Passes day one. Sonnet 5 took one tangled project call transcript and split it into two separate, correctly formatted outputs from a single prompt, in under two minutes, without one bleeding into the other and without inventing anything to fill a gap.
- The wiki output stayed on-topic: correctly formatted markdown, right headings, no project status content bled in, didn't name the client
- The status report correctly attributed all six attendees by company, named who was driving the session manager decision, and flagged the blocked IP ranges still waiting on the client's network team
- Caught the CTO's hosting deadline bombshell dropped at the end of the call and gave it its own flagged section, not buried in a bullet list
- Didn't hallucinate any dates. None were given for the new deadline, and it didn't invent one
- Correctly read a one-off security aside as non-repeatable, kept it out of the wiki page, and kept it in the status report with the reasoning attached, not just a keyword match
- Whole run took under two minutes
- No risk table. The new risk was called out clearly in prose but not pulled into a structured table alongside the action table. Minor, but worth logging.
Verdict
Passes day one. Sonnet 5 took one tangled project call transcript and split it into two separate, correctly formatted outputs from a single prompt, in under two minutes, without one bleeding into the other and without inventing anything to fill a gap.
Editor's note: this is a single test on one task, run the day Sonnet 5 launched. I'm running it for a full week across real work, home and content tasks, in chat, Cowork and Claude Code, and against other models on the same briefs. This page gets updated with that result. Treat this as a first pass, not the last word.
Setup
Twenty years of running IT rollouts taught me the number that matters isn't a benchmark, it's whether a tool can do two jobs from one brief without mixing them up. So instead of chasing the SWE-bench score every other channel was posting within hours of launch, I built a transcript shaped like a real project call.
The transcript tangles two topics together: a technical walkthrough from an engineer, then a project status conversation with decisions and blockers. It's an anonymised composite based on a scenario I've actually been through, not a real client recording. I can't use client data for a test like this, so it has the shape and the mess of a real call and nothing more.
The task
One prompt, two required outputs, kept separate:
This tests two of Anthropic's claims for Sonnet 5 at once: that it's their most agentic model yet, meaning it should handle multi-step tasks without being walked through each one, and that it gives near-Opus output in a fraction of the time.
- A Confluence-ready wiki page built from the technical walkthrough section
- A project status report built from the status conversation, decisions and blockers
The input
A markdown file containing the full call transcript, attendees from both sides (Cliffinkent Consulting and a fictional fashion retail client), technical setup detail, project decisions, open blockers, and a late-call disclosure from the client's CTO that the deadline has moved up hard.
The output
Two documents, generated in one pass in under two minutes. The wiki article: a clean setup guide in markdown, correctly headed, no project detail in it, ready to paste into Confluence. The status report: attendees, decisions, blockers, a new risk section, and an action table at the bottom, each item pointed at the next date the transcript actually mentioned.
What worked
The separation held. No project status content showed up in the wiki page, no technical walkthrough detail showed up in the status report. Attendees were correctly attributed by company. The blocked IP ranges were logged against the right team. The CTO's late-call bombshell, that the current hosting provider is pushing the client out and the deadline has moved up, got its own flagged section rather than getting buried. No dates were invented to pad it out.
The sharpest bit: a one-off security-related aside in the transcript got read correctly as non-repeatable. It stayed out of the wiki page and stayed in the status report with the reasoning attached. That's closer to understanding what the call was actually about than to matching keywords.
What broke
No risk table. The new risk was called out clearly in the text but not pulled into a structured table the way the actions were. If you need risks in a specific table format every time, check the output before you paste it in.
Safety and trust check
No hallucinated dates, no invented client detail, no bleed between the two outputs. The model correctly declined to name a client it was never told the name of. On this single run, nothing was fabricated to fill a gap.
Who should use this
Anyone who takes messy call recordings and regularly needs to turn them into two distinct deliverables, a technical doc and a status update, in one pass. Consultants and PMs doing exactly this kind of split are the target case.
Who should avoid it, for now
Anyone who needs a guaranteed structured table for every category of output every time. And anyone weighing this for repeated, unsupervised use over a full working week: that's not tested yet. Check the output on anything that reaches a client before it ships.
What would change this verdict
A side-by-side against another model on the same transcript where Sonnet 5 comes off worse. Contamination between the two outputs on a longer or messier call. A week of real use across chat, Cowork and Claude Code turning up a slip this single test didn't catch.
Disclosure
No sponsorship, no early access. Tested on public release day using a self-built, anonymised transcript, not real client data. This is a first test, not a final word: a week-long follow-up against real work, home and content tasks, and against other models, is planned and will update this page.
Real task
Split one real-shaped project call transcript into two separate outputs from a single prompt: a Confluence-ready wiki page built from the technical walkthrough section, and a project status report built from the decisions, blockers and risks conversation.
Safe input or excerpt
A composite, anonymised transcript built to match a real project call: an engineer's technical walkthrough followed by a project status conversation covering attendees, decisions and open blockers, ending with a late disclosure that the client's hosting provider is pushing them out and the deadline has moved up. No real client name, no real dates, no real data.
Below is a transcript of a project call. It contains a technical walkthrough a colleague could reuse as a how-to, and a separate discussion about project status, blockers and decisions. From this single transcript, produce two separate outputs.
One, a Confluence-ready wiki article documenting the technical walkthrough as a repeatable process, with proper heading structure.
Two, a project status report covering current progress, blockers, key decisions and action items with owners.
Keep the two fully separate. The wiki article should contain only the reusable technical process. The status report should contain only project-specific status, decisions and actions. Do not duplicate content between the two.Output or excerpt
The wiki page opened with a clean markdown heading structure for the technical setup, no project status content, no client name. The status report opened with attendees listed by company (Cliffinkent Consulting; the client's team), then key decisions, blockers including the unconfirmed IP ranges with the client's network team, a flagged "new risk" section for the hosting deadline, and an action table at the end with items dated against the next relevant date mentioned in the call.
Uncertainty
One prompt, one transcript, one session, in Claude's normal chat interface. This doesn't cover a week of real work, home or content tasks, doesn't cover Cowork or Claude Code, and doesn't cover a side-by-side against Opus or ChatGPT on the same brief.
What would change the view
A side-by-side against another model on the same transcript where Sonnet 5 loses ground. Contamination between the two outputs on a longer or messier transcript. A week of real use turning up the kind of quiet slip a single test can't catch.
- No sponsorship, no early access. Tested on public release day using a self-built, anonymised composite transcript, not real client data. This is a single-session first test. A week-long follow-up, run across chat, Cowork and Claude Code and compared against other models, is planned, and this verdict will be updated once that's done.