Anthropic shipped two new models on September 1, 2026: Claude Fable 5.1 and Claude Mythos 5.1. They are, by Anthropic’s own description, the same underlying model wearing two different sets of guardrails. Fable 5.1 is open to anyone with an API key. Mythos 5.1 is locked behind Anthropic’s trusted-access programs, reserved for vetted organizations working in cybersecurity and life sciences. The split is the headline. It tells you where Anthropic thinks the real risk in frontier AI now lives: not in the model’s raw capability, but in who gets to use the sharpest version of it.
The release lands five months after Anthropic first previewed the Mythos line and roughly three months after Fable 5 and Mythos 5 went to general availability. A point release that arrives this fast, with this much benchmark movement, is itself a data point about how quickly the coding-agent race is moving in late 2026.
What Anthropic actually announced
Anthropic published the announcement under the title “Introducing Claude Fable 5.1 and Claude Mythos 5.1” on its own site. The core claim: Fable 5.1 and Mythos 5.1 share one underlying model. Fable 5.1 runs with the full stack of consumer-facing safety classifiers. Mythos 5.1 runs with those classifiers relaxed for cybersecurity and biology work, and only for organizations Anthropic has separately vetted.
Both models carry a 1,000,000-token context window and support up to 128,000 tokens of output per response, matching the ceiling Anthropic set with Fable 5 back in June. Pricing holds at $10 per million input tokens and $50 per million output tokens, the same headline rate as Fable 5 according to Anthropic’s own model documentation. The money is in the fine print: cache-read pricing dropped 75%, cutting typical workload costs by roughly 25% and highly agentic, long-running coding sessions by up to 45%, per VentureBeat’s reporting.
That cache-pricing change is a bigger deal than it sounds for anyone running Claude Fable 5.1 in an agentic loop. A cache hit on input tokens now costs roughly 2.5% of the base input rate rather than the 10% multiplier Anthropic has used on other Claude models. For a coding agent that rereads the same repository context hundreds of times during a long task, that’s the difference between a session costing dollars and costing cents.
The benchmark numbers, and why they matter
Anthropic and independent trackers published a cluster of benchmark results alongside the release. On Terminal-Bench 4.0, a test that scores how well a model handles real command-line and coding tasks end to end, Fable 5.1 scored 55.8%, up from 42.0% for Fable 5 and ahead of Opus 5’s 52.3%. Mythos 5.1, running with its more permissive cybersecurity safeguards active, reached 60.9% on the same benchmark.
The jump is sharper on Terminal-Bench-Science 0.1, a newer benchmark aimed at multistep scientific and research workflows: Fable 5.1 scored 52.6% against Fable 5’s 24.7%, more than double the prior model’s result. That gain lines up with Anthropic’s own positioning of Fable 5.1 as built for long-horizon agentic coding and dense document work (spreadsheets, slides, research corpora) rather than single-turn chat.
Anthropic also claims Fable 5.1 beats OpenAI’s GPT-5.6 Sol across multiple benchmarks, a comparison that shows up consistently in outlet coverage of the launch, though as with any vendor-selected benchmark suite, that comparison should be read as Anthropic’s framing rather than an independent third-party audit.
| Benchmark | Fable 5 (June 2026) | Fable 5.1 (Sept 2026) | Mythos 5.1 | Opus 5 |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 42.0% | 55.8% | 60.9% | 52.3% |
| Terminal-Bench-Science 0.1 | 24.7% | 52.6% | Not separately disclosed | Not separately disclosed |
| Context window | 1,000,000 tokens | 1,000,000 tokens | 1,000,000 tokens | Not disclosed in this release |
| Max output | 128,000 tokens | 128,000 tokens | 128,000 tokens | Not disclosed in this release |
| Cybersecurity false positives (Claude Code) | Baseline | ~60% fewer | Relaxed classifiers | Baseline |
Source: Anthropic’s own release notes and benchmark disclosures, cross-referenced with Decrypt’s coverage of the Terminal-Bench scores and officechai’s benchmark breakdown.
