Reference table. Vendor rates read between and ; each row carries its date.
2026 AI vendor pricing and TCO report
What this is: 32 public pricing data points across the four cost categories a solo AI founder actually pays for — LLM APIs, vector databases, AI customer-support agents, and LLM observability. Every rate below was read from the vendor's own pricing page, cited to that page (22 official sources) and dated on its row; the Anthropic, OpenAI and Google rows are the current lineup of the model catalog the hub's calculators share. Sales-gated vendors are marked not publicly documented — this report never estimates a number a vendor refuses to publish. Underneath the tables, two total-cost-of-ownership scenarios are computed live by our calculator engines and recomputed in CI, so no cost figure here is hand-typed.
LLM APIs
Pricing unit: Per million tokens
| Vendor / plan | Type | Published rate | Source | Verified |
|---|---|---|---|---|
| Anthropic Claude Fable 5.1 | Flagship; rate for 5-minute cache writes | $10.00 in / $50.00 out | platform.claude.com/docs/en/about-claude/pricing | |
| Anthropic Claude Opus 5.5 | Flagship; rate for 5-minute cache writes | $4.00 in / $20.00 out | platform.claude.com/docs/en/about-claude/pricing | |
| Anthropic Claude Sonnet 5.5 | Workhorse; rate for 5-minute cache writes | $2.00 in / $10.00 out | platform.claude.com/docs/en/about-claude/pricing | |
| Anthropic Claude Haiku 5.5 | Budget; rate for prompts up to 100K tokens; 5-minute cache writes | $0.10 in / $0.50 out | platform.claude.com/docs/en/about-claude/pricing | |
| OpenAI GPT-6 Astra | Flagship; rate for prompts up to 272K input tokens | $10.00 in / $50.00 out | developers.openai.com/api/docs/models/gpt-6-astra | |
| OpenAI GPT-6.1 Sol | Workhorse; rate for prompts up to 272K input tokens | $2.00 in / $10.00 out | developers.openai.com/api/docs/models/gpt-6.1-sol | |
| OpenAI GPT-6 Luna | Budget; rate for prompts up to 272K input tokens | $0.10 in / $0.50 out | developers.openai.com/api/docs/models/gpt-6-luna | |
| OpenAI GPT-5.5 | Flagship; rate for prompts up to 272K input tokens | $5.00 in / $30.00 out | developers.openai.com/api/docs/models/gpt-5.5 | |
| OpenAI GPT-5.4 Mini | Workhorse | $0.75 in / $4.50 out | developers.openai.com/api/docs/models/gpt-5.4-mini | |
| Google Gemini 4 Argon | Flagship; rate for promotional price through November 30, 2026 | $2.00 in / $10.00 out | ai.google.dev/gemini-api/docs/pricing | |
| Google Gemini 3.8 Flash | Workhorse; rate for promotional price through December 31, 2026 | $0.75 in / $3.75 out | ai.google.dev/gemini-api/docs/pricing | |
| Google Gemini 3.5 Flash-Lite | Budget | $0.30 in / $2.50 out | ai.google.dev/gemini-api/docs/pricing | |
| Google Gemini 3.1 Pro Preview | Flagship; rate for prompts up to 200K tokens | $2.00 in / $12.00 out | ai.google.dev/gemini-api/docs/pricing | |
| DeepSeek V4.1-Flash | Budget; peak rate, off-peak half | $0.30 in (cache miss) / $1.20 out | api-docs.deepseek.com/quick_start/pricing | |
| DeepSeek V4-Pro | Flagship; peak rate, off-peak half | $1.32 in (cache miss) / $3.96 out | api-docs.deepseek.com/quick_start/pricing | |
| Mistral Medium 3.5 | Flagship | $1.50 in / $7.50 out | docs.mistral.ai/models/mistral-medium-3-5-26-04 | |
