Augure vs Qubittron Bastion: two Canadian-hosted APIs

Augure and Qubittron Bastion both offer an OpenAI-compatible API from a Canadian company with inference in Canada. The difference is what sits behind the endpoint. Augure serves its own reasoning models (Ossington 5, Rosedale One, Tofino 3) and ships chat, desktop and MCP surfaces around them. Bastion, per its published documentation, serves open-weight models on a hosting provider's GPUs in Quebec, alongside embeddings, image generation, transcription and text-to-speech. Read this if you are a Canadian developer or privacy officer weighing the two, or you run Bastion today and want to know what a move changes.

At a glance

	Augure	Qubittron Bastion
Company and jurisdiction	Canadian company, no US parent or investors.	Qubittron Consulting Inc., Oakville, Ontario, per its privacy policy. An SAP Silver Partner.
Where customer data is stored	Canada.	"Company servers are located exclusively in Canada", per the privacy policy (last updated March 2024).
Where inference runs	Canadian infrastructure (Ontario) and vetted EU partners under zero-data-retention; never US providers.	"All inference runs in Canada on OVH AI Endpoints (BHS region)", which its FAQ places in Quebec.
Your data used for training	Never.	Not published as of October 2026.
Prompt logging	Prompts never logged.	Audit log records a SHA-256 prompt hash, token count and per-key trace. Prompt retention not published.
Canadian residency option	Storage always in Canada. Accounts can be set Canada-only: requests then stay in Canada or fail with 503 residency_unavailable, never served elsewhere. tofino-3 and the Augure Open models never leave Canada.	All routes are described as Canadian. No per-request residency metadata is documented.
API	OpenAI-compatible, https://api.augureai.ca/v1	OpenAI-compatible, https://api.qubittron.ai/v1, keys prefixed qbt_. Also /v1/responses (stateless only), embeddings, images, transcription and TTS.
Models	ossington-5 (256K), rosedale-1 (up to 1M), tofino-3, Augure Open (GLM-5.3, Qwen3.8-Flash-Next)	Open weights: GPT-OSS 120B and 20B, Llama 3.3 70B, Llama 3.1 8B, Qwen3 32B, Qwen3 Coder 30B, Mistral Small 3.2, Mistral Nemo, Mistral 7B, Qwen2.5-VL 72B, Qwen3Guard. Context 32K to 262K.
Tool calling and JSON	Supported; response_format json_schema on supported models.	Passed through to the upstream model unchanged; unsupported models may return 502.
Images and PDFs	Images on any model via auto; PDFs read on the gateway.	Images via Qwen2.5-VL 72B. PDF input not documented.
Certifications	SOC 2 and ISO 27001 in progress (readiness underway with Vanta, per https://augureai.ca/security). TLS 1.3 in transit, AES-256 at rest.	FAQ: "SOC 2 Type I is attested. SOC 2 Type II audit is underway with completion expected in Q3 2026." The developer site also displays SOC 2 Type II and ISO 27001 badges; no report or certificate is linked.
Languages	English and Quebec French natively.	Not published for Bastion. TTS lists French-Canadian voices.
Billing currency	CAD.	CAD. Credits at 1 credit = C$0.001. Free C$0 (10,000 credits/month), Pro C$49/month (500,000 credits), Enterprise custom.
Best fit	Regulated Canadian teams that want reasoning models, long context, bilingual work and residency controls.	Developers who want cheap open-weight inference in Canada, or SAP shops on Qubi AI Suite.

Where Qubittron Bastion wins

• Price and a free tier. Bastion publishes C$0.11 input and C$0.59 output per million tokens for GPT-OSS 120B, and C$0.13 and C$0.39 for Mistral Small 3.2. The free plan includes 10,000 credits a month with no card. Augure's cheapest model, tofino-3, is C$0.60 input and C$2.00 output, and developer keys need a payment method.
• More endpoint types. Bastion documents eight endpoints: models, chat and legacy completions, Responses, embeddings (BGE-M3 at 10 credits per million tokens), image generation, transcription and text-to-speech. Augure's published API covers chat completions, web search, link reading and memories.
• Open-weight model choice. Bastion lets you call a validated open model by name. Augure Open serves two open-weight models in preview; the rest of Augure's catalogue is its own.
• A self-hosted sister product. Qubittron's Qubi AI Suite runs "inside your boundary with zero telemetry leaving the customer environment", with SAP BTP and SaaS modes, and can point its model layer at your own GPUs or at Bastion.
• SAP depth. Qubittron is an SAP Silver Partner and positions Qubi as an augmentor for SAP Joule. Augure publishes no SAP integration.

