Anthropic cuts Claude 5.2 API prices by 20%

Two models, different access

Fable 5.2 is the main model for a broad range of tasks. Mythos 5.2 uses the same base model with different safeguards; access is intended for vetted teams working in cybersecurity and the life sciences.

On Cognition’s FrontierCode 2.0 benchmark, which assesses whether code changes are good enough to merge into a real open-source project, Fable 5.2 at High averaged $6.72 per task. Opus 5.5 at Max cost $12.82 per task and delivered a nearly identical result. But setting Fable 5.2 to its highest effort level, Max, raises the average cost to $44.70, while the result is lower than with the cheaper Xhigh setting.

That makes Fable a possible fit for tasks where failed attempts already consume time and money—for example, porting an unfamiliar codebase or finding a bug that has survived several rounds of review. Anthropic still recommends Opus 5.5 for most coding and knowledge-work tasks. Fable can handle a task end to end or help a cheaper model when it gets stuck.

Fable is available under the model ID claude-fable-5-2. Mythos 5.2 remains closed to most users.

Results against competitors

In a comparison published by Anthropic, one of its two new Claude models leads in 13 of the 15 rows of results. The AA-Briefcase knowledge-work benchmark uses two separate scoring methods.

On Google’s business automation test, Gemini 4 Carbon scored 53.9%, while Fable 5.2 scored 51.4%. In another comparison, a model from the Claude family came out ahead: Opus 5.5 beat Fable in the AA-Briefcase Elo ranking, 1,800 to 1,785, although Fable more often met the benchmark’s written requirements.

These results cover tests selected by Anthropic, but they do not provide a full comparison with every competitor: Carbon has no results in several rows. Artificial Analysis also gave Carbon a score of 59 on its separate Intelligence Index, which measures other aspects of models.

Fable’s advantage is more pronounced on coding tasks:

Early user feedback is about more than benchmark scores. Vercel engineer Joe Haddad told Anthropic that projects the team had spent weeks working on with other models were completed and brought into alignment in a few days with Fable 5.2. Datadog senior staff engineer Matt Lee compared the model to an incident commander: it delegates information gathering to faster models and coordinates their work. Anthropic selected both testimonials. Teams can test these claims by measuring time to approve changes and how often a person has to step in and redirect the agent.

At standard rates, Fable 5.2 and Mythos 5.2 cost $8 per million input tokens and $40 per million output tokens, excluding caching. That is 20% below GPT-6 Astra’s standard price for short contexts and twice the standard rates for Opus 5.5 and Meta’s Muse Wonder.

The cost of completing a task changes the comparison. On FrontierCode 2.0, Fable at High averaged $6.72 per task—almost the same as Astra at Max ($6.68), and about half the cost of Opus at Max ($12.82) for a similar result. Xhigh gave Fable a better score but raised the average cost to $14.19.

Meta released Muse Wonder on October 12 through an API and a coding app. Its DeepSWE and SWE-Marathon results make it a possible lower-cost option for extended work on codebases, although a direct comparison of costs on identical tasks remains an open question.

Google announced introductory rates for Gemini 4 Carbon of $2 per million input tokens and $10 per million output tokens. After the introductory period, the price is set to rise to $4 and $20, respectively. Carbon looks inexpensive on paper, but most buyers cannot yet assess costs on their own tasks until they get access to the API endpoint.

DeepSeek offers a different pricing option. Its cloud-hosted V4 Pro 0813 costs $1.32 per million input tokens and $3.96 per million output tokens at peak rates; prices are half as high off-peak. Pro works with text, while the cheaper V4.1 Flash 1017 also accepts images. Companies can run Pro’s MIT-licensed weights on their own infrastructure and pay for it themselves.

DeepSeek has already been barred from use by US government agencies. That could make it harder to sell the model, and other low-cost Chinese models, to companies that have government organizations among their customers or prospective customers.

Effort settings affect price and results

The current version of FrontierCode has 132 tasks. Fable 5.2’s score rises with the effort setting: 37.6% at Medium, 45.3% at High, and 49.1% at Xhigh. The average cost per task at those settings is $2.48, $6.72, and $14.19, respectively.

