Your invoice fits in one cell
Flat monthly fee. The invoice fits in a single cell. No surprises on Black Friday or end of quarter. Your CFO signs without flinching and you stop defending variance to the finance committee.
OpenAI alternative · same way you integrate it
Your AI bill, fixed every month.
Put AI in your product with a flat monthly cost, data in Europe, LATAM or the US, and the same code you already use with OpenAI. No surprises on the next invoice, no lock-in.
python · three lines to try
from openai import OpenAI
client = OpenAI(
base_url="https://api.tesseraai.cloud/v1",
api_key="sk-tessera-…",
) Why Tessera
Three pillars. No fine print.
Data in your region, not theirs
GPU physically in EU, LATAM or US, your choice. Your data never crosses a jurisdiction you didn’t sign for. GDPR and AI Act by architecture for EU; data residency guaranteed for US and LATAM. DPA available across all tiers, public subprocessor list.
If you leave, you leave in an afternoon
OpenAI v1-compatible API. Open-source models with Apache 2.0 license. If you decide to leave, you leave in an afternoon. We earn the renewal every month.
How it works
Three steps. One afternoon.
-
Step 1: Pick your tier and region
EU for GDPR-bound companies, LATAM for South-American sovereignty, US for companies that prefer American residency. Residency is contracted, not discovered on a status page.
-
Step 2: Change one line in your code
Your OpenAI client stays the same. It just points to api.tesseraai.cloud. The rest of the SDK, your Langchain or LlamaIndex code and your prompts stay untouched.
-
Step 3: Pay a flat monthly fee
One invoice, one cell. No surprises on traffic spikes. We earn the renewal every month; no exit clauses to negotiate.
Trying takes 5 minutes and 3 lines of code. Full migration with tests, one day with founder-led support.
Honest comparison
How we compare with the rest.
No asterisks. If something doesn’t apply, we put a dash.
Measured benchmarks
Numbers we measured, not promised.
Metrics taken in real production, with 25 simultaneous customers running for almost 2 hours. No per-token competitor publishes their own infrastructure benchmarks. We do.
TTFT P95
350 ms
qwen3.6-35b-a3b
Success rate
99.95 %
8,000 requests · 0 generation errors
Included capacity
100 RPM
per Pro customer · burst 200
Tokens validated
5.91 M
on a 1 h 50 min sustained run
Measured on 2026-04-27 on RTX PRO 6000 Blackwell with 25 simultaneous customers. Full report (saturation curve, noisy neighbor, long context) available under NDA.
Speech · new
A 40-minute meeting, transcribed with every voice separated, in one minute.
New GPU transcription models, included in every plan. On real Spanish calls they make fewer mistakes than Whisper, the market standard, they don’t invent sentences, and they tell speakers apart without being told how many there are.
Error on Spanish calls
6.6 %
Canary 1B v2 · Whisper large-v3: 8.7%
Invented text
0.5 %
Canary and Parakeet · Whisper large-v3: 2.3%
Faster than Whisper
100 ×
Parakeet v3: 1 h of audio in about 12 s
Right speaker
96 %
of words, in conversations of 2 to 4 people
Measured by Tessera in October 2026, with every model on the same GPU: real Spanish calls of 8 and 41 min, English interviews of up to 79 min and the public FLEURS and VoxConverse sets. Error = words that differ from a reference transcript.
See the transcription benchmark and API Models included
One bundle. Eight open models. Zero upsell.
Every tier accesses the full catalogue. We don’t bill per model. There’s no "premium" tier hiding the good model behind a paywall.
chat · fast
Qwen 3.6-35B-A3B
The fast one. It activates only 3B parameters per word, so it answers sooner and handles more volume. For assistants, answers over your documents, classification, extraction and summaries. Text only, direct and thinking modes.
Apache 2.0
chat · agents · vision
Qwen 3.8-27B
The most capable: it nearly doubles Qwen 3.6 on the Artificial Analysis intelligence index (34 versus 18). For agents that use tools, code, multi-step tasks and complex instructions. It also reads images: invoices, screenshots, photos.
Apache 2.0
audio · transcription
Canary 1B v2
The most accurate on Spanish: on real calls it misses 6.6% of words, versus 8.7% for Whisper. Tells speakers apart in calls, interviews and meetings. 25 European languages.
CC-BY-4.0
audio · transcription
Parakeet TDT 0.6B v3
The fastest: transcribes an hour of audio in about 12 seconds and detects the language itself. Beats Canary on English, and also tells speakers apart. 25 European languages.
CC-BY-4.0
audio · transcription
Whisper large-v3 + turbo
The market standard, for integrations that already use it: 99 languages, Catalan included, and translation into English. large-v3 for accuracy, turbo for speed.
MIT
audio · synthesis
Kokoro 82M TTS
Natural voice synthesis with strong Spanish coverage. Sub-200 ms latency, ideal for IVR and conversational agents.
Apache 2.0
embeddings
Qwen3-Embedding-8B
Embeddings for retrieval, clustering and semantic search. 4,096 dimensions, multilingual, optimized for long contexts.
Apache 2.0
reranker
Qwen3-Reranker-4B
Second-stage RAG reranking. Trained jointly with Qwen3-Embedding-8B (same family, no cross-vendor penalty). Cohere-compatible response shape — drop-in migration from Cohere / Voyage / Jina.
