Claude Fable 5.1: Same Price, 75% Cheaper Cache

Anthropic released Claude Fable 5.1 on September 1, 2026 at Fable 5's list price, but cut cache reads 75% to $0.25 per million and posted benchmark gains that now beat Opus 5 across the board. The real headline isn't a new sticker price, it's a quiet cost cut hiding in the cache line. These are Anthropic's own launch numbers, so treat them as a strong signal rather than independent proof. At Van Data Team, we read a launch like this for what it does to your monthly bill, not your benchmark envy.

Key Takeaways

• On September 1, 2026, Anthropic released Claude Fable 5.1 and a gated variant, Mythos 5.1, keeping Fable 5's $10 input and $50 output list price per million tokens.
• The one price that moved is cache reads, cut 75% to $0.25 per million, which Anthropic estimates lowers typical workloads about 25% and highly agentic ones up to about 45%.
• On Anthropic's launch benchmarks, Fable 5.1 now leads both Fable 5 and Opus 5 on all seven, most dramatically on Terminal-Bench-Science, at 52.6% versus Fable 5's 24.7% and Opus 5's 29.0%.
• Mythos 5.1 is the same model with different safety filters, offered under restricted access for vetted cybersecurity and life-sciences organizations.
• Van Data Team's recommendation: if you run Fable 5, upgrade for the cache savings; if you run Opus 5, weigh capability against Fable's double token price, measured on your own real tasks rather than a launch table.

What Did Anthropic Actually Launch?

Anthropic shipped an updated flagship model and a gated sibling, with the real news sitting in the pricing details rather than the model name.

Reported fact: On September 1, 2026, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. Fable 5.1 is generally available to anyone with a Claude account, with production safeguards in place. Mythos 5.1 is, per Anthropic, the same underlying model with different safety filters, offered through restricted-access programs for vetted cybersecurity and life-sciences organizations that need capabilities normally constrained by those safeguards. The base price stays at $10 per million input and $50 per million output, matching Fable 5.

Van Data Team analysis: Notice the split. Fable 5.1 is the version almost everyone will actually use, and Mythos 5.1 is a narrow, gated release for organizations whose legitimate work bumps into the standard safety limits. For a normal engineering team the practical model here is Fable 5.1; Mythos is worth knowing exists, but it isn't the one you'll be calling from an agent.

What Changed in Fable 5.1?

Two things changed, and the more important one is easy to miss because the list price didn't move at all. The headline change is to cache reads; the second is a jump in benchmark scores.

Reported fact: The base price is unchanged from Fable 5, but cache reads were cut 75% to $0.25 per million tokens, which is 2.5% of Fable's $10 input price rather than the roughly 10% multiplier most Claude models apply. Anthropic estimates the cut reduces typical workload costs by about 25% and highly agentic workloads by up to about 45%. On capability, Anthropic reports Fable 5.1 now leads both Fable 5 and Opus 5 on all seven of its published launch benchmarks, reversing cases where Opus 5 had previously beaten Fable 5.

Van Data Team analysis: The cache cut is the story, and it rewards a specific shape of workload. Agentic pipelines re-read a large, mostly-stable context on every step, system prompt, tools, prior turns, and that re-read is billed as cache. Slashing cache reads by 75% directly lowers the cost of exactly the agent loops teams are building most. If your usage is one-shot prompts with little reuse, you'll see less; if it's long-running agents, this is a real reduction the pricing page's headline number hides.

How Big Are the Fable 5.1 Benchmark Gains?

Large on Anthropic's numbers, and largest on the hardest agentic tasks. The standout is a science-in-a-terminal benchmark where the score more than doubled.

The clearest figures line up side by side. Every number here is from Anthropic's launch harness, comparing Fable 5.1 to Fable 5 and to Claude Opus 5.

Van Data Team analysis: Read the Terminal-Bench-Science row and the cost-per-task row together. The first shows Fable 5.1 more than doubling Fable 5 on hard agentic science tasks; the second, from Cognition rather than Anthropic, shows cost per task falling from $5.84 to $2.68, below Opus 5's $3.51. That combination, higher scores and a lower cost per task, is the rare kind of upgrade that improves both axes at once. Still, these are launch benchmarks, and a leaderboard win of a few points rarely survives contact with your specific workload.

How Does the Fable 5.1 Cache-Read Cut Actually Work?

Prompt caching lets a model store part of a prompt it has seen before, so repeated context is billed at a steep discount instead of full input price. Fable 5.1 makes that discount much steeper.

Here's the mechanic. When you send the same large context repeatedly, the system prompt, tool definitions, and prior turns of an agent, the model can serve those tokens from cache rather than reprocessing them at the full $10 per million input rate. On most Claude models, a cache read costs about 10% of the input price. Fable 5.1 drops that to 2.5%, or $0.25 per million, a 75% reduction on the cache line specifically.

• Fresh input tokens: still billed at the full $10 per million, unchanged from Fable 5.
• Cache reads: now $0.25 per million, down from roughly $1.00, the one number that moved.
• Output tokens: still $50 per million, also unchanged.

