Claude Fable 5.1: What's New in Anthropic's Latest Model

Anthropic released Claude Fable 5.1, a refinement of its Fable 5 model that improves agentic performance and readability while cutting cached prompt read costs by about 75%, resulting in roughly 25% cheaper typical workloads and up to 45% cheaper agentic workloads. Independent benchmarking from Artificial Analysis suggests cost-per-completed-task can be higher at max settings due to increased output tokens. The model also saw cybersecurity-related refusals drop about 60% and biology/medical fallback refusals drop around 85%.

Claude Fable 5.1: What's New in Anthropic's Latest Model Anthropic's Fable 5.1 brings agentic benchmark gains, cheaper cached prompts, and better readability. Here's what actually changed. What is Claude Fable 5.1? Fable 5.1 is Anthropic’s newest frontier model, released alongside a more permissive sibling called Mythos 5.1 that’s currently limited to trusted partners. Fable 5.1 is not a full retrain so much as a refinement pass: it keeps roughly the same architecture and per-token pricing as Fable 5 but squeezes out meaningfully better results per dollar, mostly by needing fewer tokens to finish the same task and by cutting the price of cached prompt reads. The headline gains show up in long-horizon agentic work, coding benchmarks, and how readable its output is when it explains what it’s doing. TL;DR Token efficiency , not a straight price cut, is the real story. Anthropic reduced the cost of cached prompt reads reportedly by around 75% and the model uses fewer tokens overall, which nets out to roughly 25% cheaper typical workloads and up to 45% cheaper agentic workloads. Low reasoning effort now punches above its weight : on benchmarks like Terminal Bench and Cursor Bench, Fable 5.1 at low or medium effort matches or beats Fable 5 at high or max effort, at a fraction of the cost. Mythos 5.1 and Fable 5.1 are reportedly the same underlying model with different safety guardrails applied, and the version with fewer restrictions consistently scores higher on the same benchmarks, illustrating how much guardrails cost in raw performance. Readability took a real jump . Early testers say Fable 5.1’s reasoning and research output is much easier to scan than Fable 5’s, which had a reputation for being dense and hard to follow. Refusal rates dropped : Anthropic says cybersecurity-related refusals fell about 60% and biology/medical fallback refusals fell around 85%, though refusals haven’t disappeared entirely. Independent benchmarking from Artificial Analysis complicates the “cheaper” claim : because Fable 5.1 reportedly uses more output tokens than Fable 5 to reach its intelligence level, cost-per-completed-task can actually be higher for max settings, even with cheaper cache pricing. It shipped the same day OpenAI was expected to release a major new model , which several observers read as a deliberate scheduling move by Anthropic. Remy doesn't build the plumbing. It inherits it. Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something. Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want. What actually changed from Fable 5 to Fable 5.1? The core upgrade is agentic capability over long-running tasks. Anthropic and independent reviewers point to Terminal Bench a benchmark for coding and system tasks and Terminal Bench Science a benchmark for real scientific work like data analysis, model fitting, and proving theorems as places where Fable 5.1 shows a clear step up over Fable 5, at lower cost for a given score. On Cursor Bench, Fable 5.1 at max reasoning effort scored higher than Fable 5 at max effort while costing meaningfully less per task. At lower reasoning settings, the gap is even more favorable: multiple testers found Fable 5.1 on medium or low effort produced results comparable to Fable 5 on high or extra-high effort, at roughly a third of the price. The model is also being described as much faster in practice, particularly for building multi-step, agentic outputs like small games or generated apps, though “faster” here is more about needing fewer tokens and fewer retries than a change in raw inference speed. Why is Fable 5.1 cheaper if the token prices didn’t change? This is the part that trips people up. Anthropic did not lower the per-token price for input or output tokens. What changed is the price of cache reads, tokens the model has already processed and stored so it doesn’t have to reprocess them from scratch. Anthropic cut cache read pricing substantially reported around 75% cheaper, down to about 25 cents per million tokens , which matters enormously for agentic workflows that repeatedly reuse similar prompt templates, system instructions, or long context windows. Combine that with the model needing fewer total tokens to complete equivalent work, and Anthropic’s estimate is about 25% lower cost for typical workloads and up to 45% lower cost for heavily agentic ones. But this savings depends on your workflow actually preserving the cache, meaning you need to structure prompts so the cached portion stays consistent between calls. Break the cache pattern and you lose the discount. Independent analysis from Artificial Analysis pushed back on the “cheaper” framing at the high end: they found Fable 5.1 uses around 1.7 times the output tokens of Fable 5 to reach its top intelligence score, which means at maximum settings the cost per completed task can actually come out higher than Fable 5, not lower, despite the cheaper cache pricing. So