TokenMix Research Lab · 2026-04-10
Vertex AI Pricing Guide: Gemini, Claude, and Llama Model Costs on Google Cloud (2026)
Last Updated: 2026-04-29
Author: TokenMix Research Lab
Vertex AI Gemini 2.5 Pro at $1.25/$10 — 50% cheaper than GPT-4o, 58% cheaper than Sonnet. Claude priced at parity with Anthropic. Llama 70B at 55% premium ($1.36 vs Together $0.88). Google AI Studio free tier offers same Gemini for prototyping.
Vertex AI pricing is Google Cloud's gateway to enterprise AI, but understanding the actual cost requires navigating multiple pricing layers. Between on-demand and provisioned throughput, regional pricing variations, and the critical difference between Vertex AI and Google AI Studio's free tier, teams routinely overpay by 20-40% on their AI inference budget. TokenMix.ai pricing monitors show that Google AI Studio offers free Gemini 2.5 Pro access that many developers never discover, while Vertex AI's Claude pricing carries a measurable premium over Anthropic's direct API.
This guide covers Vertex AI pricing for Gemini, Claude, and Llama models -- with provisioned throughput economics, regional pricing, and direct comparisons against Google AI Studio and Anthropic's API.
Table of Contents
- Quick Comparison: Vertex AI vs Google AI Studio vs Direct API
- How Vertex AI Pricing Works
- Gemini Models on Vertex AI: Pricing Details
- Claude on Vertex AI: Pricing vs Direct API
- Llama and Open Models on Vertex AI
- Provisioned Throughput on Vertex AI
- Regional Pricing Differences
- Google AI Studio Free Tier: What You Get
- Cost Analysis for Different Workloads
- Which Vertex AI Setup Should You Pick?
- What's the Bottom Line on Vertex AI Pricing?
- FAQ
Quick Comparison: Vertex AI vs Google AI Studio vs Direct API
AI Studio = free for Gemini up to 25 RPM (Pro) or 500 RPM (Flash). Vertex AI = paid Gemini access + enterprise features. Direct Anthropic = same Claude price as Vertex but lower latency. Llama always cheaper outside Vertex.
| Model | Vertex AI (On-Demand) | Google AI Studio | Direct API | Notes |
|---|---|---|---|---|
| Gemini 2.5 Pro (input/1M) | $1.25 (<=200K) | Free tier: $0 | N/A | AI Studio free up to 25 RPM |
| Gemini 2.5 Pro (output/1M) | $10.00 | Free tier: $0 | N/A | Pay-as-you-go after free tier |
| Gemini 2.5 Flash (input/1M) | $0.15 | Free tier: $0 | N/A | Extremely cheap |
| Gemini 2.5 Flash (output/1M) | $0.60 | Free tier: $0 | N/A | Best budget option |
| Claude 3.5 Sonnet (input/1M) | $3.00 | N/A | $3.00 (Anthropic) | Same price as direct |
| Claude 3.5 Haiku (input/1M) | $0.80 | N/A | $0.80 (Anthropic) | Same price as direct |
| Llama 3.3 70B (input/1M) | $1.36 | N/A | $0.88 (Together) | +55% vs dedicated providers |
How Vertex AI Pricing Works
Four pricing modes: pay-as-you-go, provisioned throughput (20-45% off with 1-month/1-year commitment), batch prediction (50% off, 24h SLA), context caching (90% off cached input). Beyond tokens: image $0.001315/img, grounded search $35/1K queries, tuning extra.
Vertex AI uses a multi-layered pricing structure that varies by model family, usage tier, and commitment level.
Pricing Models
Pay-as-you-go (On-Demand): Per-token pricing, no commitments. Standard option for most teams.
Provisioned Throughput: Reserved capacity with guaranteed tokens-per-minute. Requires commitment but offers 20-40% savings at sustained usage.
Batch Prediction: 50% discount for async processing. Results within 24 hours. Available for Gemini and select models.
