Anthropic Releases Claude Fable 5.2 and Mythos 5.2, Token Prices Cut 20%

Just before its IPO, Anthropic released Claude Fable 5.2 and Mythos 5.2.

Both use the same base model, with the core difference being safety permissions: Fable 5.2 is aimed at general users and carries safety guardrails in the cybersecurity and biological domains; Mythos 5.2 is open only to vetted institutions, with the relevant restrictions lifted to support compliant cybersecurity and life sciences research.

The biggest upgrade in the new version is on long tasks, especially those related to robot training. At the same time, Fable 5.2's token price has dropped 20% compared with Fable 5.1.

Moreover, in terms of aesthetics and design, Fable 5.2 has also improved considerably compared with Fable 5.1.

Within half an hour of Fable 5.2's release, someone used the model to develop an FPS game with very refined visuals.

However, the independent evaluation firm Artificial Analysis published a long post complaining that Fable 5.2's intelligence level is indeed higher and its token price is indeed cheaper, but the average cost per task has risen. What Artificial Analysis found hardest to accept was that Fable 5.2 is far more "talkative" than before.

So where exactly did performance improve?

First, the scores.

Robot control RoboDojo Sim went from 27.6% to 64.9%, a direct 2.35-fold surge; chart recognition from 44.8% to 73.4%; remote professional tasks from 17.9% to 41.3%; research terminal from 52.6% to 74.6%; FrontierCode 2.0 from 32.4% to 49.1%; HLE with tools from 65.6% to 74.2%.

Computer Use OSWorld 2.1 only went from 83.5% to 89.3%, a modest gain; knowledge work AA-Briefcase only from 61.5% (ELO 1669) to 66.1% (ELO 1785).

In other words, on those long tasks that take more time and involve more tools, Fable 5.2's gains are large. But on questions that can be answered in a single round, there is almost no growth.

So how long is this long task, exactly?

Anthropic provided a set of controlled experiments: they used 1,000 Fable 5.2 agents and spent nearly 30 hours to complete a formal proof of Fermat's Last Theorem. Previously, an unreleased version, equipped with a dedicated "add-on" (a purpose-built harness for tool orchestration) plus full guidance from researchers throughout, still needed 11 days to do the same work.

And Fable 5.2 used only one-third of the action tokens of that unreleased version.

Switching to 1,000 Opus 5.5 agents, the results showed 26% completed in 18 hours, 27% in 30 hours, and then the task got stuck and made no further progress.

Besides obvious long tasks like mathematical reasoning, among all the long-task benchmarks in the performance table Anthropic provided, RoboDojo Sim is the item with the largest gain. The blogger @himanshu specifically posted a reminder that in Fable 5.2, particular attention should be paid to the two items of robotics and protein design.

This public benchmark has 42 simulated tasks and 18 real-machine tasks, with dual arms, and the model can only read the environment from camera images. The five scoring dimensions are generalization, memory, precision, long-horizon execution, and open-vocabulary instructions.

It tests looking at the images, remembering what came before, and continuously completing a series of actions; if the robot "goes crooked" during the process, it must correct itself.

Anthropic officially stated that it surpasses "most AI systems built specifically for controlling robots," because these tasks are ones the robots have never seen in the real world.

In the past, everyone thought robots needed specially trained embodied intelligence models to solve problems such as spatial understanding and motion planning. But Fable 5.2 is a general model, and any company can use it directly to train robots. This is equivalent to greatly lowering the training threshold for embodied intelligence companies.

And in the programming field where Anthropic is strongest, the official blog provided this chart, showing the scores of several of the strongest current large models on the FrontierCode 2.0 benchmark.

Anthropic also ran scores using different effort levels of Fable 5.2, and the results were very surprising—I even once suspected a typo when the table was filled in.

Fable 5.2's Xhigh mode is the strongest in the world, scoring 49.1% at $14.19 per task. But when effort is set to the highest max level, the score actually drops, to Max 45.6%, at an average of $44.70 per task.

That is to say, it spent more time thinking and more tokens, yet the result was worse than Xhigh mode.

Anthropic's official explanation is: high effort produces over-deliberation.

Anthropic believes that higher effort, when handling routine work, may "gather context beyond what the task requires and over-deliberate," and it also prepared prompts to prevent "over-planning." FrontierCode tests whether a patch is qualified enough to be merged into a real repository; the context is already in the diff, and turning a two-line fix into a refactor with claims is unqualified in this benchmark, hence the lower score than Xhigh.

In addition, Anthropic also said it may be because the model thinks too long and cannot submit on time, and timeouts lead to point deductions.

Fable's biggest feature is... cheap?

Fable 5.2's token price dropped from $10 per million input to $8, and output from $50 to $40, a 20% reduction; output speed is up to 30% faster than 5.1.

Fast mode is separate: input 16, output 80, double the price, with token generation up to 2.5 times faster.

Cache read is $0.20, versus $1 for Fable 5.

Anthropic said that setting Fable 5.2's effort to High yields results roughly comparable to Opus 5.5 at max (Max), while the cost per task is only about half of the latter.

But Anthropic's official post never explained why the price was cut.

So where did this 20% come from?

On September 12, Amodei posted "We Must Pace the Frontier," saying that slowing down "does not mean stopping model training or technological progress, but ensuring the company has enough time to align and add safeguards to models."

Ten days later, on September 22, Anthropic released Opus 5.5, with list prices cut 20% to 4 and 20, and said operating costs were about 40% lower than Opus 5. About two hours later, OpenAI released GPT-6 Sol and Luna, about 50% lower than GPT-5.6's promotional price.

Ramp economist Ara Kharazian said: "This is not some technological innovation; this is actually a price war, pushing down AI prices and also pushing down one's own profit margins."

There is another thing that makes people very curious: the version number is 5.2, while the previous Opus has already been updated to version 5.5. The New Stack said, "It's really a bit surprising that it wasn't called Fable 5.5." However, Anthropic gave no explanation for the version number.

Less than an hour after Fable 5.2 was released, the slap in the face came.

The independent evaluation firm Artificial Analysis published a long post. It said they put Fable 5.2 on Intelligence Index v5, where the max tier scored 65 and topped the list, 6 points higher than Opus 5.5's 59 and about 15 points higher than the previous generation Fable 5.1's 49, taking the highest score in five of nine evaluations.

Hold on—the Fable series has always been the ceiling of intelligence; that is beyond doubt. Artificial Analysis then wrote that the max tier costs an average of $26.5 per task, about 2.6 times Opus 5.5's $10 at max, and about twice that of Fable 5.1.

The unit price dropped 20%, but the total cost doubled. And Artificial Analysis "twisted the knife" further, saying that Fable 5.2 generated about 275 million tokens in the evaluation, while the median for similar models is only 72.8 million—"it talks too much."

One last point: performance scores can be measured numerically, and Anthropic said Fable 5.2 also upgraded and iterated on design.

So less than half an hour after Fable 5.2 was released, someone in the community tried using Fable 5.2 to develop an FPS game.

In the demo, guns, shooting feedback, and scenes were all there. Some netizens in the comments, after watching it, felt it could be released directly.

If you only know what kind of game you want to play and haven't even written the first line of code, then you can let Fable think a few more steps.

This article comes from the WeChat public account "Letter AI", author: Miao Zheng, published by 36Kr with authorization.