Fable at High costs nearly as much as Astra at Max but scores 2.5 percentage points higher. Cognition has not published confidence intervals. Excluding runs that exceeded the time limit lowers Fable’s Max score from 45.7% to 44.4%.

The Claude and Claude Code apps default to Medium, so users need to change the setting manually to get Fable’s best FrontierCode result.

Comparing generations shows how effort settings change the economics. On the same benchmark, Fable 5.1 peaked at 32.4% on Xhigh, at an average cost of $8.68 per task. Fable 5.2 at Medium beat that result by 5.2 percentage points while cutting task cost by about 71%. At Xhigh, the new model improved on its predecessor by 16.7 points, but average task cost rose by about 63%—despite the lower base rate. An upgrade with the Xhigh setting unchanged could therefore increase spending, while moving to the new model and lowering the effort level could reduce it.

Opus 5.5 remains competitive where the gap is small or results depend on how they are measured. In the offline OSWorld 2.1 benchmark, Fable 5.2 completed all required steps in 54.6% of tasks, compared with 53.2% for Opus and 47.6% for Fable 5.1.

On Artificial Analysis’s AA-Briefcase v1.1 knowledge-work benchmark, Fable 5.2 scored higher on the percentage of criteria met: 66.1%, versus 61.2% for Opus. But Opus came out ahead in the model-comparison Elo ranking, 1,800 to 1,785. Teams choosing a model for analytical work can separately check whether its output meets specified requirements and whether reviewers prefer it.

Other Anthropic benchmarks also suggest that the right effort level depends on the task. On WANDR, a benchmark for large data-gathering tasks, Fable at Medium came close to Opus’s best result while using about 40% of the cost. On Humanity’s Last Exam with tools, Fable at Low scored about the same as Opus at Max for less than a tenth of the price.

The picture is reversed for hardware engineering. On BenchCAD 2.1 with tools, Fable at Max scored 83.5 out of 100, compared with 65.3 for Opus; Anthropic’s chart shows Fable’s score rising at each higher effort level. The extra processing may be worthwhile when checking a circuit-board design. A search agent that only needs to return a list of sources may be fine at Medium.

Zapier’s business automation test checks whether models can complete tasks across different applications and leave those applications in the required state. In the AutomationBench analysis in Anthropic’s system card, Fable’s score rose from 39.7% at Low to 51.4% at Max. That is a notable gain, but almost half the tests still fell short of all success criteria. For workflows that write data to a customer relationship management system or trigger follow-up actions, teams should measure the share of tasks that complete every update and rules check, as well as how often a person has to make corrections.

Token prices are only part of the cost

According to Anthropic’s Claude Platform pricing table, Fable 5.1 and Mythos 5.1 cost $10 per million input tokens and $50 per million output tokens, excluding caching. For both 5.2 models, the rates fell to $8 and $40. Among the prices for advanced models listed in the source, they remain in the upper range, but are below their previous rates and the prices of OpenAI’s highest-performing models.

The price list also includes Gemini 4 Carbon at its introductory rate, Grok 4.7 Fast in Cursor for contexts above 256,000 tokens and up to 500,000, Gemini 4 Argon after its introductory period, and Gemini 4 Carbon after its introductory period.

Cache write and read prices fell by the same proportion. So the same volume of input and output tokens and cache operations costs 20% less at standard rates. Anthropic also says that on some difficult tasks, Fable 5.2 at Medium uses fewer tokens than Fable 5.1, which could lower costs further.

Opus 5.5’s rates for input and output tokens without caching remain half those of Fable 5.2. The price per million input tokens read from cache is now the same for Fable 5.2, Mythos 5.2, and Opus 5.5: $0.20.

Anthropic’s batch API cuts input and output rates for asynchronous requests in half. For Fable and Mythos 5.2, they are $4 and $20 per million tokens—the same as Opus 5.5’s standard interactive rates. In batch mode, Opus’s rates fall to $2 and $10.