| Mistral Small 4 | Budget | $0.15 in / $0.60 out | docs.mistral.ai/models/mistral-small-4-0-26-03 |
Vector databases
Pricing unit: Per-GB plus operations
| Vendor / plan | Type | Published rate | Source | Verified |
|---|---|---|---|---|
| Pinecone (Serverless Standard) | Per read/write/storage | $16–$18/M reads, $4–$4.50/M writes, $0.33/GB-mo, $50/mo min | pinecone.io/pricing | |
| Pinecone Starter | Free tier | 2 GB storage, 1M reads/mo, 2M writes/mo | pinecone.io/pricing | |
| Qdrant Cloud | Usage-based (compute/memory/disk, hourly) | Flat per-GB rate not publicly documented (calculator only) | qdrant.tech/pricing | |
| Qdrant Cloud | Free tier | 0.5 vCPU / 1 GB RAM / 4 GB disk, single node | qdrant.tech/pricing | |
| Weaviate (Flex) | Per 1M stored vector dimensions | from $0.00465/M dimensions, $45/mo minimum | weaviate.io/pricing | |
| Weaviate Free | Free tier, one cluster per user | 100,000 objects / 1 GB memory / 10 GB disk | weaviate.io/pricing |
AI customer-support tools
Pricing unit: Per resolution or per seat
| Vendor / plan | Type | Published rate | Source | Verified |
|---|---|---|---|---|
| Intercom Fin AI Agent | Outcome-based (per resolution) | $0.99 per resolved outcome, 50-outcome/mo minimum off-platform | fin.ai/pricing | |
| Zendesk Suite | Per-seat base plans | from $19/agent/mo; AI-agent per-resolution rate not on the public pricing page | zendesk.com/pricing | |
| Decagon | Enterprise, outcome-based | Not publicly documented (no public pricing page) | decagon.ai | |
| Sierra | Outcome-based (per successful resolution) | Per-outcome rate not publicly documented (no public pricing page) | sierra.ai |
LLM observability
Pricing unit: Per event or per seat
| Vendor / plan | Type | Published rate | Source | Verified |
|---|---|---|---|---|
| Langfuse | Per-unit (observation), platform tier | Hobby free 50k units/mo; Core $29/mo; Pro $199/mo; overage from $8/100k units | langfuse.com/pricing | |
| Braintrust | Per-GB processed + per-score | Starter free (1 GB, 10k scores); Pro $249/mo; overage from $3/GB, $1.50/1k scores | braintrust.dev/pricing | |
| Arize Phoenix | Open-source (ELv2), self-host | Free self-host; 2 free Phoenix Cloud instances (limits not published) | arize.com/phoenix | |
| Arize AX | Per-span + per-GB ingestion | AX Free 25k spans/mo; AX Pro $50/mo; overage $0.0008/span, $3/GB | arize.com/pricing | |
| Helicone | Per-request, platform tier | Hobby free 10k requests/mo; Pro $79/mo; Team $799/mo (per-request overage not published) | helicone.ai/pricing |
Computed TCO scenarios
Published rates tell you the unit price; total cost of ownership tells you the bill. The two scenarios below are computed live by the shipped engine bundles — not by us — and every input and output is rendered straight from the engine return value and recomputed in CI. Expand each block to see the exact inputs and the per-line cost breakdown.
Scenario 1 — Full AI SaaS stack on Claude Opus 5.5
A seed-stage AI SaaS on Vercel Pro hosting, Supabase Pro, Clerk auth, Resend Pro email, and Sentry Team monitoring, calling Claude Opus 5.5 at 1,200 input / 600 output tokens, eight calls per user per day. The engine projects total monthly cost and per-user cost across 100, 1k, 10k, and 100k users — and shows how completely the LLM API line dominates the bill as you scale.