Context Caching: Reduced pricing for cached context (system prompts, repeated document context). Available for Gemini models.
Key Billing Components
- Token-based charges: Input and output tokens billed separately
- Image/video/audio input: Billed per image ($0.001315/image for Gemini 2.5 Pro) or per second of audio/video
- Context caching: Storage ($4.50/1M tokens/hour) + reduced input pricing ($0.315/1M cached tokens for Gemini 2.5 Pro)
- Grounding with Google Search: $35 per 1,000 grounded queries
- Tuning: Per training token + hosting costs
Gemini Models on Vertex AI: Pricing Details
Pro: $1.25/$10 (≤200K context), 2x premium over 200K. Flash: $0.15/$0.60. Thinking tokens billed at input rates — adds 40-60% to reasoning queries. Cache discount 75% off input. Batch 50% off both. Cheapest flagship in the market.
Gemini 2.5 Pro Pricing
| Metric | Price |
|---|---|
| Input (<=200K context) | $1.25/1M tokens |
| Input (>200K context) | $2.50/1M tokens |
| Output | $10.00/1M tokens |
| Thinking tokens (<=200K) | $1.25/1M tokens |
| Thinking tokens (>200K) | $2.50/1M tokens |
| Context cache (input) | $0.315/1M tokens |
| Context cache (storage) | $4.50/1M tokens/hour |
| Batch input | $0.625/1M tokens (50% off) |
| Batch output | $5.00/1M tokens (50% off) |
Important detail: Gemini 2.5 Pro uses "thinking tokens" for its reasoning process, and these are billed at input rates. For complex reasoning tasks, thinking tokens can add 30-100% to your effective input cost. This is comparable to how OpenAI o3 bills for reasoning tokens.
Gemini 2.5 Flash Pricing
| Metric | Price |
|---|---|
| Input (<=200K context) | $0.15/1M tokens |
| Input (>200K context) | $0.30/1M tokens |
| Output (non-thinking) | $0.60/1M tokens |
| Thinking output | $3.50/1M tokens |
| Context cache (input) | $0.0375/1M tokens |
| Batch input | $0.075/1M tokens (50% off) |
| Batch output | $0.30/1M tokens (50% off) |
Gemini 2.5 Flash is aggressively priced. At $0.15/1M input tokens, it is cheaper than GPT-4o mini ($0.15/1M), Claude 3.5 Haiku ($0.80/1M), and Amazon Nova Lite ($0.06/1M on output comparison). For high-volume, cost-sensitive workloads, Flash is one of the cheapest capable models available.
Gemini 2.0 Flash and Older Models
| Model | Input (per 1M) | Output (per 1M) | Status |
|---|---|---|---|
| Gemini 2.0 Flash | $0.10 | $0.40 | Available |
| Gemini 1.5 Pro | $1.25 | $5.00 | Available, being deprecated |
| Gemini 1.5 Flash | $0.075 | $0.30 | Available, being deprecated |
Gemini vs OpenAI vs Anthropic
| Model Tier | Gemini (Vertex AI) | OpenAI | Anthropic |
|---|---|---|---|
| Flagship (input) | Gemini 2.5 Pro: $1.25 | GPT-4o: $2.50 | Claude 3.5 Sonnet: $3.00 |
| Flagship (output) | Gemini 2.5 Pro: $10.00 | GPT-4o: $10.00 | Claude 3.5 Sonnet: $15.00 |
| Budget (input) | Gemini 2.5 Flash: $0.15 | GPT-4o mini: $0.15 | Claude 3.5 Haiku: $0.80 |
| Budget (output) | Gemini 2.5 Flash: $0.60 | GPT-4o mini: $0.60 | Claude 3.5 Haiku: $4.00 |
Gemini 2.5 Pro is 50% cheaper on input than GPT-4o and 58% cheaper than Claude 3.5 Sonnet. Output pricing matches GPT-4o and undercuts Claude by 33%. On paper, Gemini 2.5 Pro offers the best price-performance ratio among flagship models.
