Claude Fable 5 and Mythos 5 are here. The tier above Opus is finally in general availability.
Table of Contents
Hello. On June 9, 2026, Anthropic announced new models: "Claude Fable 5" and "Claude Mythos 5."
When I introduced Opus 4.7 and Opus 4.8 on this blog previously, the existence of a superior model called Claude Mythos Preview frequently made an appearance. Some of you might remember me writing, "Opus is already more than strong enough, yet there's something even higher?" This announcement is about that Mythos-class finally descending into our hands in the form of general availability.
Honestly, I'm quite excited about this news. The fact that the top-tier class, which was previously limited, is now available via API and Claude subscriptions is an event that cannot be ignored for someone like me who uses Claude Code daily. Until now, the highest tier selectable in Claude Code was the Opus class, but if Fable 5, which sits above it, becomes available for daily work, the ceiling of performance for practical tasks will rise by another level.
In this article, based on the official announcement, I will organize what Fable 5 and Mythos 5 are and why they are beneficial for Claude Code users.
What is the relationship between Fable 5 and Mythos 5?
First, it's important to understand the relationship between these two models. According to the official explanation, Fable 5 and Mythos 5 are the same foundation model, and the only thing separating them is the presence or absence of safeguards.
- Claude Fable 5 - The Mythos-class model made safe for general use.
- Claude Mythos 5 - The same foundation model with safeguards removed in specific domains. It is provided exclusively to trusted partners of Project Glasswing, and it has been reported that its deployment is being coordinated with the US government.
The origin of the names was also introduced: Fable comes from the Latin fabula ("that which is told") and shares the same root as the Greek mythos (myth). The naming distinguishes the two models based on the presence of safeguards while sharing a common etymological root of "storytelling."
And regarding the term "Mythos-class," the official footnotes explain it as a new tier (layer) positioned above the Opus class. The first Mythos-class model was the Claude Mythos Preview released via Project Glasswing in April 2026. In other words, Fable 5 represents the first case where a model of a class exceeding Opus has been widely opened to the general public.
To summarize, the hierarchy of capabilities places the Mythos-class above the existing Haiku, Sonnet, and Opus lines. Fable 5 is the version of that Mythos-class made safe for general use, while Mythos 5 has some safeguards removed for trusted uses such as cybersecurity and life sciences research.
From here, we will look at the capabilities of Fable 5 in each domain as cited by the official announcement.
Software Engineering
As a Claude Code user, this is likely the first thing you're curious about. The official announcement explains that Fable 5 demonstrated state-of-the-art performance across nearly all verified benchmarks.
A specific case study introduced is Stripe. With Fable 5, months of work were compressed into days, completing the migration of a 50-million-line Ruby codebase in a single day. This task would reportedly have taken two months if done manually.
Additionally, in Cognition's FrontierCode evaluation, it was announced to have recorded the highest score among frontier models (the latest models from various companies). Furthermore, token efficiency has also improved compared to previous Claude models, suggesting progress toward handling the same tasks with fewer tokens.
Comments from various companies providing development tools were also introduced. Summarizing the main points:
- The CEO of Cursor evaluated it as the state-of-the-art model on their CursorBench, capable of solving classes of long-horizon problems that were previously out of reach.
- The CPO of GitHub stated that it handled complex, long-term coding tasks with autonomy and reliability exceeding conventional benchmarks.
The direction of "running long and autonomously" was a point emphasized in Opus 4.8, but it seems the Mythos-class has raised that bar even further.
Vision (Image Recognition)
Significant evolution in image recognition was also announced. According to the official report, Fable 5 has reached state-of-the-art performance in vision tasks.
The examples introduced are quite impressive:
- Accurately reading numerical values from charts and tables in scientific papers.
- Reconstructing the source code of a web app using only screenshots as a guide.
- Clearing Pokemon "FireRed" using a vision-only harness (image input only).
The last Pokemon example is an extension of previous efforts to measure capabilities by having Claude play games. While previous models required a combination of complex auxiliary tools, Fable 5 reportedly cleared it just by looking at the screen. The "harness" here refers to the peripheral mechanisms for the model to operate the game; the evolution is shown in the fact that it could proceed with image recognition alone even when that harness was minimized. Considering use cases like generating implementations from UI screenshots, this high level of precision is likely to be very effective in practice.
Long Context and Memory
The ability to handle long contexts has also improved. The official announcement mentions that Fable 5 can maintain focus over millions of tokens and effectively use persistent file-based memory.