Why two models with the same brain
The Fable/Mythos split did not start with this release. Anthropic previewed Claude Mythos back on April 7, 2026, describing it as one of the most capable models it had built, with particularly strong results on computer-security tasks. The company chose not to make that preview generally available. Its own assessment found the model could automatically develop functional cyberattacks at what Anthropic called a highly professional level, a capability serious enough that broad release was judged too risky.
Anthropic’s answer was to split the model in two. Claude Fable 5, announced in June 2026, is the same Mythos-class model wrapped in safeguards that block the highest-risk cybersecurity and biology responses. Claude Mythos 5 kept the full capability and went only to vetted partners inside Project Glasswing, a coalition Anthropic runs alongside AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks to harden critical software against AI-accelerated attacks, according to Anthropic’s Project Glasswing page.
That program has already produced a measurable result: Anthropic and roughly 50 partners used Claude Mythos Preview to find more than ten thousand high- or critical-severity vulnerabilities across widely used software, and by June 2, 2026, Anthropic had extended Glasswing to around 150 new organizations across more than fifteen countries, per TechCrunch’s reporting on the program’s expansion. Fable 5.1 and Mythos 5.1 inherit that same architecture: one model, two doors, and a vetting process that decides who walks through which one.
Inside the safeguards
Fable 5.1 runs three real-time safety classifiers layered on top of the base model: one for cybersecurity, one for biology and chemistry, and one aimed at blocking model distillation attempts. When a query trips one of these, it gets routed to a more conservative model rather than answered directly by Fable 5.1 itself. Anthropic says this release tuned the cybersecurity classifier specifically, cutting false positives inside Claude Code by around 60%. That matters for working developers: a security researcher running legitimate vulnerability-discovery work on Fable 5 previously risked getting flagged and redirected mid-task, breaking the workflow. Anthropic says Fable 5.1 can now be used to find software vulnerabilities but is still blocked from producing working exploit code.
Mythos 5.1 doesn’t carry that same cybersecurity ceiling, which is exactly why it stays behind a vetting wall. Anthropic has also signaled a separate biology trusted-access track is in development, aimed at giving life-science researchers access to Mythos-level biological reasoning while keeping the cybersecurity classifiers intact. Availability there is described as limited to trusted organizations rather than open enrollment, and Anthropic has not published a public waitlist process for either track.
Market impact: pricing pressure on the whole coding-agent stack
The 75% cut to cache-read pricing is the part of this release that will move the most money in the short term. Coding agents built on top of Claude, whether that’s Claude Code itself or third-party tools wrapping the API, tend to burn through cache reads far faster than fresh input tokens, since agentic loops repeatedly re-send large chunks of file and conversation context. Cutting that cost by three-quarters without touching the headline $10/$50 rate is a way to make agentic workloads dramatically cheaper without discounting the sticker price competitors compare against.
That’s a direct shot at OpenAI, whose GPT-5.6 Sol model is the one Anthropic explicitly benchmarks against in this release, and at Google, which has been shipping Gemini updates on a similarly tight cadence throughout 2026. Anthropic previously said it planned to compress its own model release cycle to roughly four to six weeks for smaller updates, a pace point release cadence this launch fits neatly inside. A model house that can cut effective agentic-workload cost by up to 45% while also beating its own prior release by double digits on coding benchmarks puts pressure on rivals to match on both axes at once, not just one.
For enterprise buyers already running large Claude Code deployments, the math is straightforward: teams running long agentic sessions against large codebases stand to see the biggest savings, since cache-read volume scales with context reuse. Teams doing mostly short, single-turn queries will see closer to nothing change, since the base per-token price didn’t move.
