Fable 5.1 puts Anthropic back on top in the agent race

Same headline price, very different economics

Anthropic’s launch of Claude Fable 5.1 is best understood as a bet on the next unit of competition in AI: not the cheapest token, but the cheapest completed long task. The company introduced Claude Fable 5.1 and Claude Mythos 5.1 on September 1, presenting them as its most advanced models for coding and knowledge work, with Fable broadly available and Mythos reserved for trusted-access programs with looser safeguards in sensitive domains [1].

The striking part is that Anthropic did not lower the base list price. Fable 5.1 still costs $10 per million input tokens and $50 per million output tokens, matching Fable 5 and remaining well above cheaper Claude tiers such as Opus 5, Sonnet 5 and Haiku 4.5 [2]. On a conventional price table, that looks conservative. In actual agent workloads, however, the decisive change is the cache-read line: rereading cached context now costs $0.25 per million tokens, down from $1 for Fable 5 [1].

That 75% reduction is narrower than it first sounds. It applies to cached reads, not to fresh input, output, cache writes or subscription seat prices. Anthropic says the change should reduce typical Fable workloads by around 25% and complex coding or highly agentic workloads by up to around 45%, because those jobs repeatedly reuse large prompts, tool definitions, code context and prior work history [1]. TokenCost’s pricing analysis makes the arithmetic explicit: since only one line item changed, a customer’s total saving is roughly 75% of whatever share of the old bill came from cache reads [3].

That matters because agents are repetitive by design. A coding agent that checks a repository, opens files, runs tests, edits code, inspects errors and tries again may send the same project instructions and accumulated context many times. A research agent may keep rereading the same dossier while adding new evidence. Fable 5.1 does not make every token cheap; it makes persistent memory-like reuse less punishing. That is why the model can be simultaneously expensive on the rate card and more competitive on the cost of a completed enterprise task.

The benchmark story: long work, tools and recovery

Anthropic’s own benchmark table shows why the company believes the new economics are enough to move Fable back to the top. On Terminal-Bench-Science 0.1, Fable 5.1 scored 52.6%, compared with 24.7% for Fable 5, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol, according to Anthropic’s launch materials [1]. On Terminal-Bench 4.0, Fable 5.1 reached 55.8%, while Mythos 5.1 reached 60.9% under its different safeguard regime; Fable 5 scored 42.0% and Opus 5 scored 52.3% [1].

Those are company-reported results, and they should be read with the usual caution. Anthropic notes that Fable 5.1 was evaluated with production safeguards enabled, and that some cyber and biology interventions routed tasks to other Claude models or scored as zero on certain evaluations [1]. Still, the pattern across the cited benchmarks is consistent with the product message: the upgrade is strongest where the model must sustain a plan, use tools, work across a large context and recover from failed steps.

The public positioning also reinforces that interpretation. Anthropic’s platform documentation describes Fable 5.1 as a model for “demanding reasoning and long-horizon agentic work,” while still advising most users to start with Opus 5 and move up only when evaluations show that Opus is not enough [2]. That is a revealing recommendation. Anthropic is not pretending Fable 5.1 is the everyday default. It is reserving the model for the jobs where a higher per-token price can be offset by better success rates, fewer retries and cheaper cached context.

Early customer examples in Anthropic’s announcement are chosen to highlight this same frontier. Millennium said Fable 5.1 found the cause of a rare crash that had eluded previous models and engineers for years; MongoDB described a multi-day prototype built after the model researched code and documentation; Ramp described an unattended 38-hour machine-learning run that revisited a conclusion, launched experiments and returned recommendations [1]. These are not chat-demo claims. They are stories about AI systems becoming long-running coworkers inside software, finance, infrastructure and enterprise workflows.

