Blog: Product
Claude Sonnet 5.5 is now supported on Open Voice Shield
By the TelecomsXChange engineering team ยท
Anthropic released Claude Sonnet 5.5 today. It is selectable on Open Voice Shield today: as the analysis model of any AI tier, or to re-analyze a past call and compare verdicts. It costs the same as Sonnet 5, customers see the same verdict shape, and every new Anthropic and OpenAI model shows up in the pickers on its own.
What it is
Anthropic's newest Sonnet, at Sonnet 5's price
Claude Sonnet 5.5 is the mid-tier model of Anthropic's Claude 5 generation, released on 28 September 2026. Anthropic lists it at $2 per million input tokens and $10 per million output tokens, the same as Claude Sonnet 5. Claude Opus 5.5, at $4 and $20, arrived the same way and is selectable too. What an analysis costs is measured below: a model's own reasoning tokens are billed as output, so the token prices alone understate it.
The verdict does not change shape with the model: a probability from 0 to 100, a category, the entity the caller claimed to be, numbered red flags quoting the call, a TCPA note, and block, review or allow. A model only changes how well those are filled in.
Where to switch it on
Three places, no restart
- Platform default. Admin, AI settings, Anthropic, Claude Sonnet 5.5, save. Applies to the next call processed.
- Per tier. Under AI modes, set Sonnet 5.5 as the analysis model of Economy, Power or Max only. The tier economics panel shows our cost beside the fee before you save.
- One call, to compare. Open any past call and re-analyze it with Sonnet 5.5. The new verdict is stored beside the previous one, with the model, tokens, cost and latency of each, which is the quickest way to see what a model change does on your own traffic.
Measured
On our labelled set
We ran both new models through the verdict harness that ships with the platform, over our labelled set of 17 transcripts: scam scripts that must be blocked, legitimate calls that must not be, and thin calls with no basis for a verdict either way.
| Model | Expected verdict | Median answer | Cost per analysis |
|---|---|---|---|
| Claude Sonnet 5.5 | 16 of 17 | 5.5 s | $0.017 |
| Claude Opus 5.5 | 17 of 17 | 9.4 s | $0.038 |
Sonnet 5.5's one miss was the grandparent-emergency script, which it scored review at 50 rather than block. The request for cash by courier and the demand for secrecy are in the transcript, and a reviewer would catch it, but the model hedged. Two of its seventeen answers took over two minutes to come back, which looks like release-day load on the API rather than the model. At $2 and $10 per million tokens the token prices suggest a fraction of a cent, but both models spend reasoning tokens that are billed as output, so an analysis lands at about 1.7 cents on Sonnet 5.5 and 3.8 cents on Opus 5.5. The tier economics panel shows this figure beside the tier's fee before you save a change.
The set is small and written by us, so it ranks models and is not an accuracy claim. The white paper says what was measured and how, and the harness runs on your own calls with one command.
Why it was ready today
The pickers read the model list from the API
Until this week the analysis models were a table in our code. Now Open Voice Shield asks the Anthropic API for its model list, and the OpenAI API for its chat models, and merges what comes back into the admin pickers, newest first. A release like Sonnet 5.5 appears the day it ships with no update on our side. OpenAI's image, audio, realtime, transcription, codex and pro variants are filtered out, because none of them can score a transcript. Published list prices are on file for the current families; a model without one is still selectable, and its card says calls record no cost until the operator sets ours.
For a customer nothing changes: the tier you chose keeps its fee, which covers transcription and analysis whatever the model behind it. For an operator on an on-premises licence the whole menu is there under your own keys, with open-weight models through OpenRouter beside the frontier labs.
Questions
Frequently asked questions
Does Open Voice Shield support Claude Sonnet 5.5?
Yes, since its release on 28 September 2026. It is selectable in the admin AI settings under Anthropic, can be set as the analysis model of any AI tier, and can be used to re-analyze any past call.
How do I switch call analysis to Claude Sonnet 5.5?
Admin, AI settings, Anthropic, choose Claude Sonnet 5.5, save. It applies to the next call processed. To move only one tier, edit that tier under AI modes. To try it first, open any call and re-analyze it with Sonnet 5.5. The new verdict is stored beside the old one.
What does Claude Sonnet 5.5 cost per call?
Anthropic lists it at $2 per million input tokens and $10 per million output tokens, the same as Claude Sonnet 5. On our labelled set one analysis cost about 1.7 cents, more than the token prices alone suggest because the model's own reasoning tokens are billed as output. The customer's per-call fee is unchanged. The fee covers transcription and analysis whatever model the tier runs, so the operator sees the cost beside the fee before choosing.
Is Claude Sonnet 5.5 more accurate at detecting scam calls?
On our labelled set of 17 scam and legitimate transcripts, Sonnet 5.5 returned the expected verdict on 16, scoring one grandparent-emergency script review rather than block; Opus 5.5 returned all 17. The set is small and written by us, so it ranks models and is not an accuracy claim. Any operator can run the same harness on their own calls, and re-analysis lets you compare two models on one real call.
Will new Anthropic and OpenAI models appear automatically?
Yes. Open Voice Shield reads the model list from the Anthropic API and the chat models from the OpenAI API, so a new release appears in the pickers the day it ships. A model without a published price on file is still selectable and its card says so until the operator sets the cost.