Claude Fable 5 and Mythos 5, the New Anthropic Models Explained for Builders
Vasim Gujrati
Solutions Architect, AI & Platforms, Unico Connect
In this article
On June 9, 2026, Anthropic released Claude Fable 5 and Claude Mythos 5, its most capable generally available models so far and the first public release of its frontier Mythos class. Software engineering is where the release stands out. On SWE-bench Pro, a test of difficult engineering tasks, Anthropic reports Fable 5 scores 80% and the restricted Mythos 5 scores 80.3%, against 58.6% for GPT-5.5 from OpenAI (Claude Fable 5 and Mythos 5 system card, section 8.1). For teams that build software with AI, that gap is the number to watch. Below are the two model configurations, the full benchmark table, the early production reports and the jobs where we think the premium price pays off.
Quick Answer
Claude Fable 5 launched as the most capable generally available model from Anthropic, built for long, autonomous coding and knowledge work and released on June 9, 2026 as the first public model in the Anthropic Mythos class. Claude Mythos 5 is the same underlying model with some safeguards lifted, and it is restricted to vetted cybersecurity partners in Project Glasswing. Both cost $10 per million input tokens and $50 per million output tokens, less than half the price of the earlier Mythos Preview. Fable 5 led the major models on agentic coding benchmarks at launch with 80% on SWE-bench Pro, and you can use it now through the Claude API, Amazon Bedrock and GitHub Copilot.
Unico Connect runs frontier models like these in production through its AI development practice, with a senior engineer reviewing the code.
Key Takeaways
- Fable 5 and Mythos 5 are one model shipped in two configurations. Fable 5 is the safeguarded, generally available version, and Mythos 5 is the less restricted version released only to approved partners.
- Coding is its strongest area, so run it on your own coding workloads before you standardize on it. Fable 5 scores 80% on SWE-bench Pro against 58.6% for GPT-5.5, and 29.3% on FrontierCode Diamond from Cognition against 13.4% for Claude Opus 4.8 (figures reported by Anthropic).
- Point it at long, autonomous work, because that is where the real shift is. Stripe reported a migration of a 50 million line Ruby codebase completed in a day. Cursor says it opened up long running problems that were out of reach for earlier models, and GitHub says it beat previous benchmarks on long running coding tasks.
- Keep a cheaper model in your stack for routine calls. At $10 and $50 per million tokens for input and output, the new model costs less than half the price of the Mythos Preview, but it still costs more than GPT-5.5 at $5 and $30.
- Safety is built into the public model, and regulated teams should still add their own guardrails on top. It ships with new classifiers, routed most biology and chemistry requests to Opus 4.8 at launch, and produced zero harmful single turn completions across 30 public jailbreak techniques in external testing.
What Changed in September 2026
Since this launch, Anthropic has shipped two releases that change the picture for Fable 5. Claude Fable 5.1 arrived on September 1, 2026, and the Fable 5 model page now carries a legacy label with a note to migrate, although Fable 5 is still available. Fable 5.1 keeps the same $10 input and $50 output price per million tokens, and Anthropic says it costs an estimated 25% less than Fable 5 for typical workloads, through cheaper cache reads. Its restricted counterpart, Claude Mythos 5.1, is available only through the Anthropic trusted access programs, with safeguards designed for cybersecurity and life sciences work.
Claude Opus 5.5 followed on September 22, 2026, at $4 per million input tokens and $20 per million output tokens. Anthropic says it performs at the level of Fable 5.1 on most work, and the Claude models overview now suggests starting with Opus 5.5 for most workloads and keeping Fable 5.1 for demanding reasoning and long horizon agentic work. The benchmarks and production reports below describe Fable 5 as it shipped in June, so run your own evaluation on Fable 5.1 and Opus 5.5 before you choose.
Fable 5 vs Mythos 5: What Is the Difference?
Anthropic built one frontier model and shipped it in two configurations.
Claude Mythos 5 is the full strength model. Anthropic has lifted certain safeguards on it and released it only to approved users, starting with its Project Glasswing cybersecurity partners.
