Claude Fable 5.1, released by Anthropic on September 1, 2026, is the successor to Claude Fable 5 and Anthropic's most capable generally available model. List prices are unchanged ($10 per million input tokens, $50 per million output), but cache reads cost 75% less, which Anthropic estimates cuts real-world cost by about 25% on typical workloads and up to roughly 45% on heavily agentic ones. The gains concentrate in long-running work: multi-hour coding, research that follows its own leads, and analysis that ends as a finished document, spreadsheet, or deck. It is available today to Pro, Max, Team, and Enterprise users and on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. The headline for business teams is not the benchmarks; it is that frontier-class reasoning got cheaper to run continuously, and the governance posture around it changed at the same time.
Editor's note: this article was drafted with Claude Fable 5.1 on its release day and reviewed by the Vantage Point team. Every product claim is sourced to Anthropic's announcement and platform documentation, linked under Sources.
Model releases arrive often enough that most business leaders have learned to ignore them, and that is usually the right call. Claude Fable 5.1 deserves twenty minutes for a narrower reason: it changes the economics of leaving a frontier model running for hours, and it arrives with governance changes your security, legal, and compliance teams need to hear about whether or not anyone asks for the upgrade. This guide covers what changed, who is affected, what to do this week, and the timeline, in that order.
Claude Fable 5.1 is Anthropic's successor to Claude Fable 5, released September 1, 2026, and built for what Anthropic calls "ambitious, long-running, asynchronous work": tasks that take hours, span several applications, and end in a deliverable rather than a chat answer.
Anthropic released two models at once. Claude Fable 5.1 is generally available. Claude Mythos 5.1 is the identical model with more permissive safeguards, offered only through trusted access programs for vetted cybersecurity and life-sciences organizations and currently limited to a set of US organizations. If your firm is not in one of those programs, Fable 5.1 is the model that matters.
| Specification | Claude Fable 5.1 |
|---|---|
| Context window | 1M tokens (default and maximum), standard pricing across the whole window |
| Maximum output | 128K tokens |
| Thinking | Adaptive, always on; depth controlled by an effort setting |
| Default effort | High on the API and in Claude Code; Medium in Claude Cowork and on Claude.ai |
| Knowledge cutoff | June 2026 |
| Where it runs | Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS |
| Who gets it | Pro, Max, Team, and Enterprise plans; all Claude API customers |
| Retirement | Not sooner than September 1, 2027 |
One detail that spares your finance team a surprise: the tokenizer is unchanged from Fable 5, so the same text still produces roughly 30% more tokens than on models older than Claude Opus 4.7. Adjust for that before comparing bills against an older baseline.
Anthropic's documentation names six areas where Fable 5.1 pulls ahead, with the gap widest at higher effort levels: agentic coding over long sessions (multi-file features, large refactors, debugging, and review across hours); knowledge work with documents, spreadsheets, and slides, taking an analysis from first question to finished deliverable; research and search that follows up on what it finds; vision, including dense charts, filings, and tables nested in PDFs; long-context reasoning across the full 1M window; and computer use, operating browsers and desktop applications and recovering from failed steps.
The published benchmarks support that description. A selection, as reported by Anthropic with production safeguards enabled:
| Benchmark (Anthropic-reported) | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 55.8% | 42.0% | 52.3% | 37.3% |
| CursorBench 3.2.0 (agentic coding) | 73.4% | 70.5% | 70.0% | 67.2% |
| Terminal-Bench-Science 0.1 (agentic research) | 52.6% | 24.7% | 29.0% | 22.4% |
| GDPval-AA v2 (knowledge work) | 1853 | 1723 | 1824 | 1711 |
| AutomationBench (business workflows) | 31.4% | 17.1% | 26.9% | 19.6% |
| OSWorld 2.0, strict (computer use) | 41.7% | 36.1% | 39.6% | — |
| Humanity's Last Exam, no tools | 60.9% | 57.8% | 56.6% | — |
| Humanity's Last Exam, with tools | 65.0% | 63.8% | 63.6% | — |
Three footnotes from Anthropic's page belong next to that table. The scores were produced with Fable 5.1's production safeguards enabled; where a safeguard intervened, cyber tasks were completed by Claude Opus 4.8 and biology tasks by Claude Opus 5, which Anthropic says likely lowers Fable 5.1's own numbers. OSWorld 2.0 was run on the benchmark authors' August 2026 task release, so no competitor score is shown. And Claude Mythos 5.1 scored 60.9% on Terminal-Bench 4.0 against Fable 5.1's 55.8%; Anthropic attributes the gap to tasks where its earlier, less precise cyber safeguards intervened and expects it to narrow with the safeguard changes shipping today.
Two readings matter more than the numbers. First, the business-workflow and computer-use gains are the ones most likely to show up in an operations team's day, because they measure the model finishing multi-step tasks inside real software. Second, Anthropic notes that at Low or Medium effort, Fable 5.1 lands at or above Fable 5's results at much lower cost, so the upgrade path for an existing Fable 5 workload is often "switch and turn the effort down," not "switch and pay more."
