Fable 5.2 Is Live: What Anthropic’s Newest Model Release Means for You

Reading Time: 5 minutes

Anthropic released Fable 5.2 and Mythos 5.2, with Fable 5.2 now ranked first on Artificial Analysis and priced 20% lower than its predecessor at $8/$40 per million tokens. Simon Willison's analysis reveals a regression: the max reasoning level exhausts its token budget and returns no output, making xhigh the practical ceiling for now.

Anthropic Just Released Fable 5.2 and Mythos 5.2

Anthropic released two new models on 9 November 2026: Fable 5.2 and Mythos 5.2. The launch is notable for more than just a version bump. According to Anthropic, Fable 5.2 has “solved two open problems in the physics of black holes” — a signal that Anthropic’s internal benchmarks have grown considerably more ambitious since earlier Fable releases. On the external rankings front, Artificial Analysis currently lists Fable 5.2 as the top-ranked model across the models it tracks.

Simon Willison’s analysis at simonwillison.net documents both the release and his process of adding the new models to his open-source plugin llm-anthropic, which reached version 0.31 with this update. Willison’s notes are worth reading carefully because they surface both what improved and what broke — and the honest accounting of both is what makes the release interesting for professional users.

What Is llm-anthropic, and Why Does It Matter?

Before getting into the model specifics, a quick explanation for non-technical readers: llm is a command-line tool that lets you send prompts to AI models from your computer’s terminal. llm-anthropic is a plugin that connects that tool to Anthropic’s Claude family of models. Simon Willison maintains both.

This matters because it represents a category of access that is different from the Claude.ai chat interface you may already use. With llm, developers and technically curious professionals can automate prompts, pipe text from files directly into Claude, and run the same query across multiple models in one step — all without opening a browser.

For most non-technical professionals in India, this is not something you will set up yourself today. But understanding that this layer exists helps you appreciate why version numbers like llm-anthropic 0.31 generate real discussion among the developer community that ultimately shapes the tools you will use.

What Changed in Fable 5.2

A Meaningful Price Drop

The clearest change for anyone paying attention to costs is pricing. According to Simon Willison’s analysis, Fable 5.2 is priced at $8 per million input tokens and $40 per million output tokens. The previous version, Fable 5.1, was priced at $10 per million input and $50 per million output.

That works out to a 20% discount across both input and output, which is significant at scale. In INR terms (using 1 USD ≈ 85 INR), you are looking at roughly ₹680 per million input tokens and ₹3,400 per million output tokens for Fable 5.2, compared to ₹850 and ₹4,250 respectively for Fable 5.1.

For a fintech startup in Pune running automated document analysis on thousands of loan applications per month, that 20% reduction translates directly into lower API bills — without any change to the quality of output, assuming the model performs comparably or better.

Five Reasoning Levels

Fable 5.2 supports five reasoning effort levels: low, medium, high, xhigh, and max. This is a continued feature from the Fable series, not a new introduction, but it is worth understanding what it means in practice.

Reasoning levels allow you — or the application you are using — to tell the model how much computational thinking to apply before returning an answer. At low, the model responds quickly and cheaply. At max, it spends significantly more tokens working through the problem before replying. The tradeoff is speed and cost versus depth of analysis.

For a contract manager at a law firm in Delhi reviewing a complex vendor agreement, using a higher reasoning level might surface ambiguous clauses that a quicker pass would miss. For a marketing executive in Mumbai who just wants a subject line for a campaign email, low or medium is entirely sufficient.

The Bug That Matters: Max Reasoning Runs Out of Tokens

This is where Simon Willison’s analysis becomes genuinely useful, because he documents a real limitation rather than glossing over it.

Willison tested Fable 5.2 by asking it to generate SVG images of a pelican riding a bicycle across all five reasoning levels. The xhigh level produced the best result. The max level, however, failed to return an image at all — the model exhausted 128,000 tokens in the thinking process and never produced an output.

Willison notes that this is the same behaviour he observed with Opus 5.5, but it is a regression for the Fable series specifically. Both Fable 5 and Fable 5.1 handled max reasoning prompts for the same task without hitting this limit. Fable 5.2 introduced the problem.

What does this mean practically? If you are building or using an application that relies on max reasoning in Fable 5.2 for complex tasks, you may hit a wall where the model consumes its token budget thinking and returns nothing. This is not a theoretical edge case — it happened on a creative generation task, not an unusually demanding one.

The honest takeaway: xhigh appears to be the practical ceiling for Fable 5.2 in its current state. max is available as a setting but carries a meaningful risk of producing no output at all on complex requests.

A Concrete Indian Scenario: Research Analyst in Bengaluru

Imagine a senior research analyst at an asset management firm in Bengaluru. Her team regularly synthesises earnings call transcripts, quarterly reports, and regulatory filings to build investment memos. These documents are long, dense, and require careful reading — the kind of task where a higher reasoning level genuinely earns its cost.

With Fable 5.2 accessed through the API (either directly or through a tool her firm’s tech team has built), she could send a 50,000-word batch of filings and ask for a structured summary of risk disclosures across five companies. At high or xhigh reasoning, Fable 5.2 would think through the material carefully before responding.

At the new pricing of $8/million input tokens, processing 50,000 words (roughly 65,000 tokens) costs approximately $0.52 — about ₹44. That is a meaningful reduction from what the same query would have cost on Fable 5.1. Over a month of daily analyses, those savings accumulate.

She would want to avoid max reasoning for now, given the documented token-exhaustion bug. But xhigh delivers depth without the risk of a silent failure.

Limitations and Honest Tradeoffs

• The max reasoning bug is real. As documented by Simon Willison, max effort in Fable 5.2 can exhaust its token budget and return nothing. This is a regression from Fable 5 and Fable 5.1, and Anthropic has not issued a public fix as of this writing.
• API access requires a developer setup. The llm-anthropic plugin and the pricing discussed above apply to API usage, not the Claude.ai consumer interface. Non-technical users in India accessing Claude through the browser are not directly affected by or exposed to these settings.
• Pricing is in USD. Indian teams using the API pay in dollars, which means rupee costs fluctuate with the exchange rate. The ₹680/million input figure assumes 1 USD = 85 INR and will shift with currency movements.
• Regional availability. Anthropic has not announced any India-specific access restrictions for Fable 5.2, but API access from India is subject to the same terms-of-service and payment infrastructure requirements as any other market. Teams should verify their account setup independently.
• Mythos 5.2 details are sparse. The release includes Mythos 5.2 alongside Fable 5.2, but Simon Willison’s analysis focuses primarily on Fable 5.2. Pricing and capability specifics for Mythos 5.2 are not detailed in the source material.

What to Watch For Next

The token-exhaustion bug at max reasoning is the most actionable item to monitor. If Anthropic issues a patch or update to Fable 5.2 that restores max reasoning reliability — as it worked in Fable 5 and 5.1 — that will meaningfully expand what the model can do on complex, long-form tasks.

For non-technical professionals in India whose firms use Claude through third-party applications or the API, the 20% price reduction is already in effect and requires no action on your part. If your team is exploring API access for the first time, Fable 5.2’s lower price point makes it a more economical entry point than its predecessor was.

For now, the practical guidance from Simon Willison’s analysis is clear: use xhigh for demanding tasks, leave max alone until the bug is resolved, and treat the pricing improvement as a straightforward benefit of staying on the latest model version.