Analysis
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 this week, with the headline change centered on cache-read pricing rather than a new headline per-token rate. Cache reads -- inputs the model has already processed and stored from earlier in a session -- now cost $0.25 per million tokens, a 75% reduction from Fable 5, VentureBeat reported. For typical workloads that cuts total cost by an estimated 25%; for highly agentic workloads -- coding agents, research agents, anything that repeatedly re-reads large amounts of cached context across a long session -- the savings can reach roughly 45%.
Claude Fable 5.1 is generally available at $10 per million input tokens and $50 per million output tokens. Claude Mythos 5.1 is described by Anthropic as the same underlying model with a different set of safeguards, available only through the company's trusted access programs designed specifically to support cybersecurity and life-sciences work -- a similar restricted-access structure to the one OpenAI is now using for Astra's advanced cyber capabilities. On Terminal-Bench-Science, a benchmark measuring scientific and technical agentic task performance, the new release scored 52.6%, per independent testing cited by MarkTechPost.
Why cache pricing, not headline pricing
The choice to cut cache-read costs rather than headline input/output pricing is a signal about where Anthropic sees its actual competitive pressure. Pulse covered Anthropic making its Sonnet 5 introductory pricing permanent rather than raising it as planned -- a defensive move against multi-model routing infrastructure that lets developers shift traffic to a cheaper competitor within days of a price change. Targeting cache-read costs specifically hits agentic and coding-assistant use cases hardest, precisely the workloads where developers run the same context through a model repeatedly and where competitors like OpenAI's GPT line and Google's Gemini models compete most directly on effective cost per completed task rather than sticker price per token.
The competitive landscape
Anthropic, OpenAI and Google have all been cutting or restructuring pricing this year rather than raising it, even as compute costs for training frontier models keep climbing -- a dynamic that shows up starkly in Anthropic's own roughly $90 billion of compute commitments signed in the past month. That combination -- rising infrastructure costs, falling or restructured customer-facing prices -- only works if inference costs are falling faster than headline pricing suggests, which is plausible given continued efficiency gains in serving infrastructure, but it does mean margin per query is under real pressure across the industry, not just at Anthropic.
What to watch next is whether OpenAI or Google respond with their own cache-pricing cuts specifically, rather than headline price moves -- that would confirm agentic workload cost is becoming the primary battleground for enterprise AI spend, not simple per-token pricing.
What the benchmark score adds
The 52.6% Terminal-Bench-Science result matters less as an absolute number than as a data point in a fast-moving benchmark race where labs leapfrog each other every few weeks on scientific and technical reasoning tasks. Combined with the pricing changes, it signals Anthropic is trying to win on both axes simultaneously -- capability and cost per completed task -- rather than picking one lane, a strategy that only works if the underlying model improvements are real rather than benchmark-optimized specifically for the headline number.
The Mythos 5.1 restricted-access structure is also worth noting on its own: pairing a capability upgrade with a narrower distribution model for the highest-risk use cases mirrors exactly the caution OpenAI is now applying to Astra, suggesting gated release for dual-use capabilities -- rather than broad simultaneous availability -- is becoming standard practice across frontier labs rather than a one-off policy choice at either company.
Additional reporting: TechCrunch.
