Anthropic Claude Fable 5.1 and Sonnet 5, Two Pillars of the Agent Era

Fable 5.1 vs Sonnet 5 on the Same Criteria

The table compares the two models’ prices, specs, and evaluation results on the same items. Prices and specs come from Anthropic’s official documentation; performance and evaluation cost come from Artificial Analysis (AA), an independent evaluator that measured both models the same way.

Price, context, and availability are from Anthropic’s documentation (checked October 6, 2026); the latency labels are from the models overview as of September 25, before Sonnet 5 moved to the legacy list. Index score, cost per task, output tokens, and output speed are from Artificial Analysis (checked October 6, 2026), with both models at maximum effort. Output speed varies from one measurement to the next.

The price list says 5x, yet in this evaluation a task cost $7.63 on Fable 5.1 and $5.09 on Sonnet 5, about 1.5x. Fable 5.1 produced roughly half as many output tokens, but input costs, caching, and run conditions also differ, so that alone does not explain the gap. For context, Fable 5.1 charges $0.25 per million tokens for cache reads and Sonnet 5 charges $0.20, almost the same, and Anthropic says cache reads make up most of the cost of agentic and coding work. One more wrinkle: Fable 5.1’s safeguards hand some cybersecurity and biology requests to other models, and AA’s score includes the answers those fallback models gave.

Sonnet 5.5, released on September 28, scrambled the picture again. It costs the same per token as Sonnet 5, yet scored 56 on the same AA index, ahead of Fable 5.1’s 53. It also used about 420M output tokens, so its cost per task came to $7.67, almost exactly Fable 5.1’s. Anthropic says Sonnet 5.5 costs up to 30% less per task than Sonnet 5 for most work, but AA’s maximum-effort run found it pricier than Sonnet 5 ($5.09). Change the effort setting and the cost ranking can flip. On Terminal-Bench 4.0, an agentic coding test, Anthropic’s Opus 5.5 and Sonnet 5.5 announcements report 55.8% for Fable 5.1, 10.3% for Sonnet 5, and 70.6% for Sonnet 5.5. Neither the token price nor a vendor’s chart can tell you what your own workload will cost.

Choosing between Fable and Sonnet is not just “hard tasks versus easy tasks.” Pick a few jobs you actually do and log output quality, response time, how often you had to re-run, and total cost. Then work upward from the cheapest option: see whether Sonnet 5.5 handles well-scoped everyday tasks, start complex work on Opus 5.5 as Anthropic recommends, and move to Fable 5.1 only when that still falls short. The real question is whether the pricier model earns its extra cost on your task.

Historical snapshot: Fable 5.1 reasoning evaluations

Anthropic positions Fable 5.1 for demanding reasoning and long-horizon agent work. In OpenAI’s GPT-6 Astra comparison table, which this article relied on when first written (Sep 22), it scored 65.7 on the Artificial Analysis Intelligence Index v4.1.1, ahead of GPT-6 Astra (61.2). On the FrontierCode 1.1 Extended coding test it scored 63.6%, close to GPT-6 Astra (64.5%).

Artificial Analysis has since moved to index v4.3.2, where Fable 5.1 scores 53 (checked October 6). Index versions use different test sets and scales, so scores from different versions should not be compared or mixed into one ranking.

Claude Sonnet 5, and Sonnet 5.5 Three Months Later

Anthropic released Sonnet 5 on June 30, 2026, calling it its most agentic Sonnet yet. According to the company, it can plan, use tools such as browsers and terminals, and run autonomously at a level that a few months earlier required larger and more expensive models, with performance close to Opus 4.8. It costs $2 per million input tokens and $10 per million output tokens, and at launch it became the default model on the Claude Free and Pro plans.

Three months later, on September 28, Sonnet 5.5 took over as the current Sonnet: same price, and more than 30% faster output, according to Anthropic. Anthropic’s docs now list Sonnet 5 as a legacy model, but it remains available on the API, with retirement no sooner than June 30, 2027. Code that runs on Sonnet 5 today will not suddenly break.

Anthropic’s Differentiators

• Engineering blog: Regularly publishes practical guides such as “Building Effective Agents” and “Contextual Retrieval.”
• MCP (Model Context Protocol): A standard way to plug AI into external tools and data. Anthropic open-sourced it in November 2024 and donated it to the Linux Foundation’s Agentic AI Foundation in December 2025, where it is now run as a vendor-neutral standard.
• Claude Code: A coding agent that works in the terminal and keeps widening Anthropic’s developer base.

Points to Consider

• Even in OpenAI’s own comparison table, Fable 5.1 and GPT-6 Astra trade wins: Fable 5.1 leads on Humanity’s Last Exam with tools (65.0% vs 57.2%) and trails slightly on Terminal-Bench 4.0 (55.8% vs 57.9%). For cybersecurity work, remember that Fable 5.1’s safeguards route many such requests to other models, such as Opus 4.8, in the first place. A single ranking can’t capture that.
• Opus 5.5 ($4 input, $20 output) arrived on September 22. Anthropic says it performs at the level of Fable 5.1 on most work, and its models overview suggests starting most workloads on Opus 5.5 and moving to Fable 5.1 only when higher effort settings still fall short. We compare its real-world costs in Claude Opus 5.5 vs GPT-6 Sol.