Anthropic releases Claude Fable 5.2 and Mythos 5.2, pairing new, record-high benchmarks with 20% API cost reduction

What's better than a new frontier model? Try two, both of them achieving new high scores on a variety of third-party benchmarks — all with a 20% cost reduction over API for enterprises and developers.

That's exactly what Anthropic has done today: less than an hour ago, the fast-growing generative AI startup announced Claude Fable 5.2 and Claude Mythos 5.2, its latest, most powerful frontier language models for general usage and cyber workflows, respectively, and the results across the board are, once again, nothing short of impressive — dare we say, potentially industry-changing.

Both models now cost $8 per million uncached input tokens and $40 per million output tokens, down from $10 and $50 for their predecessors in the 5.1 generation, respectively.

Fable is available right now on paid Claude plans, in Claude Code and through the Claude Platform (model ID claude-fable-5-2), AWS, Google Cloud and Microsoft Azure.

Mythos uses the same underlying model with different safeguards and is reserved for vetted cybersecurity and life sciences users in Anthropic's Project Glasswing program.

The most useful number for enterprise engineering teams is likely $6.72. That was the average cost of a task on Cognition’s FrontierCode 2.0 test with Fable 5.2 at High effort, according to Anthropic's System Card on the new models.

Opus 5.5, which charges half as much for standard uncached input and output, averaged $12.82 at Max for a nearly identical score. The trade can also go the other way. Turning Fable all the way up to Max drove the average cost to $44.70 and brought its score below what it achieved at the cheaper Xhigh setting, according to Anthropic’s system card.

That makes Fable an interesting option for jobs that already burn time and money in retries, such as a migration across an unfamiliar codebase or a bug that survives several rounds of diagnosis. Anthropic itself recommends Opus 5.5 for most coding and knowledge work. Fable can take over an entire job or serve as an advisor that a less expensive model consults when it gets stuck. Each setup buys a different amount of Fable’s time.

How Fable and Mythos compare with rival models

Less than two weeks after Google unveiled Gemini 4 Carbon, the search giant’s latest frontier model, Anthropic has put up a set of scores that reshuffles the field.

But the two releases look different from the buyer’s side. Google announced Carbon on October 28 and began by giving access to selected cybersecurity defenders through its Fairwind program. Carbon remains in limited release as of November 9.

Most developers who want to put a Gemini 4 model to work today can use Gemini 4 Argon. Anthropic has opened Fable 5.2 much more broadly, while reserving Mythos 5.2 for vetted cybersecurity and life sciences teams.

In Anthropic’s launch comparison, one of the two new Claude models leads on 13 of 15 reported score lines, counting the two ways of scoring the AA-Briefcase knowledge-work test separately.

Carbon takes the business automation test, 53.9% to Fable’s 51.4%. The other win stays in the Claude family. Opus 5.5 edges Fable on AA-Briefcase’s head-to-head ranking, 1,800 to 1,785, even though Fable more often meets the test’s written requirements.

The tally shows the breadth of Anthropic’s gains across its chosen tests. Several rows have no score for Carbon, and Artificial Analysis gave Carbon 59 points on its separate Intelligence Index, a broader measure that answers a different question.

Coding is where Fable’s lead is easiest to see. On FrontierCode 2.0, which asks whether a patch is good enough to merge into a real open-source repository, Fable scores 49.1%, ahead of GPT-6 Astra at 42.8% and Carbon at 41.1%. Fable ran at Xhigh effort and Carbon at Medium.

Fable also leads both on FrontierSWE v2, scoring 69.9% against 65.5% for Astra and 61.2% for Carbon. Mythos leads the reported cyber and protein-design results, though Anthropic’s system card notes that labs used different tasks for CyScenarioBench, making that score a looser comparison with GPT-6 Cyber.

Early users describe gains that are harder to capture in a table. Vercel engineer Joe Haddad told Anthropic that projects his team had spent weeks pushing along with other models were finished and approved in a few days with Fable 5.2. Datadog senior staff engineer Matt Lee likened the model to an incident commander that could send faster models to gather information while it coordinated the work.

Both accounts come from testimonials Anthropic selected for the launch. They point to two things engineering teams can measure for themselves: how long it takes to get a change through review and how often someone has to step in to redirect the agent.

At Anthropic’s standard API rates, Fable and Mythos each cost $8 per million uncached input tokens and $40 per million output tokens. That’s 20% below GPT-6 Astra’s standard short-context rates and twice the standard rates for Opus 5.5 and Meta’s Muse Wonder.

The price per assignment can tell a different story. On FrontierCode 2.0, Fable at High averaged $6.72 per task, almost exactly what Astra spent at Max ($6.68) and roughly half what Opus spent at Max ($12.82) for nearly the same score. Turning Fable up one notch to Xhigh bought the top score and raised the average bill to $14.19.