Anthropic's announcement quotes twenty-two early-access partners, and the mix is worth noticing: trading firms and finance-focused AI vendors make up a large share of the list. These are Anthropic's published customer claims, not independent tests, but they describe the model in operational terms that benchmarks cannot.
Two more speak to unattended work. Browserbase reports Fable 5.1 completing 82% of tasks on its hardest browser-agent benchmark against 74% for Opus 5 and 57% for Fable 5, using fewer tokens than either, and says it "never crossed a critical stop point across hundreds of measured tasks." Ramp describes a single unattended 38-hour run that diagnosed a prior result as a label artifact, corrected it, launched six parallel experiments overnight, and returned with a result and next steps. The pattern across the quotes is consistent: the improvement is less about a smarter single answer than about the model staying coherent, verified, and within bounds over a long job.
A large part of Anthropic's announcement is about science, and it is tempting to skip. Claude Mythos 5.1, given open-source protein design and folding tools, designed binders whose affinities on three targets were ten times higher than the best entries in Adaptyv Bio's protein design competitions, with a hit rate Anthropic describes as nearly 50% in external lab validation. Claude Fable 5.1 trained a neural network on NASA Magellan radar data more than 30 years old to produce a new elevation map of a third of Venus, resolving features of two to three kilometers rather than ten to twenty, which Anthropic has released under a Creative Commons license. And Mythos 5.1 wrote custom GPU kernels that sped up seven open-source biology models by up to 2.5 times with identical outputs, cutting estimated GPU costs on genome-wide analyses by 30% to 60%, work Anthropic says would normally take a team of performance engineers weeks.
None of that is a CRM use case, but the shape of the work is. Each example is a long, unattended job with a verification loop: design, test, correct, repeat, report. That is the same shape as a data-migration validation that has to reconcile every record, an integration mapping that has to trace a field from an inbound API call down to a table and row (Plaid describes Fable 5.1 doing exactly that across more than eight services and three codebases), or a month-end analysis that has to end as a finished deck. The research section is Anthropic's evidence that the model can hold that shape for hours, and it is why the cache-read price cut matters more than the unchanged list price.
Input and output list prices are unchanged from Fable 5: $10 per million input tokens and $50 per million output tokens, with the Batch API at a 50% discount. The change is in cache reads, where the model re-uses context it has already processed instead of paying to read it again. Cache reads now cost $0.25 per million tokens, 75% less than Fable 5.
That is a bigger deal than it sounds. In a long agentic session the model re-reads its accumulated context (your documents, tool results, prior turns) at every step, so on tool-heavy work cache reads make up most of the bill. Measuring four weeks of actual August 2026 usage at default effort, Anthropic reports the same workloads cost about 25% less for typical usage and up to about 45% less for highly agentic usage. A team using Claude for short drafting will see far less of that saving than one running unattended research or multi-hour build sessions. Two smaller notes: US-only inference is available at 1.1x pricing, and the cache-read change applies wherever usage is billed by token, such as the API and cloud marketplaces, not to plan-priced seats.
If your team is on a Pro, Max, Team, or Enterprise plan, Fable 5.1 is available now. Anthropic describes it as built for backlog work in Cowork, requests picked up in Slack through Claude Tag (beta), browser operation, and unattended runs as a managed agent. The change most users will notice is behavioral: Fable 5.1 writes fewer progress updates during long tool runs, uses less bold-and-bullet formatting in chat, and writes denser prose in places. If a workflow depended on the model narrating its progress, ask for it explicitly.
This is where the release carries real work. Anthropic's what's-new documentation lists three breaking changes for anyone already calling Fable 5:
tool_choice to any or a named tool returns an error. Use auto with strict tool use or structured outputs, and tell the model in the prompt when a tool applies.Anthropic also documents seven behavior shifts that appear with no code change, including more variable parallel tool calling, answering from memory more often at Low effort, unmarked quotations when summarizing, and rewriting whole files for small edits, each with a documented prompting fix. The migration path: update the model ID to claude-fable-5-1, remove forced tool_choice, keep history append-only, re-tune effort from the default, and re-run your evals.
| Date | What happens |
|---|---|
| August 31, 2026 | Append-only history check enforced for newly created API accounts |
| September 1, 2026 | Claude Fable 5.1 released on all platforms; Mythos 5.1 access begins through trusted access programs, with the first Life Sciences Verification Program participants enrolled and Claude Security now running on Mythos 5.1 |
| "Near future" and "over time" (Anthropic's phrasing) | Cyber Verification Program adds access to Mythos-class models; the EU AI Act watermark detection API expands beyond private preview |
| Fall 2026 | Enterprise Frontier Safeguards begin phased rollout; eligible customers use zero data retention until then |
| Future model releases | Anti-distillation history-edit restriction extends to all accounts; Anthropic says it is working to bring many of Fable 5.1's improvements to the rest of the Claude model family |
| Not before September 1, 2027 | Earliest possible retirement date for Fable 5.1 |
If you followed the Fable 5 and Mythos 5 export-control suspension earlier this year: Anthropic's Mythos 5.1 access language is explicitly US-organizations-first, with expansion described as in progress. Nothing in the release limits Fable 5.1's general availability.
