Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, the first update to its frontier-tier model since the original Fable 5 launched in June.
The headline change is pricing, not performance. Cache reads on Fable 5.1 and Mythos 5.1 cost $0.25 per million tokens, down from $1.00 on the previous generation. Base input and output prices remain unchanged at $10.00 and $50.00 per million tokens respectively. The 75 percent cut to cache reads means that agentic workloads with long-running system prompts and repeated context see the biggest savings.
Anthropic said the cost reduction comes from architectural changes that improve cache hit rates and reduce the computational overhead of maintaining cached context. The company claimed 25 to 45 percent total cost savings on typical agentic workloads, with the exact number depending on how much of the prompt is cached versus fresh input. For applications that send the same system prompt on every request, the savings are closer to the upper end of that range.
Benchmarks and Performance
Fable 5.1 scored higher than Fable 5 across every published benchmark, though Anthropic did not disclose specific numbers for most of them. The model supports a 1 million token context window and up to 128,000 tokens of output, matching the specifications of its predecessor.
The new model also introduces adaptive reasoning with configurable effort levels. Users can set the model to low, medium, high, or extra-high effort depending on the task. Lower effort settings produce faster responses at the cost of accuracy, while higher settings take longer but produce more careful reasoning. Artificial Analysis, an independent benchmarking firm, measured Fable 5.1 at between 58 and 65 on its intelligence scale depending on the effort setting. At the highest effort level, the model produced 69 tokens per second, slightly slower than the 56 tokens per second at high effort.
Mythos 5.1 is the same underlying model as Fable 5.1 but with a restricted access regime. It is available only to approved organizations working in cybersecurity and life sciences, where Anthropic has reduced some safety filters to allow more capable offensive and defensive analysis. The pricing is identical to Fable 5.1. Organizations must apply for access and demonstrate a legitimate use case in one of the two approved domains. Anthropic said Mythos models are subject to additional monitoring and usage restrictions even after approval.
Forced Tool Use Breaking Change
One change that may trip up existing users is a modification to forced tool use. In Fable 5.1, the model will no longer automatically invoke tools when they are defined in the API request. Instead, tools must be explicitly requested through a new parameter. Anthropic said the change gives developers more control over when and how the model uses external tools, but it requires updates to existing code that relied on the previous behavior.
The change is backward compatible in the sense that Fable 5 continues to work as before. Anthropic classified Fable 5 under legacy models alongside Opus 4.8 and Sonnet 4.6, with a retirement commitment of no sooner than September 2027. Existing users can continue running Fable 5 at its current pricing while they update their tool-use logic. Anthropic encouraged developers to test their applications against Fable 5.1 before migrating, given the tool-use behavioral change.
Cost Comparison
The cache read pricing puts Fable 5.1 in an interesting position relative to competitors. OpenAI GPT-6 Astra, released the same week, costs $15.00 per million input tokens and $60.00 for output. Google Gemini 3.8 Flash, also launched recently, is significantly cheaper on base pricing but offers less context and lower benchmark scores on complex reasoning tasks.
The real comparison is on total cost of ownership for production workloads. For an agentic system that processes 200,000 tokens of system prompt and runs 40 queries per session, Fable 5 costs roughly $15.80 per session. Fable 5.1 costs roughly $9.80, a 38 percent reduction driven almost entirely by the cheaper cache reads. At that volume, the savings compound quickly across thousands of daily sessions.
Anthropic positioned the pricing change as a response to customer feedback. Many enterprise users had complained that cache costs made agentic deployments prohibitively expensive at scale. By cutting cache reads by 75 percent, Anthropic is making a clear play for the production agentic market, where models run continuously and context reuse is the norm rather than the exception.
The timing matters too. OpenAI GPT-6 Astra launched on September 3 with strong cybersecurity capabilities but at a higher price point. Google Gemini 3.8 Flash targets a different segment with lower pricing and faster inference. Anthropic response with Fable 5.1 is to compete on total cost for long-context, tool-using workloads rather than on raw benchmark scores or output speed. That is the segment where cache pricing matters most, and where Anthropic says most of its enterprise revenue comes from.
The competitive landscape for AI models is shifting from pure capability comparisons to production economics. Enterprise buyers care less about which model tops a leaderboard and more about total cost per task at scale. Anthropic bet with Fable 5.1 is that cutting cache costs by 75 percent will matter more to production teams than incremental benchmark gains. Whether that bet pays off depends on whether the cost savings translate into real deployment decisions over the coming quarters. Anthropic has raised over $13 billion to date, and the pricing pressure from Fable 5.1 suggests the company is willing to sacrifice short-term margin to lock in enterprise customers.
discussion