Meta released Muse Wonder on October 12 through its API and coding app. Its results on DeepSWE and SWE-Marathon make it a plausible lower-priced option for sustained repository work, with the head-to-head cost still an open question.

Google announced Carbon at $2 input and $10 output per million tokens during its introductory period, rising to $4 and $20 afterward. Those rates make Carbon look inexpensive on paper. For most buyers, the missing piece is access to an endpoint where they can measure what it spends on their own work.

DeepSeek offers a different bargain. Its hosted V4 Pro 0813 costs $1.32 input and $3.96 output per million tokens at peak rates, with off-peak rates half as high. The Pro model handles text, while the cheaper V4.1 Flash 1017 can also take images. Companies can run the MIT-licensed Pro weights on their own infrastructure and pay to operate it themselves.

However, DeepSeek has already been hit with a U.S. government ban for agency work, likely making it (and other low-cost Chinese models) a tougher sell for enterprises and developers who count the government among their clients or prospective clients.

Benchmarks favor different effort settings for different jobs

FrontierCode evaluates whether a patch meets the standard for merging into a real open-source repository.

Across the 132 tasks in the current suite, Fable’s score rose from 37.6% at Medium to 45.3% at High and 49.1% at Xhigh.

The system card’s cost breakdown puts those three settings at $2.48, $6.72 and $14.19 per task, respectively.

Fable High cost almost exactly what Astra Max did on this test, with a score 2.5 points higher. Cognition reported no uncertainty intervals, and the Fable Max result would drop from 45.7% to 44.4% if timed-out runs received zero credit.

Medium remains the default in Claude apps and Claude Code, so the headline result takes a deliberate setting change.

The 5.1 comparison shows how far those settings can move the economics. In the same FrontierCode evaluation, Fable 5.1 peaked at 32.4% at Xhigh, averaging $8.68 per task.

Fable 5.2 at Medium beat that score by 5.2 percentage points at about 71% lower cost per task. At Xhigh, the newer model improved on its predecessor by 16.7 points and cost about 63% more per task, despite the lower list price. An upgrade with the effort setting pinned to Xhigh could increase spend.

An upgrade with a quality target and room to dial effort down could cut it.

Opus 5.5 remains competitive where the gains are smaller or depend on how the test grades the work. On the system card’s offline OSWorld 2.1 test, Fable 5.2 completed every required step on 54.6% of tasks, versus 53.2% for Opus and 47.6% for Fable 5.1.

On Artificial Analysis’s AA-Briefcase v1.1 knowledge-work test, reported in the same system card, Fable 5.2 beat Opus on the rubric pass rate, 66.1% to 61.2%, while Opus led the head-to-head Elo ranking, 1,800 to 1,785. Teams choosing a default model for analyst work can grade outputs both ways, checking whether they meet explicit requirements and whether reviewers prefer the finished deliverable.

The other launch charts make the case for tuning effort by workload. On Anthropic’s WANDR chart, Fable at Medium roughly matches Opus’s best result on large data-collection assignments for about 40% of the cost.

On Humanity’s Last Exam with tools, Anthropic says Fable at Low reaches roughly the same score as Opus at Max for less than a tenth of the cost. Hardware engineering shows the opposite pattern.

Fable reaches 83.5 out of 100 on BenchCAD 2.1 with tools at Max, versus 65.3 for Opus, and its score rises at every successive effort setting in the chart. A team spending hours reviewing a board design may have good reason to pay for the extra pass. A search agent returning a list of sources may hit its quality target at Medium.

On business automation, Zapier’s test asks models to handle realistic assignments across apps and checks whether they leave those apps in the expected state. The system card’s AutomationBench analysis puts Fable at 51.4% at Max, up from 39.7% at Low.

That is a meaningful gain and still leaves nearly half the test cases short of the full success criteria. For workflows that write to a CRM or trigger downstream actions, the useful pilot metric is the share of jobs that land every required update and obey every policy check, including how often a human has to repair partial changes.

A lower rate changes the bill. Agent behavior changes it more.

According to Anthropic’s Claude Platform pricing table, Fable 5.1 and Mythos 5.1 both cost $10 per million uncached input tokens and $50 per million output tokens. Both 5.2 models now charge $8 and $40.

Here's how they compare to other frontier models in listed API token price — among the higher end, but still below their predecessors and OpenAI's top-performing models.

Cache writes and reads fall by the same proportion, so an identical mix of input, output and cache activity costs 20% less at standard rates. Anthropic separately reports that Fable 5.2 can do some challenging work at Medium with fewer tokens than Fable 5.1, which can lower the bill further on those tasks.