Apache 2.0
When we ship a new model, we tell you a month in advance. 12-month model-freeze clause with opt-in free upgrade.
Pricing
One invoice. One cell.
Flat fee, dedicated GPU on Pro and above, no token-meter or seasonality surcharges.
no commitment
€55/mo
For founders and 1-2 person teams.
- Private Hermes assistant on WhatsApp and Telegram, managed by us
- 20 sustained RPM · 32 K context
- Full bundle: chat + voice + embeddings, no token meter
- Thinking mode: unlimited
- Founders community + monthly training + weekly podcast
- Cancel anytime · first-month guarantee
- Founding price locked for life
- No SLA · community support
€200/mo
For overnight processing and batch jobs.
- No RPM, job queue
- 16K context
- LLM + embeddings
- SLA <30s P95 (best-effort)
- No transcription or TTS
€450/mo
For small teams with a single use case.
- 50 sustained RPM · 100 burst (5 min/h)
- Full bundle: LLM + embeddings + transcription + TTS
- Sublimits: embeddings 100 RPM · transcription 10 RPM · TTS 10 RPM
- 8K context default · up to 32K configurable
- Thinking mode: unlimited
- Email · <24h business-hour response
- Region: EU, LATAM or US
€650/mo
For typical mid-market production workloads.
- 200 sustained RPM · 400 burst (5 min/h)
- Full bundle, no operational sublimits
- Sublimits: embeddings 400 RPM · transcription 30 RPM · TTS 30 RPM
- 8K context default · up to 32K configurable
- Thinking mode: unlimited
- Email + chat · <8h business-hour response
- Status page with usage metrics
- Region: EU, LATAM or US
€1,200/mo
For high concurrency and long context.
- 500 sustained RPM · 700 burst (5 min/h)
- Full bundle without additional sublimits
- 32K context default · up to 128K configurable
- Thinking mode: unlimited
- Scheduling priority over Pro and Lite
- Custom event webhooks
- Priority chat · <4h business-hour response
- Region: EU, LATAM or US
from€5,000/mo
For integrated products with very high concurrency.
- 5,000+ RPM (negotiable per hardware)
- Full bundle + optional LoRA fine-tuning
- 128K context default · up to 256K native
- Dedicated hardware of your choice
- Thinking mode: unlimited, high priority
- Maximum scheduling priority
- Shared Slack · <1h business-hour response
- Onboarding assisted by senior engineer
- Region: EU, LATAM or US
from€15,000/mo
Dedicated server, custom configuration, RFP-ready.
- 100% dedicated server (no multi-tenant)
- RPM, context, SLA and models tailored
- Hardware chosen per workload
- Compliance: SOC 2 Type I, ISO 27001 (in progress)
- Dedicated support: founder + senior engineer
- Negotiated roadmap commitment
- Deployment: Tessera cloud, private cloud or on-premise
- Fine-tuning on your data under NDA (optional)
New plan
Just getting started? The Founder plan is your way in.
Tessera Founder: your private AI assistant on your WhatsApp and Telegram, hosted by us, with a founders community, monthly training and a weekly podcast. $55 a month, flat.
Commercial honesty
Who Tessera fits. And who it doesn’t.
We don’t compete on raw price against the cheap end of the market. If we fit, you save money and headaches; if we don’t fit, we tell you on the first call.
Tessera fits if
- You spend €2,000–5,000 a month on frontier today (GPT-6, Claude Opus 5.5, Gemini 4) and the monthly variance complicates your reporting.
- Your DPO or legal team asks where customer data physically lives, and the answer matters.
- You want to consolidate invoices instead of running three different cloud providers.
- You need human support in English or Spanish, in European, Latin American or US business hours, with bounded response times.
- Your team builds product, not ML infrastructure — you want a drop-in that works, not a -90% optimization on prompt caching.
Tessera doesn’t fit if
- You already use GPT-6 Luna, Claude Haiku 5.5 or Gemini Flash-Lite and it works for you — that’s your win, not Tessera’s.
- Your traffic is extremely bursty (0 to 10,000 RPM in seconds). Per-token serverless beats dedicated GPU.
- You have an in-house ML team that optimizes every prompt and squeezes hyperscaler volume discounts.
- You need specific closed models (GPT-4o image, Sora, Veo) that only run in their native cloud.
If your case sits in the right column, we tell you on the first conversation. We don’t push contracts that don’t fit.
Built for developers, governed for enterprise
The piece your team and your DPO sign on the same day.
Documentation a developer reads in fifteen minutes. Compliance a DPO validates on Monday.
Docs with real examples
Python, Node.js, Go and cURL snippets per endpoint. Editable cookbook, errors documented with cause and workaround.
Public status page
status.tesseraai.cloud. Per-region latency in real time. Postmortems published within five business days, before customers ask.
Usage webhooks for your billing
Consumption events on every request. Plug your own chargeback or cost-center system without going through the dashboard.
Exportable audit logs
Signed logs, exportable to your S3 or GCS bucket. Configurable retention for DORA, SOC 2 and AI Act audits.
Trying Tessera takes 5 minutes. Full migration, one day.
OpenAI v1-compatible API, open models on dedicated GPU in EU, LATAM or US. Flat monthly invoice, no token-meter. Founder-led support in English or Spanish.