the savings are real but uneven: they show up clearly in agentic and cache-heavy workflows, less so if you’re running everything at max reasoning effort on one-off tasks. How does Mythos 5.1 differ from Fable 5.1? According to reporting on the release, Mythos 5.1 and Fable 5.1 are effectively the same underlying model, differentiated mainly by the safety guardrails applied on top. Mythos is Anthropic’s more permissive, partner-only variant. On identical benchmarks and at the same reasoning effort or cost per token, Mythos 5.1 scores consistently higher than Fable 5.1, in some cases by several percentage points at max reasoning effort. Plans first. Then code. Remy writes the spec, manages the build, and ships the app. The pattern lines up with a broader observation about guardrails and model performance: restricting a model from certain outputs appears to cost some amount of its effective intelligence, because part of its capacity goes toward deciding when to refuse rather than toward solving the task. This isn’t unique to Anthropic, but Fable 5.1 and Mythos 5.1 being released side by side, with the same architecture and different safety layers, makes the tradeoff unusually visible. Is Fable 5.1 actually good at agentic, long-horizon work? Early hands-on testing suggests yes, and this is where the model draws the most enthusiasm. Testers describe building fairly ambitious projects, fully voiced isometric games with loot systems, enemy AI, and an automatic play mode, plus a simulation modeled on the Stanford “generative agents” LM Village experiment, where multiple AI-driven characters with individual memory streams interact and gossip, letting researchers watch how information spreads through a small society. One tester reported generating a substantial game, including all voice lines and sound effects, in roughly 10 to 15 minutes. Beyond novelty projects, the model’s gains on Terminal Bench, Terminal Bench Science, and Cursor Bench point to concrete improvement in tasks that require many sequential steps and sustained judgment, rather than a single clever answer. Anthropic also reports lower rates of “reward hacking,” where a model optimizes for a stated goal by cheating or circumventing rules, which matters more as models are given longer leashes to work autonomously. On raw output quality for things like generated images or 3D visualizations, comparisons with OpenAI’s and other competitors’ latest models were mixed. Some testers rated Fable 5.1’s image generation as solid but not clearly ahead of rivals, and building more complex outputs like a real-time 3D visualization sometimes required manual fixes rather than working perfectly on the first try. Is Fable 5.1 worth switching to? For anyone running agentic workflows through the API, Claude Code, or tools like Cursor and Factory, the case is strong: lower reasoning-effort settings now get you results that used to require expensive, high-effort runs, and the cache pricing changes reward workflows that reuse context. Anthropic also introduced a cost-optimized mode for Claude Code aimed at exactly this kind of workload. For casual or one-off tasks run at maximum settings, the benefit is less obvious, and Artificial Analysis’s finding that Fable 5.1 can cost more per completed task at max effort due to higher output token usage is worth keeping in mind before assuming any use of the model is automatically cheaper. Anthropic’s own top-line claim, that Fable 5.1 sits at the frontier of overall intelligence benchmarks, does appear to hold up independently: it reportedly posted the highest score recorded on Artificial Analysis’s composite index at the time of release, ahead of Anthropic’s own Opus 5 and OpenAI’s GPT 5.6. Whether that intelligence edge is worth the price depends heavily on whether your workload is cache-friendly and agentic, or a scattered set of high-effort one-shot requests. Frequently Asked Questions What is the difference between Fable 5.1 and Mythos 5.1? They’re reported to be the same base model with different safety guardrails applied. Mythos 5.1 has fewer restrictions and is currently limited to trusted partners, while Fable 5.1 is the publicly available, more heavily guardrailed version. Did Anthropic actually lower prices with Fable 5.1? One coffee. One working app. You bring the idea. Remy manages the project. Per-token input and output pricing stayed the same as Fable 5. The savings come from a large cut to cached prompt read pricing and from the model needing fewer tokens to complete tasks, which Anthropic estimates saves 25% to 45% depending on the workload. Is Fable 5.1 cheaper than Fable 5 in every case? No. Independent benchmarking found that at maximum reasoning settings, Fable 5.1 can use significantly more output tokens than Fable 5, which can make cost per completed task higher despite cheaper cache pricing. Savings are most reliable in agentic, cache-heavy workflows. What benchmarks improved the most with Fable 5.1? Reported gains are largest on Terminal Bench, Terminal Bench Science, and Cursor Bench, all of which test coding, scientific reasoning, and multi-step agentic tasks. Improvements on broader benchmarks like Humanity’s Last Exam were smaller. Does Fable 5.1 still refuse requests as often as Fable 5? Anthropic reports meaningful reductions, roughly 60% fewer cybersecurity-related refusals and about 85% fewer fallback refusals on biology and medical questions, but refusals still occur, including for some research-related queries.