Claude Fable 5 just dropped - Anthropic's first Mythos-class model for everyone
Anthropic released Claude Fable 5 today, and it's a bit different from the usual model bump.
It's not a bigger Opus - it's a new tier above Opus, built on the same underlying model as Mythos, the one Anthropic had locked behind a government cybersecurity program since April.
Here‘s its own version of what it does:
Agentic work: plans multi-step tasks, calls tools, checks its own output, and keeps going without hand-holding.
Demos include autonomously playing Factorio and building a working CAD editor (with its own AI copilot inside).
Coding: 80.3% on SWE-Bench Pro vs 69.2% for Opus 4.8.
Knowledge work: top scores on GDPval-AA, ahead of Opus, GPT-5.5 and Gemini 3.1 Pro.
Long context + vision: noticeably stronger on long documents and multi-turn work.
The interesting twist: it ships with hard guardrails. Sensitive cyber/bio/chem requests get rerouted to Opus 4.8 instead - Anthropic says this fires in under 5% of sessions. The unrestricted version (Mythos 5) stays limited to vetted security orgs.
Availability: API now ($10/$50 per M tokens), and included free on Pro/Max/Team/Enterprise subscriptions until June 22 - after that it needs usage credits until capacity catches up
Has anyone tried it out? I‘m really curious to hear about your experiences and what you did with it.
Replies
Three early observations on Fable 5.
It shipped in-app mid-session while I was already working, so the change was instantly noticeable.
this one really burns through tokens. -Anthropic flags this also, so users should not be surprised. I run a Max plan and usually that is sufficient for all my projects. I checked my session usage was at 18% before. After shifting to Fable 5 and letting it plan and build the app extension I was working on, it delivered 9 files in roughly the same time Opus used to take for 2. My session usage jumped to 36% and after another revision to 56%. Stopped after that.
The speed of highly complex tasks is… crazy. Full stop.
When asking about its capabilities I triggered the fail safe explained above twice and got re-routed to Opus 4.8. Opus denied Fable existed :)) Those weren‘t even borderline legitimate questions, but it shows that the mechanism behind is really tight in terms of security.
It remains to be seen what the unrestricted public version will be like. A jump in AI capability is likely after what this first test showed.
working on my first few tests with Fable on both MetaScope and Armature. Will know more in the morning, starting a long running loop now with quite a large refactor on Armature. Implementing a medium sized RFC on MetaScope. So far so good, but no clear picture yet. To soon to tell
@geert_zaal Please update when you have a clearer picture! would be valuable to know.
first impressions are in on working in a mature Swift/Appkit environment (MetaScope) after using Fable 5 on a long coding arc: it raised the floor a lot. Strong coherence over many hours, empirical probes instead of doc-trust, artifact-led debugging, and much less flailing. But the ceiling was still human smoke testing. Green tests missed real UI behavior. The winning loop was model builds + measures, reviewer challenges, human tests the product. I used Fable on effort level high, will need more testing with higher thinking levels. But already burning hard through token budgets.
I am yet to experiment on this, Its free until June 22, then needs usage credits until capacity catches up, that phrasing is doing real work. Anthropic is constrained on this tier right now. The free period isn't generosity, it's a controlled ramp while they build out infrastructure. A new model tier that hits capacity at launch means the per-token compute cost is materially higher than Opus. the $10/$50 pricing reflects that.
@sanreds Agreed. Will be interesting to see whether pricing will adjust over time or stay so elevated due to the "marketing effects" of the whole Project Glasswing prelude.
I'm afraid to try it! Opus 4.8 already uses so many tokens. :-) I've even dropped back down to Opus 4.6 for a lot of my work. I'll try it eventually. It's interesting that this is just one week before they start a separate usage tracking for automated usage.
@debbie_shapiro Indeed. I just hit 11% weekly limit and 100% session limit with Opus 4.8 (medium) when I asked it for tasks suitable for Fable in my project. in under 10 minutes. Never had a token burn like this before. If it was Fable, fine, but it was Opus.
@debbie_shapiro @niklas_turunen I had the same experience with Token burn. Did your Opus 4.8 answer on Fable capabilities? My Opus downright denied Fable's existence the first time i triggered the safety guardrail with my question about its capabilities
@debbie_shapiro @michael_zorez I actually first asked Sonnet and it told me to use Opus for everything. But yea once I asked Opus about Fable the answer itself was rather good and it pointed to actual tasks worth doing with Fable. But pretty shocking that the first question with Sonnet took 10% and Opus took 90%. I can usually work for around 3 hours using mostly Opus.
I've been working on it since last night on a large application project I was developing with Opus 4.8, and Fable 5 optimized it in 12 hours, doing what Opus either missed or only minimally addressed.
However, it's extremely token-intensive. Despite having a Max plan, my daily quota was used up in 4 hours... I had to add an extra $50 to finish the task!
Thanks Michael, I will check it out. Used Opus 4.8 for a short while and didn't really notice the difference in quality of output from Sonnet.
Users said that the cost in "credits" was roughtly double compared to Sonnet and with no noticable gain I stuck with Sonnet 4.6 for the most part.
The write up for Fable promises great things, so I'll definitely take a look and see whether the output matches the hype