The 5.2 release cuts the base rates as well. Opus 5.5 still costs half as much for uncached input and output, while Fable 5.2, Mythos 5.2 and Opus 5.5 now charge the same $0.20 for a million cached input tokens at standard rates.

Anthropic’s Batch API prices halve the input and output rates for asynchronous requests, putting Fable and Mythos 5.2 at $4 and $20, the same as standard interactive Opus 5.5. Opus itself drops to $2 and $10 in batch.

Anthropic says the new models generate tokens up to 30% faster than their 5.1 predecessors. Its Fast mode can increase token generation speed by up to 2.5 times at double the token price.

Anthropic lists Fast mode on Claude Code, the Claude Platform, Claude Models on Google Cloud and the Anthropic-hosted option in Microsoft Foundry. That can make sense during an interactive debugging session or an incident where a developer is waiting for the next result. The same premium buys less when a job spends most of its time waiting for builds, tests or external services.

A simple invoice comparison shows how much the routing decision matters. Ten million uncached input tokens and one million output tokens would cost $150 on Fable 5.1, $120 on Fable 5.2 standard, $240 on Fable 5.2 Fast and $60 on Opus or Muse Wonder at standard rates, assuming each model uses the same number of tokens.

An actual agent job rarely holds consumption constant. It can run more tests, re-read logs, retry after failures or delegate to several other agents. The system card’s cost calculation for Zapier also uses a $0.20 cache-read rate and a $10 five-minute cache-write rate per million tokens, illustrating how repeated use of a stable prompt can change the bill.

For engineering leaders, a useful bake-off starts with real issues the team can grade against existing tests and review standards. Run the incumbent alongside Fable at Medium and High, then escalate a handful of unresolved cases to Xhigh. Count merged or otherwise accepted fixes, include failed attempts in the spend, and track reviewer time and regressions. Those measurements can show whether Fable pays for itself as the first model on a narrow class of issues or as the model that gets called after the first attempt fails.

Mythos competes on access as much as capability

Anthropic’s system card reports that Mythos 5.2 rediscovered 89.6% of known vulnerabilities in its evaluation, up from 78.3% for Mythos 5.1. Opus 5.5 scored 81.3% with its cyber safeguards disabled, a test configuration that matters when comparing it with the standard API offering.

The two Mythos scores show progress on Anthropic’s own test. OpenAI’s 95% result measures successful exploitation of known flaws in a controlled remote environment. Security teams evaluating both models can separate discovery from reproduction, patch quality and verification on an authorized copy of their own software. A model that turns up more leads can still cost more to operate if analysts have to validate a larger queue.

The access distinction can also matter within a Fable workflow. Anthropic’s system card says some blocked cyber requests can fall back to Opus 4.8 and some blocked biology requests to Opus 5 in first-party products and for API users who opt in. Behavior can vary across providers.

Anthropic’s data-handling announcement also specifies 30-day retention for safety monitoring for customers outside its eligible zero-data-retention arrangements. Those details affect whether an organization can give an agent sensitive logs, source code or research data, and whether two runs through different channels are directly comparable.

The same system card describes a test in which Anthropic placed credentials or files outside a sandbox and measured whether models used them. Mythos 5.2 did so in 3.1% of scenarios, close to Opus 5.5 at 3.3% and down from 21% for Mythos 5.1. Those are results from a purpose-built evaluation.

For a production coding or security agent, the immediate design question is which credentials, network destinations and write permissions the job actually needs, especially if it delegates work to other agents.

The 1,000-agent demo puts the opportunity in perspective

Anthropic’s Fermat’s Last Theorem experiment shows what a large pool of agents can do when progress is measurable. In a single run per model, roughly 1,000 agents using an earlier Fable 5.2 snapshot formalized the existing proof in Lean in 30 hours.

The equivalent Opus team finished about 27%. Lean could check the resulting proof step by step. In a separate biology demonstration, a method Mythos developed with two protein-folding tools accurately predicted 28 of 34 capsid structures against experimental references.

The robotics results offer another example of where the evaluation setup matters. Anthropic reports 64.9% on RoboDojo Sim, versus 39.4% for Opus 5.5 and 29.8% for GPT-6 Astra in its runs. The score comes from controlling a simulated two-arm robot through camera images.

Anthropic separately showed physical demonstrations. A robotics team would still have to establish cycle time, failure recovery and performance on its own hardware before treating the model as a candidate for an operating line.

Most teams can start much smaller. A migration with a test suite, an investigation with reproducible failures, or a design with measurable constraints gives an agent a way to check progress and a reviewer a way to judge the result.

That is where Fable 5.2’s price cut and better performance at selected effort levels have the clearest chance of showing up in a budget. The buying decision comes down to how often it finishes valuable jobs that the existing stack leaves for a human, and how much it costs each time it does.