According to Anthropic, the new models generate tokens up to 30% faster than the 5.1 versions. Fast mode can speed up generation by as much as 2.5 times, doubling the cost of tokens. It is available in Claude Code, Claude Platform, Claude Models on Google Cloud, and Anthropic’s deployment on Microsoft Foundry. The extra cost may be justified when debugging or responding to an incident, while a developer waits for the next output. If most of the time goes to builds, tests, or external services, a faster model will help less.

For an illustrative workload of 10 million uncached input tokens and one million output tokens, the bill would be $150 for Fable 5.1, $120 for Fable 5.2 at the standard rate, $240 for Fable 5.2 Fast, and $60 for Opus or Muse Wonder at standard rates—assuming all models use the same number of tokens.

In real agent work, token use rarely stays the same: an agent may run extra tests, reread logs, retry after an error, or delegate part of the work to other agents. In its Zapier cost calculation, Anthropic’s system card uses a rate of $0.20 per million cache-read tokens and $10 per million tokens written to cache for five minutes. This shows that reusing an unchanged prompt also affects the bill.

To compare models, engineering leaders should use real tasks whose results can be assessed with existing tests and review criteria. They can run the previous model alongside Fable at Medium and High, then send a few unresolved cases to Xhigh. The calculation should include failed attempts, accepted fixes, reviewer time, and regressions. That can show whether Fable is a good first choice for particular tasks or whether it should be brought in after another model fails.

Mythos offers both capabilities and access constraints

Anthropic’s system card says Mythos 5.2 rediscovered 89.6% of known vulnerabilities in its evaluation. Mythos 5.1 scored 78.3%. Opus 5.5 scored 81.3% with its cybersecurity safeguards disabled—a distinction that matters when comparing it with the standard API version.

These results show Mythos’s progress on Anthropic’s own test. OpenAI’s 95% figure refers to successful exploitation of known vulnerabilities in a controlled remote environment. Security teams comparing models should separately test vulnerability discovery, attack reproduction, fix quality, and validation on an authorized copy of their own software. A model that finds more potential problems may require more resources if specialists have to review a long list of results.

Access restrictions may also affect work through Fable. According to Anthropic’s system card, in the company’s own products and for API users who have opted in, some blocked cybersecurity requests may be routed to Opus 4.8 and some biology requests to Opus 5. Behavior may differ across providers.

Anthropic’s data-handling announcement also says that for customers not covered by zero-data-retention terms, data is kept for 30 days for safety monitoring. This affects which logs, source code, and research data an organization can send to an agent, as well as whether runs through different channels can be compared directly.

The system card describes a test in which Anthropic placed credentials or files outside a sandbox and checked whether models would use them. Mythos 5.2 did so in 3.1% of scenarios, close to Opus 5.5 at 3.3% and below Mythos 5.1 at 21%. This was a specially designed test.

When using a model in a real coding or security agent, organizations should decide in advance which credentials, network addresses, and write permissions the task actually needs—especially if it delegates work to other agents.

What the experiment with a thousand agents showed

In Anthropic’s experiment on Fermat’s Last Theorem, about 1,000 agents running an early version of Fable 5.2 formalized an existing proof in Lean in a single 30-hour run. The Opus team completed about 27% of the work in the same amount of time. Lean made it possible to check the proof step by step.

In a separate biology demonstration, a method developed by Mythos using two protein-folding tools correctly predicted 28 of 34 capsid structures when compared with experimental data.

Robotics results also depend on the test conditions. In Anthropic’s trials, the model scored 64.9% on RoboDojo Sim, versus 39.4% for Opus 5.5 and 29.8% for GPT-6 Astra. The evaluation is based on controlling a simulated two-armed robot using camera images. Anthropic also showed trials with physical robots, but robotics teams need to separately test cycle time, recovery from failures, and performance on their own equipment.

Teams can start with more constrained tasks: porting code with a test suite, investigating reproducible failures, or working on a project with measurable requirements. In these cases, an agent can check its own progress and a specialist can assess the result.

Dataist’s take

Lower rates alone do not determine the cost of work: effort settings, retries, and the need for human review all matter. Fable 5.2’s results at particular effort levels are a practical selling point for Anthropic, but access to Mythos is limited to vetted users, and benchmark results do not replace testing on real work and an organization’s own equipment.