Claude Fable 5 is that same model made safe for general use. It wraps the Mythos class capability in new classifiers and routing. At launch, most biology and chemistry requests fell back to Claude Opus 4.8. Offensive cybersecurity tasks are blocked, and attempts to distill the capabilities of the model into competing systems are refused.
When you call Fable 5 through the API, you get frontier capability with guardrails in front of it, and the raw model underneath is the same one the approved partners use.
The Benchmarks
These are the results Anthropic reported, with the comparisons it published.
| Benchmark | What it measures | Claude Fable 5 | Comparison |
|---|---|---|---|
| SWE-bench Pro | Difficult software engineering tasks | 80% | GPT-5.5: 58.6% |
| FrontierCode Diamond (Cognition) | High quality, maintainable agentic coding | 29.3% | Opus 4.8: 13.4%, GPT-5.5: 5.7% |
| Hebbia Finance Benchmark | Senior level financial reasoning | Highest score | n/a |
| Slay the Spire (persistent memory) | Long context decision making | 3x the gain from memory | Opus 4.8: 1x |
The SWE-bench Pro lead over GPT-5.5 is wide. On FrontierCode Diamond the new model scores more than double Opus 4.8, which suggests the gains are concentrated in long, multi step agentic work rather than single shot answers. Anthropic also calls Fable 5 the best available model for tasks involving vision, citing its ability to pull numbers out of scientific figures and to rebuild web app source code from a screenshot.
Treat vendor published benchmarks as the ceiling and validate on your own workloads before you commit. Even allowing for that, the spread here is large enough to matter.
The table below sets Fable 5 against the earlier Mythos Preview, Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro on the full benchmark suite Anthropic published, with the leading model in each row highlighted.
Anthropic published benchmarks, June 2026
| Capability | Fable 5 | Mythos Preview | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|---|
| Agentic codingSWE-bench Pro | 80% | 77.8% | 69.2% | 58.6% | 54.2% |
| Agentic codingFrontierCode Diamond | 29.3% | — | 13.4% | 5.7% | — |
| Agentic codingTerminal Bench 2.1 | 84.3% | — | 82.7% | 83.4% | 70.7% |
| Knowledge workGDPval-AA | 1932 | — | 1890 | 1769 | 1314 |
| Knowledge work, visionGDP.pdf | 29.8% | — | 22.5% | 24.9% | 16.7% |
| Spatial reasoningBlueprint Bench 2 | 38.6% | — | 14.5% | 36.2% | 26.5% |
| Tool useAutomationBench | 17.4% | — | 15.5% | 12.9% | 9.6% |
| Computer useOSWorld Verified | 85.0% | 85.4% | 83.4% | 78.7% | 76.2% |
| LegalLegal Agent Benchmark | 13.3% | — | 10.4% | 2.1% | 0.0% |
| Multidisciplinary reasoningHumanity's Last Exam (no tools) | — | 56.8% | 49.8% | 41.4% | 44.4% |
| Multidisciplinary reasoningHumanity's Last Exam (with tools) | — | 64.7% | 57.9% | 52.2% | 51.4% |
| BiologyBioMysteryBench (hard) | — | 29.6% | 40.0% | — | — |
| BiologyBioMysteryBench (human solved) | — | 82.6% | 80.4% | — | — |
| HealthHealthBench Professional | — | 64.7% | 56.9% | 51.8% | — |
How to read this: every figure in the Fable 5 column is Fable 5 alone, taken from table 8.1.A of the Claude Fable 5 and Mythos 5 system card. Anthropic publishes a separate and usually higher score for Mythos 5, the restricted model, which is not shown here. Where the card reports no Fable 5 result, the cell is left empty rather than filled with the Mythos 5 number. Mythos Preview was the earlier Mythos class model, released in April 2026 to a limited group of Project Glasswing partners. Terminal Bench 2.1 figures for GPT-5.5 and Gemini 3.1 Pro use the Codex CLI and Gemini CLI harnesses. Source: Anthropic system card, June 2026.
The publicly available Fable 5 posts these headline numbers on everyday coding, knowledge work and computer use. On cybersecurity and biology work, Anthropic reported launch results closer to Claude Opus 4.8 from the public model, because Fable 5 routes those questions through blocking safeguards and falls back. Outside those areas, Fable 5 leads almost every category Anthropic measured.
