free promotion

The historical free promotion has ended. Current input/output/cache-read prices are $0.075/$0.22/$0.015 per million tokens.

Historical Ling Sante promotion.

Try Nemotron free

Nemotron 3.5 Lightning Free is a non-production trial. All prompts and outputs are logged to improve models.

Free or Pay as you go teams with a valid payment method; use nvidia/nemotron-3.5-lightning-free.

Team provider allowlist

Requires Vercel Pro or Enterprise.

Provider order is a preference; only is an allowlist

The only filter runs before ordering. Providers omitted from order may still participate. Sort by cost, ttft or tps; health affects the order.

AI SDK Python beta

Gateway request rate limits

No Gateway rate limits; upstream provider limits still apply.

Global, US or EU inference

Default global; zone requires us/eu. Unsatisfied regional constraints return 400 without cross-region fallback. Provider inference and storage follow the selected region; Gateway TLS and processing may occur in any Vercel region.

Per-request provider only has no surcharge

Per-request ZDR has no surcharge

Requires Vercel Pro or Enterprise.

Provider ordering, filtering and failover

Trace Drains

Requires Vercel Pro or Enterprise.

Disallow prompt training filter

Volume commitments and Enterprise invoicing

Contact sales for volume-commitment discounts. Enterprise supports invoice billing without payment-processing fees.

Zero data retention filter

Requires Vercel Pro or Enterprise.

Configurable auto-reload · ✓

Auto-reload enabled by default · —

Default route training use · Not used for training

Zero data retention · ✓

Default route training use · Not used for training

Zero data retention · ✓

Commercial use · Conditional

Selected-model and provider terms apply; Stealth prohibits personal data and trial routes may prohibit production use.

Payment method required · ✓

Processing fees · May apply

Free grants stop after purchase · ✓

Default training use

Gateway does not train on content; upstream defaults depend on the provider. Enable the no-training filter when needed.

Maximum output · 128000 tokens

Maximum output · 0 tokens

Maximum output · 0 tokens

Maximum output · 8000 tokens

Maximum output · 65535 tokens

Maximum output · 32000 tokens

Maximum output · 65000 tokens

Maximum output · 128000 tokens

Deprecation date · 2026-10-31T00:00:00Z

Maximum output · 32768 tokens

Maximum output · 256000 tokens

Maximum output · 0 tokens

Commercial use · Conditional

Selected-model and provider terms apply; Stealth prohibits personal data and trial routes may prohibit production use.

Processing fees · May apply

Default training use

Gateway does not train on content; upstream defaults depend on the provider. Enable the no-training filter when needed.

Data retention

Request bodies are deleted after completion and are not used for training. Metadata retains status, model, provider, tokens, cost, duration, authentication and routing. Upstream provider policies apply.

Routing log retention · 30 days

Route data retention

The trial endpoint logs all prompts and outputs to improve models; production and business-critical use is prohibited.

Route data retention

The provider may retain prompts and outputs and use them for training. Personal data is prohibited; ZDR and no-training filters are unavailable.