Migrating from OpenAI

TL;DR

If your code uses the official OpenAI SDKs (Python, Node, Go) against the classic chat, embeddings, audio and TTS endpoints, the migration is three lines. The client stays the OpenAI client — only its target changes.

1. Swap the base_url

The official OpenAI SDKs expose the configuration parameter: base_url in Python and baseURL in Node. Pass it to the client constructor and the rest of your code is unaffected.

# Antes (OpenAI)
client = OpenAI(api_key="sk-…")

# Después (Tessera) — sólo cambia el destino
client = OpenAI(
    base_url="https://api.tesseraai.cloud/v1",
    api_key="sk-tessera-…",
)

response = client.chat.completions.create(
    model="Qwen/Qwen3.6-35B-A3B",
    messages=[{"role": "user", "content": "Hola"}],
)

- base_url="https://api.openai.com/v1"
+ base_url="https://api.tesseraai.cloud/v1"

2. Map your models

Tessera offers eight open models on dedicated GPUs. The table maps the most-used OpenAI models to their closest Tessera equivalent. It is not a 1:1 equivalence — different architectures, different training. We recommend validating against your real prompts before promoting the swap to production.

We don’t publish comparative benchmarks against OpenAI: any number outside your own dataset is marketing. The home calculator lets you simulate your cost on your real volumes.

3. Supported endpoints

Coverage targets the most-used routes of OpenAI’s v1 API. What we don’t expose yet is marked “roadmap” — if it blocks your migration, ping us and we prioritise it.

4. Documented behavioural differences

Same contract doesn’t mean same behaviour. These are the points where Tessera diverges from what you’d see against OpenAI, ordered by likelihood of affecting you.

temperature, top_p, max_tokens

Identical. max_tokens is interpreted as "output tokens" — input context does not count against it.

tool_choice and tools

OpenAI-compatible function calling. tool_choice accepts "auto", "none", "required" and { type: "function", function: { name } }. Parallel tool calls supported.

response_format

JSON mode (response_format: { type: "json_object" }) and JSON Schema (type: "json_schema") supported. Strict validation: the runtime retries internally before failing if the model leaves the schema.

stream

SSE identical to OpenAI: data: {…}, terminated by data: [DONE]. 15 s heartbeats to keep proxies alive.

seed

Accepted but does not guarantee absolute determinism across model versions. OpenAI doesn’t guarantee it either — the difference is we say so.

logprobs / top_logprobs

Not yet supported on /v1/chat/completions. On the roadmap. If your use case depends on logprobs, tell us.

n (multiple completions)

Accepted but always returns n=1. For multi-sampling, call N times — same cost, more predictable latency.

logit_bias

Not supported. Most uses fold into constrained generation; we recommend response_format=json_schema instead.

Vision (image inputs)

Supported on Qwen/Qwen3.8-27B and Qwen/Qwen3.8-27B-thinking: image_url parts in content (https URL or base64 data URI), same as OpenAI. Qwen/Qwen3.6-35B-A3B stays text-only.

Rate limits and errors

Same error envelope (error.type, error.code, error.message). 429 with Retry-After header. Detail in /docs/api/rate-limits.

5. Gradual rollout

Once it works locally, the prudent move is to send real traffic gradually and compare against the OpenAI flow before pulling the plug. Typical pattern, feature-flagged or percentage-based:

# Python — percentage routing, no changes to the product SDK
import random
from openai import OpenAI

openai_client = OpenAI(api_key=OPENAI_KEY)
tessera_client = OpenAI(
    base_url="https://api.tesseraai.cloud/v1",
    api_key=TESSERA_KEY,
)

def chat(messages, *, tessera_share=0.10):
    use_tessera = random.random() < tessera_share
    client = tessera_client if use_tessera else openai_client
    model = "Qwen/Qwen3.6-35B-A3B" if use_tessera else "gpt-4o"
    return client.chat.completions.create(model=model, messages=messages)

Start with tessera_share=0.05 on low-criticality traffic. Log both sides and diff cold for 1–2 days before turning the dial up.

What not to migrate yet

Some flows still depend on capabilities we don’t expose. If your product uses them, keep that subset on OpenAI — everything else can live on Tessera without issue.

• Realtime API (duplex voice over WebSocket).
• Batch API: use the chat model’s async queue if your workload tolerates async.
• Fine-tuning via API. Open-weights enables alternative paths — talk to us.
• Vision (image inputs): the model is text-only.
• Image generation, DALL·E, Sora.
• Assistants API and Agents SDK: OpenAI is winding them down, we recommend Chat Completions + tools.

Need help?

Migrations usually take less than an afternoon. If you get stuck — an endpoint you can’t find, output that doesn’t match, a prompt that behaves differently — write directly to the founder.