Van Data Team analysis: A rough example shows why it matters: suppose an agent re-reads a 50,000-token context on each of 20 steps, which is a million cached tokens per run. At the old cache rate near $1.00 per million, those reads cost about a dollar a run; at $0.25, about a quarter, before you count fresh input and output. Multiply across thousands of agent runs, and the cache line, not the token rate, is where the money is. That's precisely the spend Fable 5.1 targets, which is why the list price could stay flat while real bills fall for the teams running the heaviest agent loops.

What Does Fable 5.1 Mean for Agent Builders?

It means cheaper long-running agents, mostly, plus a stronger model for the hardest steps. The cache cut lands squarely on the cost structure of agentic work.

An agent isn't one prompt; it's a loop that re-reads a growing context repeatedly, so cache-read pricing quietly dominates the bill of any serious agent. Cutting that 75% is why Anthropic can claim up to 45% off highly agentic workloads while leaving the list price alone. For teams already building Claude agents, that's a direct reduction on the exact pattern they run, which is more useful than a headline price cut on tokens they use less of.

Van Data Team analysis: The lesson generalizes beyond this release: effective cost, not list price, is what you budget against. A model that looks identically priced can be materially cheaper for your workload because of one line, cache reads, that never appears in the marketing. We make the same point in token efficiency, and it's why we push teams to measure real cost per task rather than compare sticker rates, a discipline we detail in AI agent development cost.

What Should You Know About Mythos 5.1?

Mythos 5.1 is the launch's other half, and for most teams the short answer is that it isn't for you, which is by design. It's a gated release, not a general product.

Reported fact: According to Anthropic, Mythos 5.1 is the same underlying model as Fable 5.1, with different safety filters. It's offered through restricted-access programs for vetted cybersecurity and life-sciences organizations whose legitimate work needs capabilities the standard safeguards normally constrain. Fable 5.1, by contrast, is generally available with production safeguards in place, and it's the version behind a normal Claude account.

• Fable 5.1: general availability, standard safeguards, what your agents will call.
• Mythos 5.1: same model, relaxed filters, restricted to vetted organizations in specific fields.
• The overlap: identical core capability, so the difference is access and safety configuration, not raw intelligence.

Van Data Team analysis: The two-tier design is worth understanding even if you'll only ever touch Fable 5.1. It reflects a real tension: some legitimate cybersecurity and life-sciences work runs into the same guardrails meant to block misuse, so Anthropic gates the relaxed version behind vetting rather than shipping it broadly. For a typical engineering team, the practical takeaway is simple, build on Fable 5.1, and treat Mythos as a specialized program you'd apply to only if your domain genuinely requires it.

Should You Upgrade to Fable 5.1?

For Fable 5 users the answer is straightforward; for Opus 5 users it's a value judgment. The list price didn't rise, so the downside of moving is small.

• If you run Fable 5: upgrading is low-risk, since the price is the same and cache-heavy workloads get cheaper, so there's little reason not to.
• If you run Opus 5: weigh capability against cost, because Fable 5.1 leads Anthropic's benchmarks but still lists at double Opus 5's token price.
• If you run long agents: re-run your cost math, since the cache cut can change which model is actually cheapest for your loop.
• If you're unsure: test both on your real tasks and compare cost per completed task, not benchmark rows.

Van Data Team analysis: The trap is treating a benchmark lead as a reason to standardize on the priciest model. Fable 5.1 winning Anthropic's launch table doesn't mean it's the right default for everyday coding, where Opus 5 remains half the token price. The honest read is that Fable 5.1 got both better and, for agentic use, cheaper, while the Opus-versus-Fable value question we covered in our Fable 5.1 vs Opus 5 vs GPT-5.6 comparison still stands.

How Should a Team Respond to This Launch?

Respond by re-checking your own numbers, not by switching on the strength of a launch post. A model release is a prompt to re-measure, not a mandate to migrate.

• Identify your cache-heavy workloads, the long agents and repeated-context calls where the 75% cut will actually show up.
• Re-run cost per completed task on Fable 5.1 for those, and compare against Fable 5 and Opus 5 on your real work.
• Treat Anthropic's benchmark table as a signal to test, not a result to adopt, since your tasks aren't the harness.
• Keep your model choice portable, so acting on this launch, or the next one, is a config change rather than a project.
• Re-check on a schedule, because in this field a same-price, cheaper-cache release like this arrives every few weeks.

This is a measurement habit, not a migration. One re-run on your agentic workloads usually tells you whether Fable 5.1's cache cut moves your bill enough to matter, which is the only question that counts. From there, upgrades become a number you can defend, the same portability discipline behind our model portability work.

How Van Data Team Helps

Van Data Team helps teams turn a model launch into a decision backed by their own numbers. We start by finding the workloads where a change like Fable 5.1's cache cut would actually land, so you measure the thing that moves your bill.

From there, we run your real tasks through the new model, compare true cost per completed task against your current one, and keep your stack portable so acting on the result is easy. If you want help, our AI agent development and data pipeline development work covers the infrastructure this sits inside. The goal is simple: adopt what genuinely lowers your cost or raises your quality on your own workloads, and skip the upgrades that only move a leaderboard nobody bills you against.