Claude Fable 5.1 and Mythos 5.1 Land at the Top of the Chart, Then the Argument Starts
Quick verdict
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, positioned as its flagship pair for coding and knowledge work. Fable 5.1 tops Artificial Analysis's Intelligence Index at 66, cuts the cache-read price by 75 percent, and posts huge gains on agentic coding benchmarks. The catch is that it costs about 20 percent more per task than Fable 5, and community researchers say Fable and Mythos may be the exact same weights with different safety routing. If you buy AI by the task, this is the strongest Claude in a while, but the win is in cached agent workloads, not on raw intelligence-per-dollar.
What actually shipped
Anthropic announced the two models directly, calling them "the world's most advanced models for coding and knowledge work." Fable 5.1 is pitched at long-horizon, delegated work that runs on its own, while Mythos 5.1 is the paired release for knowledge work. Third-party platforms moved fast: Perplexity added Fable 5.1 to Computer for Pro and Max users, Nous Portal and OpenRouter picked it up, and T3 Code shipped support the same day.
The headline numbers came from Artificial Analysis:
- Context window of 1 million tokens, with text and image inputs
- Pricing unchanged from Fable 5 at $10 / $50 / $12.50 per million tokens for input, output, and cache write
- Cache reads cut from $1.00 to $0.25 per million tokens, a 75 percent reduction
- Intelligence Index of 66 at max effort, ahead of Opus 5 at 63, Fable 5 at 62, and GPT-5.6 Sol at 61
- Humanity's Last Exam at 59.1 percent, up from Fable 5's 55.5 percent
- Terminal-Bench v2.1 at 91.4 percent and SciCode at 62.0 percent
The coding jumps were the real story. Steven Dillmann pulled Terminal-Bench-Science numbers showing Fable 5 at 24.7 percent and Fable 5.1 at 52.6 percent, more than double. On the GDPval-AA v2 track, Artificial Analysis logged 1853 Elo, up 130 points over Fable 5.
| Benchmark | Fable 5 | Fable 5.1 |
|---|---|---|
| Intelligence Index (max) | 62 | 66 |
| Humanity's Last Exam | 55.5% | 59.1% |
| Terminal-Bench-Science 0.1 | 24.7% | 52.6% |
| Cost per task (max) | lower | $3.76 |
The cost math is not as clean as the benchmarks
The intelligence chart says Fable 5.1 is back on the frontier. The cost chart says slow down. Artificial Analysis put Fable 5.1 max at $3.76 per task, roughly 20 percent more than Fable 5, because the new model burns about 1.7 times the output tokens. The cache-read cut saves around $1.40 per task, which softens the hit but does not erase it.
Run the comparison against the cheaper frontier models and the picture gets sharper. One breakdown from nicdunz put Fable 5.1 max at 66 intelligence and $3.69 per task, against GPT-5.6 Sol max at 61 intelligence and $0.95 per task. On intelligence-per-dollar, Sol still wins by a wide margin. Fable 5.1 owns the absolute ceiling, not the value line.
Where it does pull ahead is inside agent stacks that re-read the same big prompt over and over. Theo reported that on CursorBench, the cache cut dropped costs by almost 50 percent while the model scored higher. That is the workload Anthropic optimized for, and it is why the cache-read price is the number to watch, not the sticker price.
Are Fable and Mythos even different models?
This is the part that turned a launch into a debate. Elie Bakouch argued that Fable and Mythos 5.1 are "the EXACT same weights," with internal activations used for safety classification, escalation to a bigger classifier, and fallback to Opus 4.8 for requests flagged as dangerous. If that holds, the difference between the two names is mostly the threshold on the safeguard classifier, not two separate base models.
Artificial Analysis's own note lines up with this reading. Its evaluation used Anthropic's default server-side fallback, where safety-flagged requests routed to Opus 4.8 or Opus 5, and that fallback served about 4 percent of output tokens across the Intelligence Index run. So some benchmark lines labeled "Fable" or "Mythos" may really reflect which safety path was active, not which model did the work. Bakouch repeatedly asked Anthropic to clarify when system-card rows use each name, because that changes how you read the whole table.
Why it matters
For anyone deploying Claude in production, the cache-read cut is the big change. It makes long-running agents a lot cheaper, which is the deployment economics finally catching up to the capability. The new Enterprise Frontier Safeguards, pitched as "ZDR++" for agent observability, plus zero-data-retention support, take a real blocker off the table for enterprise buyers. The routing story is the thing to stay careful about: treat "Fable" and "Mythos" benchmark numbers with some suspicion until Anthropic spells out which path produced them.
The louder signal is the split between the metrics and the experience. On aggregate benchmarks, 5.1 looks like a real jump. On day-one access, plenty of users hit walls, which is the older, more familiar Claude problem.
The rate limits and safeguards are the real complaint
The sharpest criticism was not about intelligence. It was about getting to use the thing. kimmonismus called the rollout "literally unusable" over rate limits and said 5.1 was "even worse than Fable 5" on usage caps, with no subscription benefit from the improved efficiency. Greg Kamradt reported requests rejected as "reverse engineering" during testing, which blocked a planned evaluation. Kyle Russell said a "military campaign" metaphor in a math session tripped cyber safeguards.
It was not universal. Theo pushed back hard, saying he was not seeing the rate-limit problem and had used only 14 percent of one weekly Fable limit. So even the complaint had no consensus. Some users hit a wall in an hour, others ran the model for days. If you have been tracking how Anthropic handles usage caps, this fits the pattern from the recent limits reset and the Max plan weekly limit changes.
Video: Claude Fable 5.1 first impressions
Here is a walkthrough of the release and what changed since Fable 5.
FAQ
Is Claude Fable 5.1 better than Opus 5?
On Artificial Analysis's Intelligence Index, yes, 66 to 63. On agentic knowledge work, they are close to tied on some measures, and Opus 5 runs cheaper per task. It depends on whether you need the absolute ceiling or better cost efficiency.
What is the difference between Fable 5.1 and Mythos 5.1?
Anthropic frames Fable as the coding and long-horizon model and Mythos as the knowledge-work model. Community analysis suggests they may share the same base weights, with the difference coming from safety routing and classifier thresholds rather than separate training.
Should I switch my coding agent to Fable 5.1?
If your workload re-reads a large cached prompt, the 75 percent cache-read cut makes it cheaper to run than Fable 5. If you are choosing between models on price, compare against Sol and route by task. See our guide on cutting coding agent costs with model routing.
Sources
- @claudeai - official launch of Fable 5.1 and Mythos 5.1
- @ArtificialAnlys - benchmarks, pricing, and cost-per-task breakdown
- @mikeyk - pricing and the 75% cache-read cut
- @StevenDillmann - Terminal-Bench-Science numbers
- @eliebakouch - "Fable and Mythos 5.1 are the EXACT same weights"
- @kimmonismus - rate-limit complaints
- @theo - cache cut and CursorBench cost drop
Further reading
Try all the models mentioned in this article
Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Free to start.
Start free on Admix