What It Does in Production
The launch partners Anthropic named reported production results that back up the benchmarks.
- Stripe said Fable 5 compressed months of engineering into days, completing a migration of a 50 million line Ruby codebase in a single day, work it had estimated at over two months for a team.
- Michael Truell of Cursor said it opened up a class of long running problems that were out of reach for earlier models.
- Mario Rodriguez of GitHub said it beat previous benchmarks on complex, long running coding tasks.
On the science side, the less restricted Mythos 5 reportedly accelerated aspects of drug design by around ten times, with nine of 14 protein targets yielding strong drug design candidates. In blinded comparisons, scientists preferred its molecular biology hypotheses around 80% of the time.
What these reports have in common is duration. They describe tasks that run for hours across many steps and hold context the whole way, a different kind of result from faster autocomplete.
Pricing and Availability
Fable 5 and Mythos 5 are both priced at $10 per million input tokens and $50 per million output tokens, with batch processing at $5 and $25. Anthropic notes this is less than half the price of the Claude Mythos Preview. It still costs more than GPT-5.5 from OpenAI, which lists $5 per million input tokens and $30 per million output tokens for prompts under 272K tokens.
Access at launch differs by model.
- Claude Fable 5 is live immediately on the Claude API, consumption based Enterprise plans, Amazon Bedrock, and GitHub Copilot. It was included at no extra cost on Pro, Max, Team, and seat based Enterprise plans from launch until Anthropic suspended access to both models on June 12, after the US government applied export controls to them. When Fable 5 returned on July 1, Pro, Max, Team, and select Enterprise plans included it for up to 50% of weekly usage limits in a promotion that ended on July 19. Today Max plans and premium seats on Team and seat based Enterprise plans include it for up to 50% of weekly usage limits, while Pro plans and standard seats need usage credits.
- Claude Mythos 5 is offered by invitation only to Project Glasswing partners for defensive cybersecurity work.
Safety and Access
Safety is why there are two models. Anthropic kept full Mythos 5 access narrow and shipped Fable 5 with new classifiers covering three areas. They block progress on exploitation and offensive cyber tasks, and at launch they fell back to Opus 4.8 for most biology and chemistry requests. They also block attempts to extract the capabilities of the model for competing models. An August 7 update cut biology fallbacks by about 85% in Anthropic testing, and dual use biology requests such as virology and toxicology now go to Claude Opus 5, as do flagged chemistry requests.
Anthropic reports that external red teaming found zero harmful single turn requests complied with when testers used 30 public jailbreak techniques. For teams in regulated industries that posture matters, though your own evaluations and guardrails still belong on top of the ones the model ships with.
What Fable 5 Changes for Teams Building With AI
At Unico Connect, AI native development is how we build, so a model launch lands straight in our daily work. Our engineers use Claude Code for development, reviews and tests every day. Roughly 80% of production code at Unico Connect is AI generated and engineer reviewed, and we ship AI integrations and agentic systems for clients worldwide. The question we ask of any release is what it changes in the work we already do, and for Fable 5 the answer falls in two places.
The first is large migrations and refactors. Big, mechanical, risky work that used to be quoted in months, such as framework upgrades, language migrations and dependency overhauls, becomes a supervised exercise measured in days, and the Stripe migration is the proof point. The second is longer agentic workflows, where higher reliability on long, multi step tasks raises the ceiling on what an agentic system can own from end to end before it has to hand off to a person.
As our Solutions Architect for AI and Platforms, I put it this way.
"We have built our engineering around AI assisted, long running work for two years, so a model that reliably runs multi hour tasks is not a novelty for us. It is a force multiplier on a workflow we already run. The teams that win the next year will not be the ones with access to Fable 5, because everyone has that. They will be the ones who already know how to wrap a frontier model in guardrails, evaluations, and human checkpoints. That is the part that takes real engineering."
We apply the same caution here that we apply to every agent we build. Capability is not the same as production readiness. A more capable model makes the demo easier, but you still need guardrails, cost controls, human escalation and evaluation on your own data. Closing the gap between a working demo and a system you can trust in front of customers takes engineering on top of any model upgrade, and that is how we approach AI development.
