After the US Government suspended access (globally) to Anthropic Fable 5 last month, just after release, these blocks have now been lifted, and Claude Fable 5 is now back (optional) to use in preview inside Microsoft Cowork. At the sametime, Anthropic Sonnet 5 (which is an upgrade to 4.6) is now live too.
These appear in the model selectors inside Cowork and are also available across Microsoft Copilot Studio and Foundry.
Claude Sonnet 5
Sonnet 5 is more agentic, safer, and more capable at multi‑step work than the previous 4.6 version, with tunable effort levels and better computer‑use reliability. Anthropic say it brings:
- Stronger agentic execution – completes multi‑step tasks end‑to‑end with better planning and self‑checking.
- Improved computer‑use skills – OSWorld score rises to 81.2% (from 78.5%).
- Selectable effort levels – low/medium/high/max/x‑high reasoning depth to balance cost vs accuracy.
- Closer to Opus‑grade reasoning – lands close to Opus 4.8 on several benchmarks.
- Safer and more robust – fewer hallucinations, better refusal behaviour, stronger prompt‑injection resistance.
- Better follow‑through – more likely to finish long tasks without nudging.
Claude Fable 5
This is Anthropics’ “Mythos‑class” model built for long‑horizon autonomy, massive context, and advanced coding/knowledge‑work performance. I have covered this in another blog, but in summary it brings:
- Frontier‑level autonomy – can run for days, plan stages, coordinate sub‑agents, and self‑test work. (watch those Copilot Credit costs!)
- Massive context window – Massive 1M‑token context + 128K output enables deep research across large document sets.
- Best‑in‑class coding – said to outperforms flagship coding models (GPT‑5.5, Opus 4.8, Gemini 3.1 Pro) on complex coding and agentic tasks.
- Frontier vision capabilities – understands diagrams, tables, PDFs, and uses vision to validate code outputs.
- Stronger reasoning – 1.02×–1.4× better on standard reasoning; 2×–5× better on long‑horizon agentic/security tasks.
- Enterprise‑grade workflows – handles multi‑stage research, analysis, and deliverables with minimal oversight.
Copilot Credits
Remember, Cowork is now PAYG or Copilot Credit pre-purchase paid, so watch those credit costs as you experiment with the new models.
use /cost to get a credit cost for your running tasks in Cowork.
We are still waiting for the anticipated Cowork-1 Model from Microsoft to be released in Cowork, offering high-quality, lower cost and a Microsoft First model.
Leave a Reply