Competitive comparison
| Model / Vendor | Positioning | Context window | Access model |
|---|---|---|---|
| Claude Fable 5.1 (Anthropic) | Flagship coding and knowledge-work model, general availability | 1M tokens / 128K output | Open, via API and partner platforms |
| Claude Mythos 5.1 (Anthropic) | Same model, relaxed cyber/bio safeguards | 1M tokens / 128K output | Restricted, trusted-access programs only |
| Claude Opus 5 (Anthropic) | Prior-generation flagship | Not disclosed in this release | Open, via API |
| GPT-5.6 Sol (OpenAI) | Anthropic’s named benchmark rival | Not disclosed by Anthropic | Open, via API |
The most distinctive thing on that table isn’t a benchmark score, it’s the access column. No other major lab has publicly split a single model into two access tiers gated by safety classifier strength the way Anthropic has with Fable and Mythos. OpenAI and Google both publish usage policies and rate limits, but neither runs a formal trusted-access program comparable to Project Glasswing for a specific model variant. If Fable/Mythos performs well through this cycle, expect that structure, not just the benchmark scores, to be the part competitors start copying.
Historical context: how we got two Claudes
The Fable/Mythos split traces back to a specific problem Anthropic hit in early 2026: its own safety testing showed a frontier model was good enough at cybersecurity reasoning that giving everyone unrestricted access looked reckless. Rather than sit on the capability or ship it flat, Anthropic built a two-track release model. Claude Mythos Preview came first, in April, as a closed evaluation. Claude Fable 5 and Claude Mythos 5 followed in June, formalizing the split into a public tier and a gated tier built on the same weights. Fable 5.1 and Mythos 5.1, arriving September 1, are the first point-release iteration on that structure rather than a new generation.
That’s worth noting because it changes how to read the release. This isn’t a from-scratch model launch with a new training run behind it in the way Fable 5 was. It’s closer to what other software categories would call a maintenance release: better classifiers, cheaper caching, meaningfully higher benchmark scores, same architecture and pricing floor. Anthropic has now shown it can ship that kind of update on a roughly quarterly cadence, which is fast for a frontier lab but increasingly normal for the category.
What developers should actually do with this
For teams already on Fable 5 through Claude Code or the API, upgrading to Fable 5.1 should be close to a drop-in change; pricing structure and token limits are unchanged, and Anthropic is positioning this as a direct successor rather than a parallel option. The main integration work is verifying that any workflow tuned around the old cybersecurity false-positive rate still behaves the way you expect, since the classifier retune could change which queries get routed away from the model.
{
"model": "claude-fable-5-1",
"max_tokens": 128000,
"context_window": 1000000,
"pricing": {
"input_per_million": 10,
"output_per_million": 50,
"cache_read_per_million": 0.25
}
}That $0.25 cache-read figure is the number worth putting into any cost projection for agentic tooling built on Fable 5.1. Teams running high-repetition workloads, code review bots, long-running research agents, multi-step document pipelines, should re-run their cost models now rather than waiting for the next invoice to show the difference.
Enterprise and developer reaction
Coverage of the launch spread fast across mainstream tech outlets within hours, a sign of how closely the developer and enterprise-AI press now tracks Anthropic’s release cadence. MacRumors and 9to5Mac both framed the release around the lower false-positive rate and cache savings rather than the raw benchmark jump, which suggests the practical cost and workflow-friction angle is what’s landing with working developers, not the leaderboard numbers.
Coding agents and IDE integrations
Claude Code and third-party IDE integrations that route through the Claude API pick up Fable 5.1 as the default flagship without any migration step required from users, consistent with how Anthropic handled the Fable 5 rollout in June. Teams building custom agents on top of the API get the new cache pricing automatically the moment they point their model string at Fable 5.1, with no separate opt-in.
What Mythos access actually requires
Anthropic has not published a self-serve application path for Mythos 5.1. Access runs through existing trusted-access relationships, largely built around Project Glasswing partners already working with Anthropic on critical-infrastructure security, plus the life-sciences track the company says is still in development. Organizations without an existing relationship to one of these programs have no current route to Mythos-level capability, regardless of budget.
Predictions: what happens next
1. Rivals respond on cache pricing, not base pricing. Anthropic didn’t touch its headline per-token rate, it restructured caching instead. Expect OpenAI and Google to follow with their own cache-cost cuts rather than base-price wars, since that’s the lever that actually moves agentic workload economics without disturbing sticker-price comparisons.