Where Fable 5 Fits Best
The most capable model is not automatically the right one for every job. Based on the work we do for clients, this is where Fable 5 earns its premium price and where we would reach for something cheaper.
It is a strong fit for these jobs.
- Large migrations and modernization. Framework upgrades, language migrations and legacy refactors are the long, high risk work that used to take months. The Stripe codebase migration shows what becomes possible when one model holds the whole context and runs for hours under supervision.
- Agentic workflows that span many tools. When a task needs an agent to plan, call several systems, recover from errors, and stay coherent across dozens of steps, the reliability gain is worth paying for. Most agent projects break on this kind of task, and it is also where the benchmark lead shows up.
- Repository wide engineering tasks. Test backfills, large refactors, security and dependency audits, and documentation generation across a big codebase all benefit from a model that stays focused across millions of tokens in a long running task.
- Dense knowledge work. Financial analysis, research synthesis, and reasoning over long, complex documents play to its highest scores.
It is probably overkill for the jobs below.
- Simple chatbots and FAQ assistants, whose predefined flows and short answers do not need a frontier model.
- High volume, low complexity classification such as tagging, routing and sentiment scoring, which runs cheaper and fast enough on a smaller model at scale.
- Short, single step tasks. If the job finishes in one turn, you are paying for autonomy you never use.
The rule we give clients is to match the model to the shape of the problem. Fable 5 is built for long, autonomous, high stakes work, so point it at problems with that shape and let a cheaper model run everything else.
Frequently Asked Questions
What is Claude Fable 5?
Claude Fable 5 launched on June 9, 2026 as the most capable generally available model from Anthropic. It is the first public model in the Anthropic Mythos class, built for long, autonomous coding and knowledge work, and it ships with safeguards that route sensitive requests to Claude Opus 4.8 or Claude Opus 5.
What is the difference between Claude Fable 5 and Claude Mythos 5?
They are the same underlying model, and the difference is the safeguards. Fable 5 is the safeguarded version available to everyone through the API and platforms like Bedrock and GitHub Copilot. Mythos 5 has certain safeguards lifted and is restricted to approved partners, namely the Anthropic Project Glasswing cybersecurity group.
How much does Claude Fable 5 cost?
It costs $10 per million input tokens and $50 per million output tokens, with batch pricing at $5 and $25. Anthropic says this is less than half the price of the earlier Claude Mythos Preview. It still costs more than GPT-5.5 from OpenAI, which lists $5 input and $30 output per million tokens.
Is Claude Fable 5 better than GPT-5.5?
On the coding benchmarks Anthropic published, yes. Fable 5 scores 80% on SWE-bench Pro against 58.6% for GPT-5.5, and 29.3% on FrontierCode Diamond from Cognition against 5.7% for GPT-5.5. Those are vendor reported figures, so validate on your own workloads before you standardize on either model.
How do I access Claude Fable 5?
Fable 5 is available now through the Claude API, consumption based Enterprise plans, Amazon Bedrock, and GitHub Copilot. After a June 12 to June 30 suspension, Anthropic included it on Pro, Max, Team, and select Enterprise plans for up to 50% of weekly usage limits until July 19, 2026. Max plans and premium Team and Enterprise seats still include it for up to 50% of weekly usage limits, and Pro plans and standard seats use usage credits.
Is Claude Fable 5 safe for business use?
Fable 5 was built specifically for safe general use. It includes classifiers that block offensive cybersecurity and capability distillation attempts. At launch they also routed most biology and chemistry requests to Opus 4.8, and an August 2026 update cut biology fallbacks by about 85% in Anthropic testing. External red teaming reported zero harmful single turn completions across 30 public jailbreak techniques. For regulated workloads, layer your own evaluations and guardrails on top.
The Bottom Line
Claude Fable 5 moves the frontier on long running coding work and brings Mythos class capability to anyone with an API key, at less than half the price of the preview. The benchmark leads are real and the early production reports are striking. For teams building software, the open question has moved from whether the model is capable enough to how you put that capability to work safely and reliably. Our AI development services page shows how Unico Connect builds production AI systems and agents on Claude.