The caveat: thinking tokens. For reasoning-heavy tasks, Gemini 2.5 Pro's thinking tokens (billed at input rates) can significantly increase effective costs. TokenMix.ai analysis shows that for complex analytical queries, thinking tokens add an average of 40-60% to the base input cost.
Claude on Vertex AI: Pricing vs Direct API
Same per-token rates as Anthropic direct ($3/$15 Sonnet, $0.80/$4 Haiku). Trade-offs: Vertex adds 30-100ms latency, lags 1-2 weeks on new features, requires GCP setup. Choose Vertex for GCP integration + compliance, direct for latency + new features.
Claude Pricing on Vertex AI
| Model | Vertex AI Input (per 1M) | Vertex AI Output (per 1M) | Anthropic Direct | Difference |
|---|---|---|---|---|
| Claude 3.5 Sonnet v2 | $3.00 | $15.00 | $3.00 / $15.00 | 0% |
| Claude 3.5 Haiku | $0.80 | $4.00 | $0.80 / $4.00 | 0% |
| Claude 3 Opus | $15.00 | $75.00 | $15.00 / $75.00 | 0% |
| Claude 4 Sonnet | $3.00 | $15.00 | $3.00 / $15.00 | 0% |
Per-token pricing is identical between Vertex AI and Anthropic's direct API. Like AWS Bedrock, Google maintains pricing parity for Claude models.
Why Choose Claude on Vertex AI vs Direct Anthropic API?
Same token pricing, so the decision comes down to infrastructure:
Choose Vertex AI when:
- Your infrastructure is on Google Cloud
- You need GCP IAM, VPC Service Controls, and Cloud Audit Logs integration
- Compliance requires data to stay within GCP regions
- You want consolidated GCP billing
Choose direct Anthropic API when:
- You want lowest latency (Vertex adds a proxy layer, ~30-100ms additional latency per TokenMix.ai benchmarks)
- You need the latest Claude features immediately (Vertex AI availability sometimes lags by 1-2 weeks)
- You are not locked into Google Cloud infrastructure
Choose TokenMix.ai when:
- You want to route between Vertex AI Claude, direct Anthropic API, and Bedrock Claude based on availability and latency
- You need automatic failover if any single access point has issues
Llama and Open Models on Vertex AI
Vertex Llama 70B = $1.36/M, 55% premium over Together's $0.88. Less premium than AWS Bedrock's 201%, but still significant. Mixtral premium only 8%. For GCP-native teams: defensible. For everyone else: route Llama to dedicated providers via TokenMix.ai.
Google Cloud's Model Garden hosts open-source models on Vertex AI, including Llama, Mixtral, and others. Pricing is higher than dedicated inference providers.
Open Model Pricing on Vertex AI
| Model | Vertex AI (per 1M tokens) | Together AI | Groq | Vertex Premium |
|---|---|---|---|---|
| Llama 3.3 70B | $1.36 | $0.88 | $0.59 | +55% vs Together |
| Llama 3.3 8B | $0.20 | $0.18 | $0.05 | +11% vs Together |
| Llama 4 Scout | $0.25 / $0.72 | $0.18 / $0.59 | $0.11 / $0.34 | +39% vs Together |
| Mixtral 8x22B | $1.30 | $1.20 | $0.90 | +8% vs Together |
The Vertex AI premium for Llama models (55% for 70B) is lower than AWS Bedrock's premium (201%), but still significant. GCP-native teams may accept this premium for infrastructure integration. Cost-sensitive teams should route open-source model inference to dedicated providers.
Provisioned Throughput on Vertex AI
1-month: 20-30% off. 1-year: 35-45% off. Break-even ~$50/day single-model on-demand spend. 1-year commitment locks Gemini choice for that period — start with on-demand and switch only after 4+ weeks of consistent high usage.
