Boris Cherny stated that this is the largest leap in capability since Opus 4.5 in November 2025, as the model has begun to exhibit the hallmark qualities of a large model.

Article author, source: 0x9999in1, ME News



TL;DR

• Anthropic has officially launched Claude Fable 5, the first Mythos-level model available to the public; priced at $10/$50 per million tokens, more than halving the cost of the test version.
• FrontierCode Diamond difficulty evaluation: Fable 5 achieved 29.3%, GPT-5.5 only 5.7%, and Opus 4.8 just 13.4%.
• Stripe moved 50 million lines of Ruby code in one day—previously requiring a team two months to complete.
• The visual end can screenshot its way through Pokémon FireRed and reconstruct web source code; the biological end predicts adenovirus capsid assembly better than specialized protein models; single-cell genomics outperforms the Science paper’s model with only one percent of the parameters.
• Security follows a dual-track approach: the public version, Fable 5, implements hard limits, while the institutional version, Mythos 5, is selectively available to approximately 200 institutions through the Glasswing program.
• Boris Cherny stated that this is the largest leap in capability since Opus 4.5 in November 2025, as the model has begun to exhibit the hallmark qualities of a large model.

One, Anthropic turned the "press conference" into a "triage desk"

Let’s start with the conclusion. Fable 5 is not just a routine version update. It’s the first time Anthropic has clearly defined who can use it, to what extent, and at what price.

The version available to the public is Fable 5, with safeguards in place—prohibited for cybersecurity purposes and high-risk operations.

The institution received Mythos 5. Same architecture, fewer restrictions, under the Glasswing program, currently covering only about 200 institutions, including the U.S. government.

What about the pricing? Both are the same: $10 per million input tokens and $50 per million output tokens—more than halved from the testing period.

What is this? This is making "security" a part of the product, not just a PR slogan at a launch event.

In the past, when we saw large model releases, we were accustomed to a narrative: “We’re the strongest, the cheapest, the most secure.” All three claims shouted together. Today, Anthropic has broken this apart: placing its strongest capabilities within controlled channels, while the public version sacrifices some freedom in exchange for lower barriers to entry.

There is a clear pattern behind this. Previously, Mythos was used internally by Anthropic to discover "thousands of software vulnerabilities." What does it mean when this capability is unleashed? It means the same model can both patch and exploit vulnerabilities. Anthropic didn’t pretend otherwise—it chose a harder but more transparent path: tiered access.

Does that sound a bit like vaccine distribution logic? Yes—vaccinate high-risk groups first, then expand to the general population. The difference is that here, "high-risk" corresponds to high capability.

Two: 29.3% vs. 5.7%: This isn't a lead, it's a generational gap

FrontierCode's Diamond difficulty is one of the most widely recognized benchmarks in code evaluation.

Fable 5: 29.3%.

GPT-5.5: 5.7%.

Claude Opus 4.8: 13.4%.

29.3% versus 5.7%. The difference is more than five times.

This kind of gap is uncommon in the history of large models. What has been the norm over the past two years? It’s been a fierce race on the leaderboard, where a difference of three to five percentage points warranted a tweet celebration. It’s been OpenAI, Anthropic, and Google measuring their heights down to the decimal point.

It's different now. Fable 5 isn't just outpacing GPT-5.5 by a few lengths—it's leaving it behind entirely.

Of course, a single ranking cannot tell the whole story. So let’s look at real-world scenarios.

What did Stripe do with it? Migrated 50 million lines of Ruby code in one day.

What does 50 million lines mean? The entire core trading system of a large internet company is roughly this scale. With traditional methods, the same workload would take a team two months to complete.

One day versus two months. The person-month cost is flattened, reduced by nearly an order of magnitude.

What does this mean? It means that the process of code migration is shifting from an "engineering problem" to a "computing power problem."

Engineers won't lose their jobs. But the marginal value of engineers is being revalued.

III. What Does It Mean to Screenshot a Pokémon Model?

Anthropic included a slightly "playful" detail in the demo—Fable 5 can complete Pokémon FireRed directly through a screenshot.

Sounds like a toy?

No.

To complete an RPG, you need long-term planning, state memory, visual understanding, map awareness, dialogue inference, and goal decomposition. Together, this set of abilities forms the prototype of an Agent.

A more practical application is its ability to reconstruct a webpage's source code from a screenshot alone. This means the cost of front-end reverse engineering has been reduced to a level accessible to even ordinary designers.