An interesting verification result was introduced here. In a test using the card game "Slay the Spire," the performance improvement when given access to memory was three times that of Opus 4.8. Furthermore, under the memory-enabled condition, the frequency of reaching the final chapter of the game also tripled.
It's not just about the context being long; it seems the difference lies in whether the model can effectively utilize information when given external memory. This difference might be felt in long sessions or workflows where you want the model to remember past work.
Knowledge Work and Benchmarks
High evaluations were also reported in the field of knowledge work.
- Highest score on Hebbia's Finance Benchmark (measuring senior-level financial reasoning).
- Highly evaluated by IMC regarding trade analysis evaluations.
- On ViBench, it was the highest performing model tested, nearly saturating basic use cases.
- In a certain analytics benchmark, it broke the 90% barrier for the first time in the core evaluation (a 10-point increase over Opus).
Comments regarding physics research were also introduced. According to the CTO of Replit, it is the strongest model they have tested in cutting-edge physics research, reaching in 36 hours almost the same point that GPT-5.5 took 4 days to reach. Moreover, it was explained that the reasoning tokens used were about one-third of those used by GPT-5.5.
The ability to persist in fields where verification is difficult, like finance and physics, is not irrelevant to Claude Code users. This is because it translates to being a partner that can stay with you longer not just for writing code itself, but for "outside the code" tasks like deciphering specifications, research, and tackling complex problems. While benchmarks are only one aspect, a characteristic of this release is the improvement across the board in areas close to practical work: coding, finance, and physics.
Mythos 5 Scientific Achievements
From here, we look at the achievements of Mythos 5, which has some safeguards removed. While a bit removed from the daily use of Claude Code, it's worth touching upon briefly as it shows the raw power of the Mythos-class.
According to the official report, in the fields of drug discovery and life sciences, experiments with internal experts accelerated parts of protein design by about 10x, and in some tasks, produced results comparable to or exceeding skilled human operators (obtaining promising drug candidates for 9 of 14 targets). It was also explained as the first model capable of consistently generating novel and compelling scientific hypotheses. In genomics research, it worked almost autonomously for over a week, constructing single-cell data for 138 animal species and millions of cells. The machine learning model created there was reportedly 1/100th the size but outperformed results published in the journal Science.
This gives the impression that it has entered a stage where it can advance research itself, moving beyond the framework of a "convenient companion for AI coding."
Safety Mechanisms
Because the model is so capable, the design of its safety is crucial. Anthropic has explained this point in great detail.
In Fable 5, three classifiers are operating to automatically detect high-risk requests and route them to Claude Opus 4.8.
- Cybersecurity - Targets vulnerability exploitation and offensive cyber tasks. When an external partner tested it with 30 types of public jailbreak methods, it was reported that the number of times Fable 5 responded to harmful one-off requests regarding cyberattack planning, exploit development, or defense evasion was zero.
- Biology/Chemistry - Targets requests related to high-risk biological research, with Claude Opus 4.8 responding instead. Note that Mythos 5 reportedly outperformed specialized protein language models in predicting the assembly of Adeno-associated virus (AAV) shells.
- Distillation - Prevents the extraction of capabilities to train competing models.
In addition, in over 1,000 hours of red-teaming by external organizations, no universal jailbreaks were found. While the UK's AISI (AI Safety Institute) made limited progress toward one during a short verification period, one external partner evaluated it as having the "most robust safeguards tested." In terms of daily usability, it was explained that in over 95% of Fable sessions, no safeguard-triggered fallback (switching to Opus 4.8) occurs at all.
Data handling was also mentioned. Mythos-class traffic is retained for 30 days, but its use is limited to safety purposes and it is not used for training models. All human access is logged, and in almost all cases, it is deleted after 30 days.
Project Glasswing is Anthropic's initiative to provide cutting-edge AI exclusively to those responsible for cybersecurity and critical infrastructure. Mythos 5 will first be provided to Glasswing partners (with cyber safeguards removed), then expanded to some biological researchers (with bio/chem safeguards removed), with plans for a broader trust-based access program in the future.
As someone who continues to use Claude because I like Anthropic's philosophy, I find this stance of discussing safety design with the same passion as capability announcements to be truly trustworthy.