How It Works
Provisioned throughput on Vertex AI guarantees a minimum tokens-per-minute (TPM) throughput for Gemini models. You pay a flat hourly rate regardless of actual usage, which is cheaper than on-demand at high sustained volumes.
Provisioned Throughput Pricing (Gemini 2.5 Pro)
| Commitment | Estimated Hourly Rate (per 1K TPM) | Savings vs On-Demand |
|---|---|---|
| No commitment (on-demand) | Varies by usage | Baseline |
| 1-month commitment | ~$1.50-2.50/hour per 1K TPM | ~20-30% |
| 1-year commitment | ~$1.00-1.80/hour per 1K TPM | ~35-45% |
Break-Even Analysis
For Gemini 2.5 Pro provisioned throughput to be cheaper than on-demand:
- Estimated break-even: sustained usage above ~$50/day on a single model
- At $100/day on-demand: provisioned saves approximately $25-35/day (1-month commitment)
- At $500/day on-demand: provisioned saves approximately $150-200/day (1-year commitment)
TokenMix.ai recommendation: Start with on-demand pricing. Switch to provisioned only when your Gemini usage consistently exceeds $50/day for at least 4 weeks. The 1-year commitment offers the best savings but locks you into Gemini for that period.
Regional Pricing Differences
US base price; EU +0-5%, Asia-Pacific +0-10%, South America +5-15%. Built into per-token rates (unlike AWS Bedrock's explicit 10% cross-region surcharge). Multi-region endpoints available without extra fee on Gemini.
Vertex AI pricing varies by Google Cloud region. Most model pricing is consistent across US regions, but international regions may carry premiums.
| Region Category | Price Modifier | Example Regions |
|---|---|---|
| US (standard) | Base price | us-central1, us-east4 |
| Europe | +0-5% | europe-west1, europe-west4 |
| Asia Pacific | +0-10% | asia-northeast1, asia-southeast1 |
| South America | +5-15% | southamerica-east1 |
Unlike AWS Bedrock's explicit 10% cross-region surcharge, Vertex AI's regional pricing variations are built into the per-token rates. Check the specific pricing page for your region before budgeting.
Multi-Region Considerations
Vertex AI supports multi-region endpoints for Gemini models. These route requests to the nearest available region for lower latency and higher availability. Unlike AWS Bedrock's explicit cross-region surcharge, Vertex AI does not add a separate fee for multi-region routing on Gemini models, though per-token prices may vary by the region that actually processes the request.
Google AI Studio Free Tier: What You Get
Free Gemini access at 25 RPM Pro / 500 RPM Flash / 1,500 RPM 2.0 Flash. Catch: data may be used for training, no SLA, no enterprise security. Use AI Studio for dev/prototyping, Vertex AI for production. Many teams pay Vertex when AI Studio would suffice.
Google AI Studio is the most overlooked free resource in the AI API market. It offers free access to Gemini models with generous rate limits -- and many teams are paying for Vertex AI when AI Studio would suffice.
AI Studio Free Tier Limits (April 2026)
| Model | Free Rate Limit | Context Window | Paid Rate (after free) |
|---|---|---|---|
| Gemini 2.5 Pro | 25 RPM / 50 RPD | 1M tokens | $1.25/$10.00 per 1M |
| Gemini 2.5 Flash | 500 RPM | 1M tokens | $0.15/$0.60 per 1M |
| Gemini 2.0 Flash | 1,500 RPM | 1M tokens | $0.10/$0.40 per 1M |
AI Studio Free Tier vs Vertex AI
| Feature | Google AI Studio (Free) | Vertex AI (Paid) |
|---|---|---|
| Price | Free (within limits) | Per-token billing |
| Rate limits | 25-1,500 RPM by model | Higher (scalable) |
| SLA | None | 99.9% SLA available |
| Data privacy | Data may be used for improvement | Data not used for training |
| Enterprise features | None | IAM, VPC, audit logs |
| Compliance | Basic | SOC 2, HIPAA eligible |
| Claude/Llama access | Not available | Available |
For prototyping, development, and small-scale production (under 25 requests/minute for Gemini 2.5 Pro), Google AI Studio is free. The catch: no SLA, no enterprise security features, and Google may use your data for model improvement.
