Anthropic's Fable 5.1 Boosts Coding, Cuts Cache Prices 75%

A New Flagship, Fast on the Heels of the Last One

Anthropic has released Claude Fable 5.1, positioning it as the company's most capable generally available model for coding, research and long-running, tool-using work. The launch arrived alongside Claude Mythos 5.1, a version built on identical underlying weights but with looser safety restrictions, reserved for vetted cybersecurity and life-sciences organizations through Anthropic's trusted-access programs 7817. Both are available immediately across Anthropic's own platform, the Claude API, and major cloud providers including AWS, Google Cloud and Microsoft Azure 716.

The timing is notable: Fable 5 only debuted in June, meaning Anthropic pushed out a major successor in roughly three months, underscoring how quickly frontier AI labs are now iterating on their top-tier models 10. Bloomberg and multiple regional outlets framed the release around Anthropic's own pitch — a model that is simultaneously stronger at programming and science tasks and less expensive to run, changes the company says respond directly to customer complaints about price, data retention and overly aggressive safety filters 1112.

What Fable 5.1 Is Actually Built For

Rather than targeting everyday chat, Fable 5.1 is aimed at work that unfolds over hours or days: sprawling software projects that span an entire codebase, code review, performance tuning, diagnosing root causes instead of patching symptoms, browser and computer-use tasks, and multistep scientific work such as experiment design and interpreting dense tables or diagrams 7816. Anthropic says the model is better at planning multi-step jobs, recovering when a step fails, checking its own output and keeping users updated as it works unsupervised 16.

A key feature carried over from prior releases is an adjustable “effort” setting that trades intelligence for speed and cost. Anthropic says Low or Medium effort can match Fable 5's results at a fraction of the cost, while Claude Code defaults to High effort and Claude.ai/Cowork default to Medium 79. Anthropic's own model-selection guidance, notably, still tells most users to start with the cheaper Opus 5 and only move up to Fable 5.1 when Opus falls short on demanding, long-horizon reasoning 9.

Benchmark Gains That Anthropic Says Are Substantial

Across Anthropic's published evaluations, Fable 5.1 shows sizable jumps over its predecessor. On Terminal-Bench 4.0, a measure of agentic coding, Fable 5.1 scored 55.8% versus 42.0% for Fable 5 and 52.3% for Opus 5; Mythos 5.1, running with fewer safeguard interventions, reached 60.9% 7810. On Terminal-Bench-Science 0.1, a scientific-research benchmark, Fable 5.1 more than doubled Fable 5's score, hitting 52.6% versus 24.7% 7810. It also posted gains on CursorBench, AutomationBench for business workflows, GDPval-AA v2 knowledge-work scoring, and Humanity's Last Exam 710. Artificial Analysis, which assisted in pre-release testing, separately reported that Fable 5.1 topped its Intelligence Index with a score of 66, ahead of Opus 5's 63 and GPT-5.6 Sol's 61 10.

TechSpot cautions that these are largely Anthropic's own benchmark results, not independent verification that the model outperforms rivals across every real-world use case 8. Anthropic itself notes that benchmarks were run with production safeguards active, and that safety interventions — which route flagged cybersecurity or biology tasks to other Claude models — likely dampened scores on evaluations like OSWorld and AutomationBench 78.

Early commercial users offered supporting anecdotes rather than controlled testing. Millennium said the model tracked down a software crash that had gone unexplained for four to five years by dissecting an external library and matching it against a core dump. Ramp reported running the model unattended for 38 hours on a machine-learning task, during which it revisited an earlier conclusion and launched six new experiments 816. Jane Street, Devin, MongoDB and Square offered similarly favorable testimonials about long-running autonomy and token efficiency, though these too are customer accounts rather than third-party benchmarks 16.