Moving further toward the biological side, things get even more intense.

Fable 5 can predict the adenovirus capsid assembly of Dyno Therapeutics through reasoning alone—outperforming specialized protein models.

It independently completed a single-cell genomics study—designing, training, and classifying. The final model outperforms similar models published in Science, with only one percent of the parameters.

One percent.

This is a number worth pausing to think about for three seconds.

It shows one thing: general large models have begun to reverse dominate the territory where vertical scientific research models excel—not by brute force computing power, but by superior reasoning.

The research community has long harbored a concern: will general-purpose models eventually replace specialized models? The answer was once unclear. Today, Fable 5 offers a less favorable example.

Four: More than half off—generosity or cold calculation?

$10 in, $50 out.

Sounds expensive. But Anthropic's Dianne Penn offered a key point: Fable 5 requires fewer tokens to complete the same tasks than its predecessor.

That means a higher unit price but a lower total cost.

This is a very sophisticated pricing strategy.

It doesn't play the "low price" card. It plays the "cost per unit outcome" card.

It tells customers: Don't just look at the token price—look at how much it costs you to complete one task.

This script was previously used by cloud providers; now it’s being adopted by large model companies. This is a sign of industry maturity.

Why can the testnet pricing be halved directly? Two possible reasons.

First, the inference cost has truly come down. Second, Anthropic is grabbing market share.

I tend to think both are.

GPT-5.5's collapse on the coding leaderboard has given Anthropic a rare window of opportunity. When else would be the time to make a move?

But Anthropic didn’t drop the price to the floor. It left a buffer. Why?

Because it knows it’s not just selling tokens, but also a sense of peace of mind.

The public version has safeguards, and enterprise traffic is retained for 30 days for anti-jailbreaking purposes. The cost of this compliance must be absorbed from the gross profit.

Affordable, but not discounted. This is Anthropic’s stance.

Five: "Model Taste": What Boris Cherny Saw

Boris Cherny, the creator of Claude Code, used very strong language in this evaluation.

He said that Fable 5 represents Anthropic's largest leap in capability since the release of Opus 4.5 in November 2025.

After the release of Opus 4.5, Cherny uninstalled the IDE and switched to pure terminal-based programming. This caused quite a stir in the development community at the time.

And Fable 5 made him feel that Claude was no longer just a "code agent," but a "thinking and design partner."

What is the most interesting detail?

Cherny found that Fable 5 actively adds logs, performs measurements, and validates during debugging. Before declaring itself "fixed," it first confirms whether it has truly been fixed.

Sounds like nonsense? No.

What was the most common issue with past models? "Confidently making things up." "I've fixed this bug"—then you run it, and it still fails.

Fable 5's performance on this matter was described by Cherny as having a "large model flavor."

What is "large model flavor"?

Not a sales pitch style—it's an engineer's instinct to doubt oneself and then re-verify.

More importantly, this behavior wasn't taught by Claude Code's system prompt—it's inherent to the model's personality.

That's interesting.

If a capability is induced by prompts, it is fragile, transferable, and replicable. If it is the model’s own "personality," then it is Anthropic’s moat.

Anthropic has spent the past two years selling a story—that its differentiation isn’t just “smarter,” but “more trustworthy.” Fable 5 is the first time this story has been backed by a product.

Six: When it comes to security, Anthropic did not take a third path.

Let’s return to security.

Fable 5 has deployed an independent security classifier. For high-risk queries, it automatically falls back to Claude Opus 4.8. Anthropic's data shows: average impact on less than 5% of sessions.

What does this mean?

This means that in over 95% of scenarios, users won’t perceive the safeguards.

This means that less than 5% of "borderline" requests will be quietly downgraded.

This is a "gentle" restriction.

But Anthropic isn’t playing innocent either. Enterprise traffic data is retained for 30 days, during which it is manually reviewed, and all reviews are logged.

This is a compliance design and also part of the legal evidence chain.

For Glasswing partners, cybersecurity restrictions have been lifted. Biomedical restrictions will be gradually lifted through a trusted access mechanism.

Please note the four characters "received access." It means that access is granted based on qualifications, not on a first-come, first-served basis.

After reading through the entire mechanism, Anthropic is doing one thing: positioning itself as the "licensed regulator" in the large model field.

This is the fundamental difference in its approach compared to OpenAI.

OpenAI's logic is "release first, patch later."

Anthropic's logic is "classify first, then release."