TokenMix.ai recommendation: Use AI Studio for development and testing. Migrate to Vertex AI for production workloads that need reliability guarantees, data privacy, and compliance. The price difference is zero for Gemini models -- you are paying for enterprise features, not model access.
Cost Analysis for Different Workloads
Dev 2M Pro tokens: AI Studio free vs Vertex $22 vs OpenAI $25. Production 50M Flash: Vertex $38 vs Haiku $240 (5x cheaper Flash). Enterprise 500M mixed: Vertex-only $6,800 vs hybrid $4,200 (Vertex Gemini/Claude + Together Llama).
Developer/Prototype (2M tokens/month, Gemini 2.5 Pro)
| Platform | Monthly Cost | Notes |
|---|---|---|
| Google AI Studio (free tier) | $0 | Within free limits for most dev work |
| Vertex AI (on-demand) | $22 | Enterprise features included |
| OpenAI GPT-4o | $25 | Direct API |
| Anthropic Claude 3.5 Sonnet | $36 | More expensive on output |
| TokenMix.ai (routed) | $18-22 | Cheapest model per query |
Production Application (50M tokens/month, Gemini 2.5 Flash)
| Platform | Monthly Cost | Notes |
|---|---|---|
| Vertex AI (on-demand) | $38 | Flash is extremely cheap |
| Vertex AI (batch) | $19 | 50% off for async workloads |
| OpenAI GPT-4o mini | $38 | Same input, same output price |
| Anthropic Claude 3.5 Haiku | $240 | 5x more expensive |
| TokenMix.ai (routed) | $30-35 | Includes Gemini Flash routing |
Gemini 2.5 Flash is remarkably cost-effective for budget-conscious production workloads. At $0.15/$0.60 per million tokens, it matches GPT-4o mini pricing while offering a 1M token context window.
Enterprise Mixed Workload (500M tokens/month, multi-model)
| Setup | Monthly Cost | Notes |
|---|---|---|
| Vertex AI only (Gemini + Claude + Llama) | $6,800 | Llama premium hurts |
| Vertex AI (Gemini + Claude) + Together AI (Llama) | $4,200 | Optimal cost split |
| All via TokenMix.ai | $4,000-4,500 | Smart routing across providers |
| OpenAI + Anthropic direct | $5,500 | No Gemini Flash savings |
Which Vertex AI Setup Should You Pick?
Prototyping with Gemini: AI Studio free tier. GCP-native production: Vertex AI. Cheapest capable model: Vertex Gemini Flash ($0.15/$0.60). Llama on GCP: Together via TokenMix.ai (55% cheaper). Highest reasoning quality: Vertex Gemini 2.5 Pro.
| Your Situation | Best Choice | Why |
|---|---|---|
| Prototyping with Gemini | Google AI Studio (free) | Free tier is generous |
| GCP-native production | Vertex AI | IAM, VPC, compliance integration |
| Cheapest capable model needed | Vertex AI (Gemini 2.5 Flash) | $0.15/$0.60 per 1M, hard to beat |
| Claude on Google Cloud | Vertex AI | Same price as direct, GCP integration |
| Llama on Google Cloud | Together AI via TokenMix.ai | 55% cheaper than Vertex AI |
| Multi-provider cost optimization | TokenMix.ai | Route by cost and performance |
| Data must stay in GCP | Vertex AI | Only option with GCP data residency |
| Need highest reasoning quality | Vertex AI (Gemini 2.5 Pro) | $1.25 input -- 50% cheaper than GPT-4o |
| Batch processing at scale | Vertex AI (batch API) | 50% discount, 24-hour turnaround |
Related: Compare all model pricing in our complete LLM API pricing comparison
What's the Bottom Line on Vertex AI Pricing?