Pricing
Pricing was announced as follows:
| Item | Price |
|---|---|
| Input (per 1M tokens) | $10 |
| Output (per 1M tokens) | $50 |
This price is common to both Fable 5 and Mythos 5, and according to the official report, it is less than half the price of Claude Mythos Preview. Compared to the standard rates for the Opus class ($5 / $25), it is double the level, but considering that a class above Opus is now available at this price, it feels like it has landed in a much more realistic range than expected.
In the API, it can be used with the model ID claude-fable-5.
Availability and How to Try
The provision schedule was also specifically shown:
- Claude API - Available from the announcement date, June 9, 2026.
- Claude.ai Subscriptions - From June 9 to June 22, available at no additional charge for Pro, Max, Team, and seat-based Enterprise plans. From June 23 onwards, usage credits will be required, with plans to return to standard provision as capacity stabilizes.
In other words, as of June 10, 2026, when I am writing this article, we are right in the period where you can try it at no additional charge. If you're interested, it's a good idea to touch it during this free period.
When using it from Claude Code, I have already confirmed that Fable 5 can be selected from the /model command in my environment. I immediately tried running a Workflow with Fable 5 and hit the 5-hour Rate Limit (lol). While it's that easy to put into actual work, you should also be careful about limits during long automation tasks. If you can't see it yet, try updating and restarting.
claude update
Note that visibility and default behavior vary depending on the plan, provider, and administrator settings. For those still using Claude Code via Homebrew or npm, switching to a native installation at this timing will make following new features smoother as automatic updates take effect. Detailed steps are summarized in The story of how switching Claude Code from Homebrew to native installation made it more comfortable.
Comparison of Fable 5 and Mythos 5
The differences between the two models are summarized in the table below.
| Item | Claude Fable 5 | Claude Mythos 5 |
|---|---|---|
| Positioning | Mythos-class made safe for general use | Same foundation model with some safeguards removed |
| Foundation Model | Same as Mythos 5 | Same as Fable 5 |
| Safeguards | 3 classifiers operating | Removed in specific domains |
| Provision | General availability via API/Claude.ai | Limited provision via Project Glasswing |
| Main Intended Use | General development / Knowledge work | Cybersecurity / Life sciences research |
| Pricing | $10 / $50 per 1M tokens | $10 / $50 per 1M tokens (Same as Fable 5) |
Why this is beneficial for users
Let's reorganize the content so far from the perspective of "how it affects my work."
1. A class above Opus is normally usable
The Mythos-class, which was previously the "strongest but out of reach model," is now available via API and Claude.ai subscriptions. For Claude Code users, the biggest change is being able to incorporate the top-tier class into daily workflows if the environment allows.
2. Running long and autonomously
Improvements such as maintaining focus over millions of tokens and utilizing external memory will be effective in long sessions or when delegating large tasks. The result seen in the Slay the Spire verification—where the improvement when given memory was 3x that of Opus 4.8—suggests good compatibility with workflows that assume memory.
3. Easier to generate implementations from images
Vision performance capable of reconstructing web app source code from screenshots will likely lower the hurdles for tasks like UI reproduction and reading diagrams.
4. Design you can trust
The fact that safety designs like the three classifiers, 1,000+ hours of red-teaming, and limited data retention are discussed with the same passion as capabilities is a point that makes it easy to trust as a partner for work.
5. More realistic pricing than expected
By setting the price at less than half of Mythos Preview, the psychological hurdle to trying the top-tier class has dropped significantly.
Summary
The key points of Claude Fable 5 and Mythos 5 are as follows:
- The Mythos-class, positioned above the Opus class, has been opened for general availability for the first time.
- Fable 5 and Mythos 5 are the same foundation model; the only difference is the presence of safeguards.
- Announced to have shown state-of-the-art performance in many areas: software engineering, vision, long context, and knowledge work.
- Examples include completing Stripe's 50-million-line Ruby migration in 1 day and using 1/3 the reasoning tokens of GPT-5.5 in physics research.
- Mythos 5 reported results comparable to or exceeding human experts in drug discovery and genomics.
- Safety design with 3 classifiers and 1,000+ hours of red-teaming.
- Pricing is $10 / $50 per 1M tokens, less than half of Mythos Preview.
- API available from June 9; Claude.ai is no additional charge from June 9-22 (usage credits after June 23).
Just two weeks after Opus 4.8 came out, the class above it has descended to the public. Anthropic's speed of evolution is surprising as always, but this time I'm even more excited by the fact that "the Mythos-class that was always far above is finally within reach."
Since it's also a period where you can try it at no additional charge, if you have an environment where it's available, please check its capabilities in your own workflow.
