Anthropic has released Claude Fable 5.2 and Mythos 5.2, reclaiming the AI performance lead shortly before its initial public offering while pushing peak computational costs to an average of $26.50 per task.
On the Artificial Analysis Intelligence Index v5 at the highest compute tier, Fable 5.2 takes first place with 65 points, placing five points ahead of Google's unreleased Gemini 4 Carbon and six points ahead of Claude Opus 5.5. Behind these benchmark records lies an operational trade-off: securing top-tier performance costs roughly 2.6 times as much as Opus 5.5, which runs at $10 per task, and doubles the $12.99 average of Fable 5.1.
Anthropic establishes technical dominance ahead of its market debut by relying on extreme compute allocation, altering the unit economics of enterprise deployment. When complex technical queries scale across production systems, the financial outlay directly determines whether top-tier model integration remains viable. Enterprise deployment friction increases as organizations navigate strict budget allocations to support these resource-intensive models, requiring finance and engineering teams to carefully weigh the return on investment for routine operations against the steep costs of maximum compute tiers.
The Real Price of Peak Performance
Nominal token prices present an apparent discount. Anthropic priced both Claude Fable 5.2 and Mythos 5.2 at $8 per million input tokens and $40 per million output tokens, below the $10 and $50 rates set for their predecessors.
However, the gross volume of tokens required to complete an assignment offsets those nominal rate cuts. Per-task expenses at the highest compute tier climb to $26.50, demonstrating that raw architectural efficiency diverges from per-token billing, which directly impacts enterprise ROI when routine operations demand continuous high-tier inference.
Per-task costs at the highest tier average $26.50, about 2.6 times as much as Opus 5.5.
Deploying these systems at full capacity introduces significant operational overhead and budget allocation challenges. While the new models produce standard output up to 30 percent faster than Fable 5.1 and Mythos 5.1, Anthropic offers an optional fast mode for select organizations on platforms like Claude Code, the Claude Platform, Google Cloud, and Microsoft Foundry. That feature accelerates token delivery by up to a factor of 2.5, but doubles the total cost, intensifying the friction between desired execution speed and constrained operational budgets.
Benchmark Gains and Domain Restrictions
In domain evaluations, Fable 5.2 reaches 80 percent on HLE-Diamond's expert academic questions and 73 percent on Terminal-Bench-Science. On multi-step knowledge workflows evaluated by AA-Briefcase, the model logs 1,785 Elo, trailing Claude Sonnet 5.5 at 1,813 Elo. On Terminal-Bench 4.0, which tracks developer and security workflows, Fable 5.2 records 60 percent, matching Opus 5.5.
Access structure reinforces the commercial division between the releases. Claude Fable 5.2 is available to all users but includes active safeguards that block cybersecurity and biological requests, redirecting them to Opus 4.8 and Opus 5 respectively. Mythos 5.2 removes select restrictions exclusively for entities approved through Anthropic's Verification Programs for Cybersecurity or Life Sciences.
Mythos 5.2 found 10.2 percent more vulnerabilities than Opus 5.5 and 14.4 percent more than Mythos 5.1 in the October 2026 run with 1,829 CVEs.
These specialized capabilities target narrow institutional workflows rather than unconstrained developer access.
Audit production query logs across technical pipelines this week to measure actual token consumption per workflow, and restrict high-tier routing exclusively to multi-step tasks where Opus 5.5 fails to complete the execution.
What changed
Who actually pays for Anthropic securing the top spot on benchmark leaderboards ahead of its public debut? Dropping list prices to $8 per million input tokens proved to be a convenient front, as the system simply burns through them at volumes that double the final bill for complex jobs. While in April the Claude Mythos rollout raised alarms about business models being dismantled, the current friction is purely accounting. Token rates look cheaper on paper, yet a single completed task now runs up a $26.50 tab.
Our stories: Anthropic Claude Mythos: AI Cybersecurity Threat or PR? (April 11).