2. Project Glasswing keeps expanding, slowly. Given the jump from roughly 50 partners at the Mythos Preview stage to around 150 organizations across fifteen-plus countries by June, expect Anthropic to announce another expansion tied to Mythos 5.1, likely framed around the life-sciences trusted-access track it has said is in development.
3. Other labs adopt a tiered-safeguard release model. If the Fable/Mythos split holds up through another cycle without a major incident, it becomes a template. A single high-capability model shipped with a public safety-limited version and a gated full-capability version is a cleaner regulatory story than either withholding the model entirely or shipping it flat.
4. Benchmark gaps between Fable and Mythos narrow over time. The roughly 5-point gap between Fable 5.1 (55.8%) and Mythos 5.1 (60.9%) on Terminal-Bench 4.0 reflects how much the safety classifiers currently cost in raw performance. As classifier accuracy improves, expect that gap to shrink in future point releases, reducing the performance incentive to seek Mythos access.
5. Point releases keep landing roughly every 8-12 weeks. The gap between Fable 5 (June) and Fable 5.1 (September) is close to three months. Combined with Anthropic’s stated push toward faster release cycles, expect Fable 5.2 or an equivalent update before the end of 2026 rather than a full year-long wait for the next major version.
The risk side of the ledger
None of this erases the underlying tension. A model capable enough that Anthropic itself won’t release it without a vetting program is still, by definition, a model capable of serious harm in the wrong hands. Trusted-access programs reduce that risk, they don’t eliminate it: vetted organizations can still have their own credentials compromised, and a model that can find ten thousand critical vulnerabilities for defenders is the same model that could find them for an attacker if access controls slip. Project Glasswing’s expansion to fifteen-plus countries and roughly 150 organizations also means the attack surface for compromising Mythos access itself has grown right alongside the program’s defensive value.
Anthropic’s bet is that centralized, auditable, gated access beats either open release or permanent lockup. Whether that bet holds depends less on the model’s benchmark scores than on whether the vetting process around Mythos 5.1 stays airtight as the program scales into more countries and more partner organizations.
Frequently asked questions
What is the difference between Claude Fable 5.1 and Claude Mythos 5.1?
They are the same underlying model. Fable 5.1 runs Anthropic’s full set of consumer-facing safety classifiers and is available to anyone with API access. Mythos 5.1 runs with cybersecurity and biology safeguards relaxed, and is limited to organizations vetted through Anthropic’s trusted-access programs.
How much does Claude Fable 5.1 cost?
Base pricing is $10 per million input tokens and $50 per million output tokens, unchanged from Fable 5. Cache-read pricing dropped 75%, cutting typical workload costs by around 25% and highly agentic workloads by up to 45%.
What is the context window for Claude Fable 5.1 and Mythos 5.1?
Both models support a 1,000,000-token context window with up to 128,000 tokens of output per response.
Can I get access to Claude Mythos 5.1?
Only through Anthropic’s trusted-access programs, including Project Glasswing, which is aimed at organizations working on critical software security. Anthropic has not published an open enrollment or waitlist process for general access.
How does Claude Fable 5.1 compare to Claude Opus 5?
Anthropic reports Fable 5.1 scoring 55.8% on Terminal-Bench 4.0 versus 52.3% for Opus 5, putting the newer point release ahead of the prior flagship on that benchmark.
Does Fable 5.1 fix the cybersecurity false-positive problem from Fable 5?
Anthropic says Claude Code users should see around 60% fewer cybersecurity-related false positives with Fable 5.1, thanks to a retuned safety classifier.
Is Claude Mythos 5.1 the same as a “jailbroken” Claude?
No. Mythos 5.1’s relaxed safeguards are an official Anthropic release limited to vetted organizations, not a bypass of the model’s guardrails. General users cannot access it through normal API keys.
What is Project Glasswing?
A coalition Anthropic runs with AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks to use Claude Mythos-class models to find and fix vulnerabilities in widely used software.