The Pricing Story Is More Complicated Than “Cheaper”

Anthropic's standard token prices for Fable 5.1 did not change: $10 per million input tokens and $50 per million output tokens, identical to Fable 5 and roughly double Opus 5's rates 71013. The headline savings instead come from cache reads — the cost of having the model reuse context it has already processed, such as a codebase, system instructions or prior conversation history. That price dropped 75%, from $1 to $0.25 per million tokens, while cache writes remain unchanged at $12.50 (five-minute cache) and $20 (one-hour cache) per million tokens 7816.

Because cached content can make up half or more of the tokens used in long, complex tasks, Anthropic estimates this translates into roughly 25% lower total costs for typical workloads and up to about 45% savings for heavily agentic tasks that lean on repeated context 7111214. Anthropic product lead Dianne Penn said the change was a direct response to customer feedback about cost 12. The company frames these figures as measured against actual usage patterns from August 2026 across Claude Enterprise, Claude Code and the API, not as a blanket reduction in list prices 7.

That nuance matters. The Decoder reports that Artificial Analysis, despite helping test the model pre-release, disputes the simple “cheaper” framing. At maximum effort, Fable 5.1 reportedly consumed about 1.7 times as many output tokens as Fable 5, meaning the cache-read discount — worth roughly $1.40 per agentic task — was outweighed by higher overall token usage, leaving Fable 5.1 costing around 20% more per task than Fable 5 at that setting. Artificial Analysis put Fable 5.1's cost at $3.76 per Intelligence Index task at max effort, compared with $2.34 for Opus 5, which trailed by only three points. At extra-high effort, Fable 5.1 reportedly closed the intelligence gap somewhat, scoring 65 at $2.72 per task, but still cost more than Opus 5 10.

Digital Trends and Firstpost largely echoed Anthropic's straightforward savings framing, emphasizing the 75% cache-read cut and the estimated 25–45% effective reduction, while also noting that Fable's standard rates remain pricier than Opus 5 or GPT-5.6 1314. Every, an early hands-on tester, reported that in an internal Slack-assistant comparison, Fable 5.1 used less than half as many tokens as Opus 5 for comparable results — an informal but suggestive data point on efficiency 15.

Safety Guardrails and the Mythos Split

Fable 5.1 and Mythos 5.1 share identical model weights but differ in safeguard strictness, continuing a pattern established with earlier Claude releases 71617. Anthropic says updated filtering produces 60% fewer false positives on cybersecurity queries and 85% fewer on benign biology-related questions compared with earlier versions, addressing a persistent complaint that safety systems were blocking legitimate research 10. Fable 5.1 can now identify software vulnerabilities for the first time, though exploit development remains off-limits and gets routed to Opus-tier models 1017. Mythos 5.1 access remains limited to U.S. organizations through Anthropic's Cyber Verification and Life Sciences Verification programs, with international expansion planned in coordination with the U.S. government 710.

The accompanying system card also flags a more serious note: Anthropic now assesses the risk of catastrophic misuse as “low” rather than “very low,” citing increased uncertainty tied to recent disclosures about model behavior in cybersecurity evaluations 17. Separately, Fable 5.1 and Mythos 5.1 are the first Claude models to include built-in watermarking, with a detection API being rolled out in private preview to regulators, journalists and researchers 10.

The Bigger Picture

The release lands amid intensifying competition in coding-focused AI, with Google reportedly preparing to launch a Gemini 3.8 Flash update this week touting “significantly upgraded coding capabilities” aimed at narrowing the gap with Anthropic and OpenAI 45. Anthropic's own documentation still steers most users toward the less expensive Opus 5 by default, reserving Fable 5.1 for cases where high-effort reasoning genuinely falls short 9.

Taken together, the coverage suggests Fable 5.1 is a genuine capability upgrade for long-horizon coding and research agents, but its economic pitch is conditional rather than universal. Buyers running short, simple prompts may see little benefit from the cache-price cut, while those running context-heavy, tool-heavy agents could see real savings — unless higher output-token consumption at maximum effort erodes those gains, as Artificial Analysis' testing suggests it can.