Which path is right? There’s no answer yet. But the market will vote with its wallet.

Seven: Soft Price Increase for Subscribers

The new model is fully available on API and Enterprise editions.

The handling of subscribed users is more interesting: free trial until June 22, then switch to point-based payment after June 23. Once sufficient computing power is available, they will be纳入 regular subscriptions.

What does it translate to?

It’s a gradual price increase, like boiling a frog in warm water.

First, let you taste the sweetness, then ask you to pay. Due to current insufficient capacity, high-end features are temporarily offered as "pay-as-you-go."

This approach has been used by mobile cloud providers for many years. Now, large model companies are also taking this path.

Subscriptions are no longer "unlimited monthly access." They are evolving into a structure of "basic plan + pay-as-you-go add-ons."

Is this a step backward?

No. This is a proven model that has been validated for thirty years in the SaaS industry. Large model companies are now developing mature revenue models.

Eight, How the Industry Landscape Changes After Fable 5

Let's take a step back.

In November 2025, Opus 4.5 was released. Anthropic solidified its position in the top tier of code models.

In the first half of 2026, GPT-5.5 was released. The market saw its first significant correction in expectations for OpenAI.

June 9, 2026 (Tuesday, Pacific Time), Fable 5 and Mythos 5 launch simultaneously.

Seven months. The rankings of the three leading manufacturers have been rearranged once.

This doesn't mean OpenAI lost. OpenAI still holds strong in video, voice, and consumer-facing areas. But in the enterprise market, code, and research, Anthropic has pulled ahead for a significant lead.

What about Google? No one mentioned it. That in itself is a signal.

What about domestic manufacturers? No one mentions them either. But they’re definitely pulling all-nighters.

Code migration, scientific reasoning, and visual agents—these three areas represent the key focus points of Fable 5, each corresponding to a trillion-dollar application market.

Whoever keeps up with these three areas is still at the table. Those who don’t will have a tough year ahead.

Nine. Several unanswered questions

At this point in the article, we must acknowledge that there are several questions without current answers.

First, can Fable 5's "self-verification" behavior remain consistently stable across long contexts and multi-turn conversations? Anthropic has provided only demos and early customer feedback, not statistically significant data.

Second, how has Mythos 5 performed in actual deployments at 200 institutions? This information is strictly confidential. We can currently only see the capabilities of the public version.

Third, when will the competitors strike back? OpenAI won’t stay silent, and neither will Google. The second half of the year is likely to bring another wave of releases.

Fourth, will prices drop further? At $10/50, the pricing is still expensive for small and medium-sized businesses. As computing power is further released, could we see prices of $5/25—or even lower?

No one can answer these questions right now.

But one thing is certain.

Ten, Final Words

The large model industry has entered a new phase.

In the past, it was about who was smarter.

Now it's a competition of "who is more controllable."

The future will be decided by who can make customers sleep better.

Fable 5 is not the end. It's just Anthropic's new answer to the industry.

The core message of this survey is simple:

Capabilities will continue to grow exponentially.

But permissions will become increasingly granular.

The price will gradually decline.

Trust will become the most expensive thing.

Who will ultimately prevail?

I don't know.

But right now, Anthropic is in a strong position.

It didn't shout, "We want AGI."

It simply and seriously split one model into two parts.

A gift for everyone.

A gift for "someone you trust."

This restraint is almost a luxury in today's AI industry.

And luxury goods have never been cheap.

References

1. Anthropic.Introducing Claude Fable 5 and Claude Mythos 5. Anthropic official blog, June 9, 2026.
2. Anthropic.The Glasswing Program: Restricted Access to Frontier Capabilities. Anthropic Policy Document, June 2026.
3. FrontierCode Benchmark Team.Diamond Difficulty Evaluation Results: June 2026FrontierCode Leaderboard, June 2026.
4. Stripe Engineering.Migrating 50 Million Lines of Ruby in a Day with Claude Fable 5. Stripe Engineering Blog, June 9, 2026.
5. Cherny, Boris.Claude Fable 5: From Coding Agent to Thinking Partner. Author's personal blog and X platform posts, June 9, 2026.
6. Dyno Therapeutics discloses joint research materials on control experiment results for adenovirus capsid assembly prediction, May–June 2026.
7. Penn, Dianne. Media interview remarks on customer feedback data regarding Fable 5 pricing and token consumption, June 9, 2026.
8. Anthropic.Opus 4.5 Release Notes. Officially released by Anthropic, November 2025.