Where Augure wins

• Reasoning models with long context. ossington-5 and rosedale-1 return their thinking as reasoning_content, and rosedale-1 accepts up to 1M tokens. Bastion's largest documented context is 262K on Qwen3 Coder; most of its models are 131K or less.
• Residency you can enforce and see. Every Augure response carries _augure.inference_region, and a Canada-only account fails with 503 residency_unavailable rather than being served elsewhere. Bastion's docs do not describe per-response region metadata.
• Documented data handling. Augure states that prompts are never logged, customer data is never used to train models, and EU inference partners work under zero-data-retention agreements. Bastion's published pages state no training or prompt-retention policy.
• PDFs, tools and structured output on the gateway. Augure reads PDFs and spreadsheets itself, with OCR for scans. Bastion passes tools and response_format through untouched, and warns that unsupported models return 502.
• Bilingual by design. Augure's models are built for Quebec French and English. Bastion publishes no language commitment for its LLMs.

Switch from Qubittron Bastion to Augure

1. Create a key at https://augureai.ca/developers, add a payment method, and export the key as AUGURE_API_KEY.
2. Change the base URL and key. Bastion already uses the OpenAI SDK, so the client constructor is the only code change.
3. Map model ids. OpenAI alias ids such as gpt-4o-mini are also accepted and served by the closest Augure tier.

Bastion model	Augure model	Why
gpt-oss-120b	ossington-5	General reasoning and code, 256K context, thinking can be turned off.
gpt-oss-20b, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3	tofino-3	Fast, no thinking, served in Canada with no failover outside Canada.
Meta-Llama-3_3-70B-Instruct, Qwen3-32B	ossington-5	Same class of general model with a larger context.
Mistral-Small-3.2-24B-Instruct-2506, Mistral-Nemo-Instruct-2407	tofino-3	Everyday drafting, extraction and classification.
Qwen3-Coder-30B-A3B-Instruct	ossington-5	Code with reasoning_effort control.
Qwen2.5-VL-72B-Instruct	auto	Images route to Augure's vision engine in Canada.
bge-m3 and other embeddings, images, transcription, TTS	none	Not in Augure's published API. Keep them where they are.

4. Before and after (Python, OpenAI SDK):

# Before: Qubittron Bastion
import os
from openai import OpenAI
client = OpenAI(base_url="https://api.qubittron.ai/v1", api_key=os.environ["QUBITTRON_API_KEY"])
resp = client.chat.completions.create(
    model="gpt-oss-120b",
    messages=[{"role": "user", "content": "Summarise this contract in two lines."}],
)
print(resp.choices[0].message.content)

# After: Augure
from openai import OpenAI
client = OpenAI(base_url="https://api.augureai.ca/v1", api_key=os.environ["AUGURE_API_KEY"])
resp = client.chat.completions.create(
    model="ossington-5",
    messages=[{"role": "user", "content": "Summarise this contract in two lines."}],
)
print(resp.choices[0].message.content)
print(resp.model_extra["_augure"]["inference_region"])

5. Re-check: omit max_tokens on reasoning models (cap 32,768; tofino-3 16,384); send reasoning_effort: "none" on ossington-5 for mechanical tasks; response_format is not combined with tools on reasoning models; streaming errors arrive inside the stream; the default limit is 60 requests per minute with a C$25 daily spend cap until the first charge.
6. Run both for a week on your own prompts, watch _augure.inference_region and fallback, then cut over.

Keep Qubittron Bastion if

• You need embeddings, image generation, transcription or text-to-speech from the same key.
• Your workload is validated on a specific open-weight model and per-token price is the deciding factor.
• You run Qubi AI Suite self-hosted or on SAP BTP and want its model layer on a Canadian endpoint, or you only need a free tier to prototype.

Sources

• https://developer.qubittron.ai/ (checked 2026-10-27): Bastion positioning, base URL, "OVH Beauharnois" and "BHS · QUÉBEC", compliance badges, plan prices.
• https://developer.qubittron.ai/models (checked 2026-10-27): model ids, context windows, per-model prices.
• https://developer.qubittron.ai/pricing (checked 2026-10-27): credit definition, Free, Pro and Enterprise plans, credits per model, embeddings and included models.
• https://docs.qubittron.ai/llms-full.txt (checked 2026-10-27): endpoint surface, qbt_ keys, supported model list, tools and response_format passthrough, store: false, rate limits, error codes, OpenAI migration mapping.
• https://www.qubittron.ai/faq (checked 2026-10-27): SOC 2 Type I attested and Type II in audit, deployment modes, where inference happens, no public price list for Qubi, Bastion vs Qubi.
• https://www.qubittron.ai/developer-portal (checked 2026-10-27): "OpenAI-compatible LLM API hosted in Canada", OVH AI Endpoints BHS, PIPEDA, Law 25 and ITSG-33 alignment, SHA-256 prompt hashing.
• https://www.qubittron.ai/company/about (checked 2026-10-27): SAP Silver Partner, self-hosted platform, SOC 2 status.
• https://www.qubittron.ai/product/security (checked 2026-10-27): "SOC 2 Ready", self-hosted, zero telemetry, RBAC, SSO.
• https://www.qubittron.ai/legal/privacy (checked 2026-10-27): legal name and Oakville address, servers in Canada, last updated date.
• https://www.qubittron.ai/ (checked 2026-10-27): Qubi AI Suite overview, model providers, SAP positioning.