Show the recompute-verified inputs and outputs
| hosting_index | 1 |
|---|---|
| database_index | 1 |
| auth_index | 0 |
| ai_model_index | 2 |
| avg_input_tokens | 1200 |
| avg_output_tokens | 600 |
| api_calls_per_user_per_day | 8 |
| email_index | 1 |
| monitoring_index | 2 |
| domain_cost_yearly | 15 |
| tiers › row 1 › users | 100 |
|---|---|
| tiers › row 1 › hosting | 20 |
| tiers › row 1 › database | 25 |
| tiers › row 1 › auth | 0 |
| tiers › row 1 › ai api | 403.2 |
| tiers › row 1 › email | 20 |
| tiers › row 1 › monitoring | 26 |
| tiers › row 1 › domain | 1.25 |
| tiers › row 1 › other | 0 |
| tiers › row 1 › total | 495.45 |
| tiers › row 1 › cost per user | 4.95 |
| tiers › row 2 › users | 1000 |
| tiers › row 2 › hosting | 20 |
| tiers › row 2 › database | 25 |
| tiers › row 2 › auth | 0 |
| tiers › row 2 › ai api | 4032 |
| tiers › row 2 › email | 20 |
| tiers › row 2 › monitoring | 26 |
| tiers › row 2 › domain | 1.25 |
| tiers › row 2 › other | 0 |
| tiers › row 2 › total | 4124.25 |
| tiers › row 2 › cost per user | 4.12 |
| tiers › row 3 › users | 10000 |
| tiers › row 3 › hosting | 20 |
| tiers › row 3 › database | 25 |
| tiers › row 3 › auth | 0 |
| tiers › row 3 › ai api | 40320 |
| tiers › row 3 › email | 20 |
| tiers › row 3 › monitoring | 26 |
| tiers › row 3 › domain | 1.25 |
| tiers › row 3 › other | 0 |
| tiers › row 3 › total | 40412.25 |
| tiers › row 3 › cost per user | 4.04 |
| tiers › row 4 › users | 100000 |
| tiers › row 4 › hosting | 20 |
| tiers › row 4 › database | 25 |
| tiers › row 4 › auth | 1025 |
| tiers › row 4 › ai api | 403200 |
| tiers › row 4 › email | 120 |
| tiers › row 4 › monitoring | 26 |
| tiers › row 4 › domain | 1.25 |
| tiers › row 4 › other | 0 |
| tiers › row 4 › total | 404417.25 |
| tiers › row 4 › cost per user | 4.04 |
| dominant driver | AI API |
| dominant driver percent | 99.77 |
| insight | AI API is 99.77% of your costs at 10K users. Consider caching responses, using a cheaper model for common queries, or batching requests. |
| selected › users | 1000 |
| selected › hosting | 20 |
| selected › database | 25 |
| selected › auth | 0 |
| selected › ai api | 4032 |
| selected › email | 20 |
| selected › monitoring | 26 |
| selected › domain | 1.25 |
| selected › other | 0 |
| selected › total | 4124.25 |
| selected › cost per user | 4.12 |
| selected driver | AI API |
| selected driver percent | 97.76 |
| selected insight | AI API is 97.76% of your costs at 1K users. Consider caching responses, using a cheaper model for common queries, or batching requests. |
Computed live at build time.
Scenario 2 — Vector database for a 500k-vector RAG app
A production RAG app holding 500,000 vectors at 1,536 dimensions, serving 5,000 queries and ingesting 2,000 documents per day. The engine prices Pinecone, Postgres with pgvector on Supabase, LanceDB on R2, and Turbopuffer for that exact workload and names the cheapest — a useful cross-check against the vector-database rates in the table above.
Show the recompute-verified inputs and outputs
| vector_count | 500000 |
|---|---|
| dim | 1536 |
| queries_per_day | 5000 |
| ingest_per_day | 2000 |
| vendors › row 1 › vendor | Pinecone |
|---|---|
| vendors › row 1 › monthly cost | 50 |
| vendors › row 1 › notes | Pinecone Standard 2026-10: $16/M read units, $4/M write units, $0.33/GB-mo, $50/mo plan minimum; a query here uses about 3.58 read units (1 per GB of namespace, minimum 0.25). Small workloads can fit the free Starter plan (up to 2 GB, 1M read and 2M write units a month) or Builder ($20/mo flat, up to 10 GB, 2M read and 5M write units). |
| vendors › row 2 › vendor | Postgres+pgvector |
| vendors › row 2 › monthly cost | 225 |
| vendors › row 2 › notes | Supabase Pro 2026-10: $25/mo with $10 compute credit, XL compute ($210/mo, 16 GB RAM), disk above 8 GB at $0.125/GB-mo for about 7.15 GB (table plus HNSW index). Sized from Supabase's pgvector HNSW benchmarks (1536-d: Micro 15k vectors, Small 50k, Medium 100k, Large 224k, XL 500k, 2XL 1M), scaled by floats per vector and by RAM above 2XL. |
| vendors › row 3 › vendor | LanceDB |
| vendors › row 3 › monthly cost | 0.38 |
| vendors › row 3 › notes | LanceDB on Cloudflare R2 2026-10: $0.015/GB-mo storage, $4.50/M writes (Class A), $0.36/M reads (Class B), one request per upsert and per query. Self-hosted compute not included. |
| vendors › row 4 › vendor | Turbopuffer |
| vendors › row 4 › monthly cost | 16 |
| vendors › row 4 › notes | Turbopuffer Launch 2026-10: $16/mo minimum usage; $0.33/GB-mo logical storage, $2/GB written, $0.001/TB of data queried (namespace size, 1.28 GB minimum, discounted above 32 GB), $0.05/GB returned. |
| cheapest vendor | LanceDB |
| cheapest monthly cost | 0.38 |
| cheapest all in vendor | Turbopuffer |
| cheapest all in monthly cost | 16 |
| storage gb | 3.58 |
Computed live at build time.