Vantage Point is a Claude Partner Network Member and an implementation partner for Salesforce and HubSpot, which puts us on both sides of the question most clients are asking this week: what should the model do for us, and what has to be true in our CRM, our data, and our policies before it can? We help mid-market teams choose the workloads worth running on a frontier model, configure Claude Team and Enterprise deployments, wire Claude into Salesforce and HubSpot with the permissions and governance those systems need, and write the two paragraphs of policy nobody wants to write. Senior consultants only — no junior handoffs; the experts you meet are the experts who deliver. If your access was reshaped by the July Fable 5 plan changes, this release is a good moment to revisit which tier your team actually needs.
The teams that get value from a model like this pick the right two workloads, set effort on purpose, and have their governance answer ready before the first unattended run. Vantage Point's senior consultants can help you do all three in a single working session. Talk to our team about a Claude readiness review, or explore our Claude services to see how we support Salesforce and HubSpot teams adopting Claude.
A new model release with its own model ID (claude-fable-5-1), its own retirement date (not before September 1, 2027), and breaking API changes. It shares Fable 5's tokenizer, list prices, and always-on adaptive thinking, so for business users it behaves like a stronger, cheaper-to-run Fable 5. For developers it is a migration, not a swap.
They are the same underlying model with different safeguards. Fable 5.1 is generally available with production safeguards on. Mythos 5.1 offers more permissive safeguards for vetted cyberdefense and life-sciences organizations: the Life Sciences Verification Program, run with the US government, has enrolled its first participants, and Anthropic says the Cyber Verification Program will add Mythos-class models in the near future. Access is currently limited to a set of US organizations. Anthropic's Claude Security product, which scans codebases for vulnerabilities, now runs on Mythos 5.1. On Terminal-Bench 4.0, Mythos 5.1 scored 60.9% to Fable 5.1's 55.8%; Anthropic attributes the gap to earlier cyber-safeguard interventions and expects it to narrow.
Input and output list prices are identical at $10 and $50 per million tokens. Cache reads fell 75% to $0.25 per million tokens, which Anthropic estimates cuts total cost by about 25% on typical workloads and up to about 45% on heavily agentic ones. Short, one-shot tasks see little of that saving; long, tool-heavy sessions see most of it.
Not by default. Anthropic recommends starting most workloads on Opus 5 and reserving Fable 5.1 for demanding reasoning and long-horizon agentic work, or where your evaluations show Opus 5 falling short. Fable 5.1 lists at twice Opus 5's per-token price, so the case for it rests on tasks where finishing a multi-hour job correctly is worth more than the difference. Some early-access partners are moving anyway: Cognition says it is shifting its Opus 5 traffic in Devin to Fable 5.1 on launch day because the cache-read pricing makes it economical, and Every reports Fable 5.1 running about twice as fast as Opus 5 on half the tokens. Both are Anthropic-published customer claims; the right answer for your team comes from your own evaluation set.
Probably some of them. Anthropic documents behavior changes that appear with no code change: fewer parallel tool calls in agent loops, fewer progress updates, denser prose, less chat formatting, unmarked quotations in summaries, and whole-file rewrites for small edits. Each has a short documented prompting fix. Re-run your evaluations after migrating rather than assuming parity.
It does not change the meaning, quality, or readability of the output and adds no visible characters. It is a statistical signal detectable only through Anthropic's detection API, currently in private preview. If your firm publishes AI-assisted content, the practical step is a disclosure line in your AI use policy, not a change to your editing process.
EFS gives enterprise customers the privacy of zero data retention while still detecting misuse: data stays on the customer's own cloud infrastructure, and human review is done by the customer by default. Anthropic says it was developed with more than 100 customers, including financial services firms, and it rolls out in phases starting fall 2026. Eligible customers can use Fable 5.1 under zero data retention until their EFS phase arrives.
Safer than its predecessor, with limits Anthropic states itself. The system card reports refusal of malicious agentic requests at a rate comparable to Opus 5 and Sonnet 5, Anthropic's best result to date on an external prompt-injection benchmark, and lower rates of reaching outside its environment or ignoring explicit constraints than Mythos 5. Anthropic also reports that the model can still sometimes bypass approvals and auto-mode classifiers, and that its audit sees less of very long-context and multi-agent work. Run unattended jobs with a human approval step on consequential actions, scoped permissions, and full logging.
Vantage Point is a boutique CRM consulting firm helping businesses transform with Salesforce, HubSpot, and AI — 150+ clients, 400+ engagements, and a 4.71/5 average engagement rating. Learn more at vantagepoint.io.