Why cached context is the lever

The cache-read cut is not just a discount; it changes what is rational to build. When rereading a large context is expensive, developers shorten prompts, summarize aggressively, truncate logs or split work into smaller sessions. Those tactics save money but can make agents forget constraints, miss dependencies or redo analysis. A cheaper cached prefix encourages a different architecture: keep more stable context warm, let the model revisit it, and pay mainly for genuinely new input and final output.

Anthropic’s docs say Fable 5.1 and Mythos 5.1 keep a 1 million token context window and 128,000 token maximum output, while using the same $10/$50 input-output base pricing and a $0.25 cache-read rate [2]. That combination is aimed squarely at agent loops that hold a large project state. The company’s own cost graph says its August usage analysis found larger savings for context-heavy, tool-heavy work where cache reads make up most of the cost [1].

The caveat is that the bill will still be dominated by output-heavy work. A model that writes long reports, rewrites whole files or generates large artifacts will keep paying $50 per million output tokens. TokenCost notes that the discount ceiling is 75% only for a hypothetical workload made entirely of cache reads; ordinary blends save much less when output and fresh input dominate [3]. In other words, Fable 5.1 is cheaper for the type of agent Anthropic wants enterprises to run, not cheaper in every use case.

A stronger model with a stricter contract

The upgrade also comes with integration consequences. Anthropic’s documentation says teams moving from Fable 5 cannot simply swap the model ID and assume their agent runtime will behave the same [2]. Fable 5.1 does not support forced tool use via tool_choice set to any or a named tool, and such requests return a 400 error [2]. Earlier Claude models also cannot read Fable 5.1 thinking blocks, and editing earlier turns can invalidate those blocks in conversations that preserve reasoning state [2].

For enterprise buyers, that is a meaningful boundary. The model is designed to think across long sessions, but the infrastructure around it must treat conversation history, tool definitions and preserved reasoning as protocol state rather than disposable text. Anthropic added beta features such as per-message effort, turn-scoped system messages and user-visible progress updates between tool calls, but those features require engineering work to adopt [2].

Safety is the other constraint. Anthropic says Fable 5.1’s safeguards route many flagged cybersecurity and biology requests to Opus models, while Mythos 5.1 provides the same underlying capabilities only to vetted cyberdefenders and life-sciences users through trusted-access programs [1]. The company also says it reduced false positives: cyber safeguard interventions in Claude Code should fall by about 60% per session compared with Fable 5, and newer biology safeguards fire 85% less often on benign elementary biology and medical requests than the original Fable 5 launch safeguards [1].

Axios framed the release as a bundle of new models, cost cuts, safeguards, cybersecurity changes and enterprise IP protections, while noting that Anthropic is moving in a competitive environment where frontier releases are closely watched [4]. That is the broader point: Fable 5.1 is not merely a better model. It is a platform move that tries to make autonomous work safer, cheaper to meter and easier to sell to regulated enterprises.

The new top tier

Fable 5.1 puts Anthropic back at the summit not because every number is lower, but because the company attacked the economic bottleneck that matters for long-running agents. The base token price is still high. The context window is not meaningfully larger. The migration path has breaking changes. The safety layer still shapes what the generally available model can do.

But for enterprises evaluating AI agents that can inspect a codebase, operate tools, maintain state, test their own work and keep going for hours, the equation has changed. Fable 5.1 offers stronger benchmark performance, a cheaper way to reuse context and a clearer product split between broad commercial access and restricted high-risk capability access. If the next phase of AI competition is measured by cost per completed autonomous task, Anthropic has made Fable the model to beat again.

Sources from the last 72 hours

1. [1]Introducing Claude Fable 5.1 and Claude Mythos 5.1 \ AnthropicSep 1, 2026, 12:00 AM UTC
2. [2]Claude Fable 5.1 - Claude Platform DocsSep 1, 2026, 12:00 AM UTC
3. [3]Claude Fable 5.1 Pricing: The 75% Cache Read Cut | TokenCostSep 2, 2026, 12:00 AM UTC
4. [4]Anthropic releases new models, cuts agent costsSep 1, 2026, 7:55 PM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.