How to use this report
- Find your unit first. LLM APIs bill per token, vector DBs per GB plus operations, support agents per resolution or seat, observability per event or seat. Two tools that cost the same on paper diverge once your real workload meets the meter.
- Treat "not publicly documented" as a flag, not a gap. A vendor that won't publish a rate is telling you to budget for a sales process and a negotiated contract, not a self-serve price.
- Model your own numbers. The TCO scenarios use one fixed input set; your token volume, query rate, and user count are different. Run them in the AI Stack Cost Calculator and the Embeddings DB Cost calculator.
- Re-verify before you commit. AI rates move often. Each row carries the date it was read; check the linked vendor page for the current number before signing.
For the category-by-category playbook on reading each pricing model, see the AI Vendor Pricing guide. For a deep dive on the vector layer, read Pinecone serverless pricing vs Qdrant, and to turn vendor cost into a defensible price for your own product, use the AI Product Economics guide.
Provenance & method
- Pricing tables: every rate was read from the vendor's own pricing page on the date in its row, the URL is cited in the Source column, and the CSV carries the same URL and date per row. The Anthropic, OpenAI and Google rows come from the model catalog aidevhub.io collects and verifies, which CI keeps in step with that source. Where a rate quotes a range (e.g. Pinecone reads at $16–$18/M), the range is the vendor's own published range, which varies by cloud and region.
- TCO scenarios: computed by the shipped engine bundles at
/engines/ai-stack-cost-calculator.jsand/engines/embeddings-db-cost.js. Each engine is pure — same inputs, same output — and CI independently recomputes the embedded result on every build. No TCO number was hand-typed. - No fabricated benchmarks: this report aggregates cited public numbers and computes scenarios from them. It does not claim to have run a private trading or usage experiment; nothing here is a fabricated executed measurement.
FAQ
What is the cheapest flagship LLM API in 2026?
Among the current flagship-tier models, Google Gemini 4 Argon is the cheapest at $2.00 in / $10.00 out per million tokens, $4.00 blended at a 3:1 input-to-output mix. The dearest, Anthropic Claude Fable 5.1 and OpenAI GPT-6 Astra, cost $10.00 in / $50.00 out, $20.00 blended. For most production traffic a workhorse model such as Google Gemini 3.8 Flash ($0.75 in / $3.75 out) costs far less than any flagship; check that it clears your quality bar first.
Why do some AI support vendors show no price?
AI customer-support agents split into self-serve and enterprise. Intercom Fin publishes an outcome-based rate of $0.99 per resolution, but Decagon and Sierra publish no pricing at all — every figure that circulates for them is a third-party estimate, not a vendor-documented rate. This report labels those "not publicly documented" rather than guessing.
How do you compute total cost of ownership for an AI stack?
Add every recurring line: hosting, database, auth, the LLM API bill (input plus output tokens times calls), email, and monitoring, projected at the user count you expect. For most AI apps the LLM API bill dominates as you scale, so the model and token volume you choose set your unit economics. The two scenarios above are computed by the live AI Stack Cost and Embeddings DB Cost engines.