Fable 5.1 is out. I put it to work overnight
Claude Fable 5.1 and Mythos 5.1 are out. I do not usually cover frontier models in any depth, so this is a first impression, not a deep dive: what changed, who gets which model, what it costs, and then the part I actually care about. I had a job for it.
The release, skimmed
Fable 5.1 is Fable 5 released to everyone: Pro, Max, Team and Enterprise tiers. Until now you needed the more expensive plans to run Fable 5, and there were plenty of complaints about requests downgrading to Opus 4.8. Anthropic says the new version has up to 60 percent fewer false positives, can discover vulnerabilities without exploiting them, and leans hard into biology and scientific research.
The terminology matters. Fable 5.1 and Mythos 5.1 are essentially the same model with different levels of safeguards. Fable 5.1 is generally available. Mythos 5.1 is for the cool kids in the Trusted Access program, and I have no idea how you get invited. It seems like you need to be hacked, like Hugging Face was, and then you are in.
The benchmarks in the announcement compare Fable 5.1 to Fable and Mythos, and they are all self-reported. The only one I would give real weight to is the Cursor bench. Take the rest with a grain of salt.
Two data points worth knowing. Enterprise Frontier Safeguards, which promise the same protections as a zero data retention policy, do not arrive until later this fall. Until then, zero data retention is still there and you should have it enabled, on Anthropic, on OpenAI, and on anything you route through OpenRouter. It is a flag. What happens after you set it is up in the air, and you are taking their word for it. Never trust data you care about to these companies. They are not in your best interest. They are in the business of making the models as good as possible.
The scientific research demo is genuinely cool: an elevation map of a third of Venus, generated from old mission data.
The pricing, read off the docs
The platform docs are easier to read than the press release, so I read the numbers from there, per million tokens:
- Base input: $0.10
- Cache writes: $12.50
- One-hour cache: $20
- Cache reads: $0.25
- Output: $50
Anthropic says cache reads now cost 75 percent less. OpenAI has been very good on token efficiency and cost per token, and Fable is heading in that direction. It still sits at the top of the leaderboards for price. No surprise, still insanely expensive.
The real test: my benchmark, overnight
I am not going to use Fable 5.1 much. I pay for a subscription that I mostly use for YouTube editing, so I pointed it at something I was stuck on.
harness-bench is the benchmark I have been building for local coding agents. It runs any harness, OpenCode, pi, DeepSeek Harness, against local models on my own tasks. The problem: the harnesses are too good. Four tasks, 47 of 48 runs resolved, and the one miss was a subjective 96 out of 100. I am not a benchmark or prompt person. I got it to this state with local AI and I was running in circles.
So I wrote a long prompt (priority one: make it harder, calibrate, keep it result oriented and universal because the output feeds YouTube videos), pasted it into Claude Code in plan mode, and watched. Plan mode alone ran 9 minutes and 13 seconds, 37 thousand thinking tokens, $7.24 on the usage estimator, and produced a phased plan with a 6 to 9 hour timeline. With a local model I would have had to split that into many chunks. I hit auto mode and went to bed.
Two days later
Lots of little hiccups. It ran for about two days instead of one night. Sessions died because the machine ran out of memory, or because a benchmark errored. Auto mode tried to fix things on its own, but it needed me to step in and kill tasks more than once. The estimator ended at $165.08. It is a subscription, so that is not money out of my pocket, but it is the number.
What came back is a clear step up from what I had built with local models. Suite version 0.2.4, over a hundred runs across three models and three harnesses, all on one page: mean score, pass rate, agent time, tokens generated, thinking tokens, and real diffs of what each run changed against the task. The harder tasks worked. The earlier turns (0.2.1 and 0.2.2) were still handing out passing scores, and by 0.2.4 the failures are back where a benchmark needs them.
The caveat I said on camera and will keep saying: I have AI calibrating AI. Most checkpoints are hard pass or fail, but a few are interpretive, and Qwen 3.8-27B is the one grading those. Ideally that is a human audit. With the number of runs I plan to do, it is not feasible to audit them all, so it stays labeled as what it is.
What I think
Decently impressed, and I will leave it there. With limited hand-holding, Fable 5.1 turned a benchmark I was stuck on into a harder, calibrated suite and automated running it overnight at scale. It is also the most expensive model on the board and it did not survive two days unattended without help.
Next: calibration runs on GLM 5.3 Flash, and digging into why pi, OpenCode and DeepSeek Harness make such different numbers of tool calls and turns on the same model. I am comparing apples to oranges across model sizes on purpose (Qwen Coder, Gemma, Qwen 3.8 Flash, the 3.6 models); this round was about testing the benchmark itself.