Break an LLM call. Watch Iris Control catch it.
Take a model down and watch the fallback answer, test a new model release on a golden eval set with its price, and see what every call costs. PII redaction, prompt injection protection, and data governance are in the Security tab. No signup, nothing saved.
Model failover
Every answer here runs on the same primary and fallback models our production chatbot uses. You set the chain, and Iris Control only ever calls the models on it. Turn on the outage and the primary is treated as down, so you can watch the fallback take over with no error for the customer. Nothing is saved.
Answers come from the IrisAgent help center, the knowledge base of our demo account. The question is shared with Data governance in the Security tab.
Ask a question to see each model attempt, how long it took, and which model answered.
AutoEval
Which model should you upgrade to, and what will it cost you? AutoEval grades every output in your own LLM logs for correctness and against your own rules, and replays the same prompts on every model you consider, with its price. This tab uses real results from IrisAgent's own golden support set, the one behind the Support Model Arena, last run October 15, 2026. Pick the model you run today and a new release to see whether the upgrade ships.
| Suite | Today | New release | Change |
|---|---|---|---|
| Overall | 93.3 | 90.6 | -2.7 |
| Price per 1M tokens (in / out) | $5 / $30 | $2 / $10 | 67% cheaper |
| Answer accuracy | 98.8% | 95.3% | -3.5 |
| Hallucination resistance | 95% | 95% | 0 |
| Cited sources | 76.5% | 70.6% | -5.9 |
| Intent routing | 97.5% | 95% | -2.5 |
| Ticket tagging | 90.8% | 86.7% | -4.1 |
- Overall score is not higher: 90.6 against 93.3 today.
- Answer accuracy drops from 98.8% to 95.3%.
- Cited sources drops from 76.5% to 70.6%.
- Ticket tagging drops from 90.8% to 86.7%.
For context, not part of the rule: Typical response time is about 16% slower.
A new model ships only if it passes every check on the golden set. In Iris Control, each use case sets its own rule.
- The new model's overall score must be higher than the model you run today.
- Hallucination resistance must not drop at all.
- No other suite may drop by more than 3 points.
Each call's system prompt, user prompt, and output are graded for correctness, and against rules you write such as "always include a link". To compare models, the same prompts are replayed on each one and every output is graded the same way.
| # | Model | Overall | Answer accuracy | Hallucination resistance | Cited sources | Price per 1M (in / out) | Speed vs median |
|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 (reasoning)Anthropic | 96.6 | 100% | 100% | 88.2% | $4 / $20 | 1.3x |
| 2 | Claude Fable 5Anthropic | 94.4 | 100% | 95% | 87.1% | $10 / $50 | 1.2x |
| 3 | GLM 5.3 FlashZhipu | 94.4 | 97.6% | 100% | 82.4% | $0.1 / $0.5 | 1.1x |
| 4 | Claude Opus 5.5Anthropic | 94.1 | 97.6% | 95% | 87.1% | $4 / $20 | 1.2x |
| 5 | Grok 4.5xAI | 94.1 | 97.6% | 95% | 87.1% | $2 / $6 | 0.9x |
| 6 | GPT-6 AstraOpenAI | 93.9 | 97.6% | 100% | 76.5% | $10 / $50 | 0.6x |
| 7 | GPT-5.5OpenAI | 93.8 | 100% | 95% | 78.8% | $5 / $30 | 0.7x |
| 8 | GPT-5.6 SolOpenAI · Today | 93.3 | 98.8% | 95% | 76.5% | $5 / $30 | 0.6x |
| 9 | Gemini 3.1 Pro (thinking)Google | 93.3 | 96.5% | 95% | 82.4% | $2 / $12 | 1.6x |
| 10 | Claude Sonnet 5.5Anthropic | 93 | 92.9% | 100% | 82.4% | $2 / $10 | 0.6x |
Send 10 to 20 example calls (prompt and output) from any workflow, including your own in-house model, and we will return a comparison like this one on your prompts, with price. Nothing changes in your call path.
Cost
Every call on this page goes through one path, so each one is logged with its model and cost. In Iris Control the same view groups spend by model, workflow, user, and customer, and you can set a limit on each. Here, each demo tab stands in for a workflow.
Run a few calls in the other tabs, then come back to see what they cost.
Costs are estimated from the text each model read and wrote, priced at list rates per million tokens. System prompts and tool-loop overhead aren't counted, so real cost is somewhat higher. A primary skipped by a simulated outage is never called, so it isn't counted.
Security
The same layer that routes every call also protects it: personal data and secrets are redacted before a provider sees them, Prompt Shield inspects, isolates, and verifies untrusted content, and each call reads only the data it's allowed to. Nothing is saved.
Edit it or paste your own. Nothing is saved, and no model is called.
- Email addresses
- Phone numbers
- Card numbers
- Social Security numbers
- Street addresses and ZIP codes
- IP addresses
- Usernames, passwords, and login strings
Run the redactor to see exactly what a model provider would receive.
Prompt Shield handles prompt injection in three steps. Inspect scores every input (user queries, retrieved documents, tool results, attachments) for injection attempts. Wrap isolates untrusted content in delimited blocks behind a hardened system prompt, so the model reads it as data, never as instructions. Verify checks the output for system prompt leaks, unknown citations, and links outside your approved sources. How Prompt Shield works
Below, wrap and verify run live on our demo chatbot's answer path. Further down, check any answer against your own sources.
Your message goes to the same model and prompt our demo chatbot uses, with Prompt Shield's wrap and verify steps on.
- Wrap: the answer step receives your message in a tagged block, with forged tags escaped.
- Answer: the hardened prompt tells the model to treat that block as data, never as instructions.
- Verify: the answer is checked for copied prompt text and links outside the retrieved sources. A blocked answer is withheld.
Grounding check
Before an answer reaches a customer, Iris Control checks it against the sources you approved. Every link must point at an approved source, and a judge model checks each claim against the source text. The sources below are from a fictional help center, so edit anything. Nothing is saved.
Check the made-up answer first: it invents a timeline, a credit, and a link. Then load the grounded answer and compare.
Data scoping
Multi-product and multi-tenant teams need each answer to come only from the data it is allowed to use. Pick a product and compare what retrieval finds across the whole knowledge base with what it finds once the call is scoped. The answer uses only the scoped articles. Nothing is saved.
Answers come from the IrisAgent help center, the knowledge base of our demo account. The question is shared with Data governance in the Security tab.
Run a comparison to see which articles each call can use. Articles that only an unscoped call would see are marked out of scope.
Call log
Every call on this page goes through one path, so every call is logged. Kept in this browser tab only.Run any tab above and its calls show up here.
Want to see this on your own LLM traffic?
We'll put Iris Control in front of one of your use cases and walk through failover, scoping, and grounding with your team.