What Anthropic actually shipped
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, calling them its most advanced models for coding and knowledge work. The API model ID is claude-fable-5-1 and it is generally available on the Claude API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry, Claude.ai, Claude Code, and Claude Enterprise. Mythos 5.1 is the same underlying model with reduced safeguards, and access stays restricted to US organizations that pass the Cyber Verification Program or Life Sciences Verification Program. This is the two-tier structure Anthropic introduced with Fable 5 in June, so the launch iterates on a strategy rather than replacing it.
The platform documentation supplies the operating parameters a procurement team will care about. Fable 5.1 carries a 1 million token context window, a 128K maximum output, a June 2026 knowledge cutoff, adaptive thinking that is always on, and a default effort setting of high. Anthropic positions it for demanding reasoning and long-horizon agentic work where Opus 5 at higher effort still falls short. That framing is telling. The company is telling buyers that Opus 5 is the default and Fable 5.1 is the escalation path, which is the right way for most enterprises to read it.
The pricing story is about caching and token efficiency
The list price did not move. Fable 5.1 costs 10 dollars per million input tokens and 50 dollars per million output tokens, exactly what Fable 5 cost in June and double the 5 and 25 dollars Opus 5 charges. What changed is the cache read rate, which drops 75 percent to 25 cents per million tokens. Anthropic's arithmetic on that change is a 25 percent cost reduction for typical workloads and up to 45 percent for highly agentic work, where the same system prompt, tool definitions, and accumulated context get re-read on every turn.
We think this is the most consequential line in the announcement for anyone who has watched an agent pilot's bill. Long-running agents spend most of their tokens re-reading context they already saw, so the cache rate matters more than the marginal input rate. Customers echo the efficiency point. Dan Shipper, CEO of Every, describes the model as roughly twice as fast as Opus 5 at half the token usage. Alex Wang at Rogo reports that on its internal finance benchmark Fable 5.1 matches Fable 5 on accuracy while using 20 percent fewer tokens. If those numbers hold in your environment, the comparison against Opus 5 becomes much closer than list price suggests.
Where the benchmark gains land
Anthropic's comparison table shows the largest deltas on agentic and scientific evaluations. Terminal-Bench 4.0 rises from 42.0 percent on Fable 5 to 55.8 percent, with Mythos 5.1 at 60.9 percent, and AutomationBench nearly doubles from 17.1 to 31.4 percent. Terminal-Bench-Science jumps from 24.7 to 52.6 percent. OSWorld 2.0 in strict scoring moves from 36.1 to 41.7 percent, and CursorBench 3.2.0 edges up from 70.5 to 73.4 percent. Against the model Anthropic lists as GPT-5.6 Sol, Fable 5.1 leads on every row shown, though the table is vendor-selected and we have seen no independent replication yet.
The pattern matters more than any single number. Gains on knowledge tests like Humanity's Last Exam are modest, up from 57.8 to 60.9 percent without tools, while gains on multi-step tool use and computer use are large. Customers describe the same thing. Ben Lafferty, Senior Staff Engineer at Shopify, says the model is more comfortable with long, unattended work, keeps records, reprioritizes as things change, and picks up where it left off. Dwight Temple at Ramp describes a single unattended 38-hour run that diagnosed a prior result as a label artifact, corrected it, and kicked off parallel experiments. For teams whose agent projects stall on reliability, that is the capability to test first.
The safeguards got quieter, which is what enterprises asked for
The Fable tier exists because Mythos-class capability ships with classifiers that flag cybersecurity, biology, and distillation requests, and the enterprise complaint about Fable 5 was false positives. Anthropic says it addressed that directly. Biology safeguards now produce 85 percent fewer false positives on elementary biology and medical queries, and cybersecurity safeguards trigger 60 percent fewer times per session than on Fable 5. For security teams, the model now permits vulnerability identification for defensive work, while penetration testing, exploit generation, and binary-based scanning still get redirected to the Opus models.
One new restriction deserves a flag to platform teams. As an anti-distillation measure, new API accounts can no longer manually edit prior Claude turns in multi-turn conversations. Any product that rewrites assistant history to steer behavior, a common pattern in agent frameworks and evaluation harnesses, needs checking against that rule. Anthropic also reports alignment improvements over Mythos 5, including lower reward-hacking rates, less motivated reasoning, and a reduced tendency to reach outside a test environment. Those claims are welcome for anyone letting a model run unattended for 38 hours, and they are exactly the kind of claim that deserves verification in your own sandbox.
Enterprise Frontier Safeguards move the logs into your cloud
The companion announcement is the one we would put on a board slide. Enterprise Frontier Safeguards, published the same day, pairs zero data retention with misuse detection by moving log storage into customer-controlled cloud infrastructure. Logs live in the customer's own AWS, Google Cloud, or Azure account under customer-managed encryption keys, review is fully automated with no Anthropic human in the loop, and the feature is free from Anthropic, with the cloud provider billing for storage. It is opt-in and changes nothing about model behavior, pricing, or rate limits. Rollout begins in fall 2026, with eligible customers getting zero data retention on Fable 5 and 5.1 in the meantime.
The participant list explains why this exists. Goldman Sachs, Morgan Stanley, Citi, Bank of America, Wells Fargo, Mastercard, Visa, Stripe, Comcast, KPMG, Salesforce, Snowflake, Cognition, and Factory are named, alongside the Analysis and Resilience Center for Systemic Risk. That is the regulated buyer base that has been asking for frontier capability without handing conversation logs to a model vendor. For a PE-backed SaaS company selling into those same banks, the practical implication is that your customers will soon expect the same posture from you. Read the EFS design as the template for what your own data-handling story will need to look like by next year's renewal cycle.
What this means for your model roadmap
The launch page carries more than twenty customer quotes, uniformly positive because Anthropic chose them, and a few still carry signal. Walden Yan, Co-founder and CPO of Cognition, says Devin is moving its Opus 5 traffic to Fable 5.1 on launch day because it matched or edged out Fable 5 at lower cost. Daniel Shan, Staff Engineer at Datadog, calls it a leading model for incident investigation with stronger reasoning than Opus 5 on complex production incidents. The quotes converge on three testable claims: fewer tokens per task, better long unattended runs, and better document grounding. Each is measurable in a week with your own traces.
For most enterprises the decision is about routing. Keep Opus 5 as the default for bulk production traffic at 5 and 25 dollars, and send the long-horizon agent work, incident investigations, and document-heavy analysis where Opus falls short to Fable 5.1, where the cache pricing now makes that escalation cheaper than it was in June. The strategic point is that Anthropic is shipping Mythos-class capability on a roughly quarterly cadence with a stable price and a maturing governance layer around it. Teams stuck in pilot because legal could not sign off on log retention now have a concrete path. The model is the headline. The zero data retention program and the cache price are the reasons the roadmap changes.