Cost leader for Gemini-first workloads. Parity for Claude. Premium for Llama. Strategy: AI Studio for dev → Vertex for production Gemini → TokenMix.ai for Llama and other models. Don't pay Vertex Llama premium unless GCP residency is mandatory.
Vertex AI pricing is competitive for Gemini models, at parity for Claude, and premium for Llama. The strategic play depends on your model mix and infrastructure.
Gemini 2.5 Pro at $1.25/1M input tokens is the cheapest flagship model across major providers -- 50% less than GPT-4o, 58% less than Claude 3.5 Sonnet. Gemini 2.5 Flash at $0.15/$0.60 is one of the best budget models available. These prices make Vertex AI the cost leader for Gemini-first workloads.
For Claude, Vertex AI matches Anthropic's direct pricing -- the choice is purely about infrastructure preference. For Llama, Vertex AI carries a 55% premium versus dedicated providers, making it a poor cost choice unless GCP data residency is mandatory.
The wildcard is Google AI Studio's free tier. Many developers are paying for model access they could get for free during development. TokenMix.ai recommends starting on AI Studio, graduating to Vertex AI for production, and routing non-Gemini workloads through TokenMix.ai for unified access at optimized pricing.
Compare real-time Vertex AI pricing against all alternatives at TokenMix.ai.
FAQ
Is Vertex AI cheaper than Google AI Studio?
No. Google AI Studio offers free access to Gemini models with rate limits (25 RPM for Gemini 2.5 Pro, 500 RPM for Gemini 2.5 Flash). Vertex AI uses per-token billing with no free tier. For development and low-volume use, AI Studio is free. Vertex AI adds enterprise features (SLA, data privacy, compliance) worth paying for in production.
How does Vertex AI Gemini pricing compare to OpenAI GPT-4o?
Gemini 2.5 Pro input costs $1.25/1M tokens versus GPT-4o's $2.50/1M -- 50% cheaper. Output pricing is the same at $10.00/1M tokens. However, Gemini 2.5 Pro uses thinking tokens for reasoning tasks that add 40-60% to effective input costs. For simple queries, Gemini is clearly cheaper. For complex reasoning, the cost gap narrows.
Is Claude cheaper on Vertex AI or through Anthropic's direct API?
The per-token price is identical: $3.00/$15.00 per million input/output tokens for Claude 3.5 Sonnet. Choose Vertex AI for GCP integration and compliance. Choose Anthropic's direct API for lower latency (30-100ms less overhead) and immediate access to new features.
What is provisioned throughput on Vertex AI and when does it save money?
Provisioned throughput reserves guaranteed tokens-per-minute capacity at a flat hourly rate. It saves 20-45% compared to on-demand pricing at sustained high usage. The break-even point is approximately $50/day on-demand spend on a single model. Below that, on-demand is cheaper.
Can I access Vertex AI models through TokenMix.ai?
Yes. TokenMix.ai provides unified API access to Gemini models alongside Claude, GPT-4o, Llama, and 300+ other models. This enables smart routing that sends budget queries to Gemini Flash and complex queries to Claude or GPT-4o, optimizing cost across your entire workload.
Why is Llama more expensive on Vertex AI than on Together AI?
Vertex AI charges a managed service premium for hosting open-source models on GCP infrastructure. Llama 3.3 70B costs $1.36/1M tokens on Vertex AI versus $0.88 on Together AI -- a 55% premium. The premium buys GCP integration, security features, and compliance certifications. For teams without GCP requirements, dedicated inference providers offer better Llama pricing.
Author: TokenMix Research Lab | Last Updated: April 2026 | Data Source: Google Vertex AI Pricing, Google AI Studio, Anthropic Pricing + TokenMix.ai