Use Canonn as an MCP evidence tool while Claude remains your primary coding model.
claude mcp add --transport http canonn \https://canonn.ai/mcpSetup →A 9B model trained to answer from the context you provide, instead of what it remembers. When the answer is not there, it says so.
Canonn R1 live The GPT-5.6 comparison is recorded for the preset example only. Your edited text went to Canonn alone.
91.8% from the document·$0.25 / 1M in·OpenAI-compatible API
Canonn R1 answers live, through the public demo endpoint of api.canonn.ai/v1, limited per visitor. GPT-5.6 was given the same document and question through OpenRouter on 23 Sept 2026, at temperature 0. The labels are recorded verdicts of an independent referee, TypeSafe Jev, not decided by this page (GPT-5.6: corrected the document from memory; Canonn R1 recorded 24 Sept 2026). 91.8%: FaithEval-counterfactual, the paper's protocol, 110 clean items, on production, 28 Sept 2026. $0.25 / $1 per 1M tokens in / out; OpenAI-compatible at api.canonn.ai/v1.
FaithEval-counterfactual: 110 passages that contradict what the models "know", none of them in our training data. The right answer is on the page. Measured with the paper's own prompt and scorer, through each model's live API.
One document that says something the models "know" is wrong. The box at the top of this page is the same model, live.
A 9B model against Claude Opus 5, GPT-5.6 and Gemini 3.6, measured the way the benchmark's authors measure it, through their live APIs: Canonn R1 on production on 28 Sept 2026, the rivals on 25 Sept 2026.
From the document. FaithEval-counterfactual (ICLR 2025), the paper's prompt and scorer, 110 clean items per model, live APIs: Canonn R1 on production 28 Sept 2026, rivals 25 Sept 2026. 95% interval margins of 3.8 to 9.3 points. 63% of the original counterfactual sample (190 of 300) was in our older training data; on the 110 clean items the result is the same.
Real help centers. 6 public help centers, 201 answerable questions and 67 the help center does not answer; Claude Sonnet 5 got the same passages. "Right" means the core of the answer is correct; the intervals overlap. Graded strictly, where any wrong added detail fails an answer: 85.6% against 94.5%. Details
Contradictions and injection. Self-contradicting documents go to a conflict expert in production (how); level with Claude Opus 5 and Gemini 3.6 within the 95% intervals. On our own held-out injection set it keeps to the business's data 98.8% of the time (GPT-5.6, 26 Sept 2026: 54.4%). The router's development set included FaithEval-shaped conflicts and was built with the same generators as these test items; it was held out by document, not by generator.
Same document, same question, six recorded questions. Per answer: GPT-5.6 16 to 30×, Gemini 3.6 27 to 100×, Claude Opus 5 297 to 799×. The answers were not all the same: the referee's verdict on the rival differed from its verdict on Canonn for GPT-5.6 on 6 of 6, Gemini 3.6 on 1 of 6, Claude Opus 5 on 6 of 6.
Per token, 3 to 20×. Per answer, more: on the same question the three write about 27× more tokens on average (median over the questions), and every token is billed.
Per answer: Canonn at list price against what OpenRouter billed each model for the same document and question, lowest to highest over the six questions · verdicts from the independent referee (TypeSafe Jev) · per token: OpenRouter list prices, 24 Sept 2026.
Under five minutes on why Canonn R1 answers from your documents instead of its memory. The line moves with the voices; click it to skip.
Answers are built from the documents you supply, including when they contradict what the model learned in training.
When the data does not contain the answer, Canonn says so instead of inventing one. 77.6% correct abstention on FaithEval-unanswerable, the paper's protocol, on production 28 Sept 2026; Gemini 3.6 Flash scores higher, 85.3%.
Grounded answers through the files extension carry passage-level citations, so every claim can be traced back to the line it came from.
Watch an answer form, and every fact in it trace back to a passage in your data.
If your stack speaks to OpenAI, it already speaks to Canonn. Change the base URL and the model name.
Same request shape, same streaming. Works with the OpenAI SDKs, LangChain and LlamaIndex.
Put the document in the prompt, or upload files once and let the API retrieve and cite the passages per request.
Prompts are never used for training. Logging is off by default. Inference runs in Canada.
One call scans a file and lists where it disagrees with itself, before your customers find it. Docs
curl https://api.canonn.ai/v1/audits \
--oauth2-bearer "$CANONN_API_KEY" \
-d '{"file_id": "file-..."}'
from openai import OpenAI client = OpenAI( base_url="https://api.canonn.ai/v1", api_key=CANONN_API_KEY, ) r = client.chat.completions.create( model="canonn-r1", messages=[ {"role": "system", "content": "Refund window: 45 days. In-app purchases: 5 days."}, {"role": "user", "content": "How long do I have?"}, ], ) print(r.choices[0].message.content) # 45 days from purchase, or 5 days for in-app purchases.
Claude Code and Codex keep doing the coding. Canonn joins only when a task depends on your documents, internal data, or a claim that needs a citation.
Use Canonn as an MCP evidence tool while Claude remains your primary coding model.
claude mcp add --transport http canonn \https://canonn.ai/mcpSetup →Let Codex call Canonn for source-backed answers, then continue the task with the evidence attached.
codex mcp add canonn --url \https://canonn.ai/mcpSetup →Point your OpenAI client at api.canonn.ai, or paste your own document into the playground and ask it something it should refuse.