Fifty dollars per million output tokens. That's what an enterprise now pays to run Anthropic's strongest model, roughly three times what the same workload cost on Opus 4.8 a week ago. The question every head of AI should be asking this morning is not whether Claude Fable 5 is better. It obviously is. The question is whether the workflow you're about to point it at clears the new bar.
Anthropic released Claude Fable 5 and Mythos 5 on June 9. They share the same underlying weights but ship with different safeguard configurations: Fable 5 for general availability with safety fallbacks to Opus 4.8 in less than 5% of sessions, Mythos 5 with safeguards lifted and access gated to Project Glasswing partners and US government cyber defenders. Both are priced identically at $10 per million input tokens and $50 per million output tokens, which is double the input cost and roughly triple the output cost of Opus 4.8.
That's the new bar. The benchmarks justify the bar in some cases and not in others. Here is the math, the workflows that clear it, and the workflows that don't.
The benchmarks are real
Before the cost math, the capability math. Fable 5 and Mythos 5 are the new top of the leaderboard on every benchmark that matters for enterprise agent work.
On GDPval-AA, the enterprise knowledge-work benchmark, the new models score 1932 versus 1890 for Claude Opus 4.8, 1769 for GPT-5.5, and 1314 for Gemini 3.1 Pro. On GDPpdf, the visual document reasoning benchmark, they score 29.8% without tool use versus 22.5% for Opus 4.8, 24.9% for GPT-5.5, and 16.7% for Gemini 3.1 Pro. On SWE-bench Pro, the hard software engineering benchmark, they hit 80.3% against GPT-5.5's 58.6%. On Cognition's FrontierCode Diamond for maintainable agentic coding, 29.3% against Opus 4.8's 13.4% and GPT-5.5's 5.7%.
That last gap (29.3% vs 13.4%) is the largest single-release jump in agentic coding capability in two years. The leaderboard wins are real, they are large, and they are the reason the premium exists.
The raw math, per workflow
A typical agentic enterprise task runs around 50,000 input tokens (context, retrieved documents, prior turns) and 5,000 output tokens (the agent's response and tool calls). At Fable 5 prices, that's $0.50 for input plus $0.25 for output, around $0.75 per task. At Opus 4.8 prices (roughly $3 per million input, $15 per million output), the same task runs about $0.15 plus $0.075, or $0.23.
Per task, the gap looks small. At production volume, it doesn't.
A reconciliation agent running 10,000 tasks per day on Opus 4.8 costs about $2,300 per day. The same agent on Fable 5 costs about $7,500 per day. That's a $1.6 million annual delta on a single workflow before counting input growth, retries, or any of the long-context patterns that push token counts higher. Multiply across the five to ten production agent workflows a typical enterprise runs today, and the annual delta lands in eight figures fast.
The default answer to "should we move this workload to Fable 5" should not be yes. It should be: "Show me which workflows clear the new bar."
Where the 3x premium clears the bar
Three categories where the cost math actually pencils out.
Hard software engineering agents. This is the strongest case. SWE-bench Pro at 80.3% versus GPT-5.5's 58.6% means roughly 22 additional correct pull requests per 100 attempts. At a fully-loaded engineer rate of $90 per hour and two hours saved per merged PR, that's $3,960 of engineering time recovered per 100 Fable 5 runs against an incremental token cost of roughly $48. Cost-recovery ratio: about 80x. Enterprise teams running internal coding agents, code review automation, or migration tooling will see Fable 5 pay back inside a single sprint.
Visual document reasoning at decision-grade stakes. GDPpdf at 29.8% versus Opus 4.8's 22.5% is a 32% relative jump. That maps directly to contract review (where one missed clause costs more than a year of premium token spend), insurance claims triage (where misclassification compounds across thousands of claims), and financial filings analysis (where SEC compliance penalties dwarf any plausible model cost). The premium is rational anywhere a single agent decision touches material money or regulated outcomes.
Long-horizon planning and multi-step reasoning. Workflows where the agent has to plan, execute, and self-correct over many turns benefit disproportionately from frontier reasoning quality. A workflow with eight reasoning steps and 5% error per step on Opus 4.8 fails end-to-end about 34% of the time. The same workflow at 2% error per step (Fable 5's typical advantage on long-context benchmarks) fails 15% of the time. The cost of getting these workflows right is rounding error compared to the cost of getting them wrong.
Where the premium does not clear the bar
Three categories where moving to Fable 5 wastes money.
High-volume classification, routing, and triage. Workflows that sort, label, or route at scale. Most tickets, most emails, most document categorization. Opus 4.8 already hits production accuracy targets here. The GDPval-AA jump from 1890 to 1932 is a 2.2% relative lift on enterprise knowledge work, and a 2.2% accuracy lift does not justify 3x the per-token cost when each output is a one-sentence label worth pennies. Keep these on Opus 4.8, Haiku, or smaller open models.
Customer-facing chat and summarization. Volume is high, accuracy bars are met by mid-tier models, and the marginal value of each additional correct word is small. Premium models here are pure waste.
Workflows where the bottleneck isn't the model. Most production agent failures in enterprise environments are not reasoning failures. They are integration failures, data quality issues, prompt design issues, or governance gaps. Throwing a more expensive model at a non-model problem doesn't fix anything except your token bill. Diagnose the actual failure mode before reaching for Fable 5.
The 13-day window
From June 9 through June 22, Fable 5 is included free on Anthropic's Pro, Max, Team, and seat-based Enterprise plans. That's a 13-day window where enterprise teams can run real A/B comparisons against Opus 4.8 logs at zero incremental cost.
Use it. Pick the three workflows where you have the strongest hypothesis that frontier reasoning quality matters: the highest-stakes financial review, the most complex engineering automation, the contract or claims pipeline where errors compound. Reroute them to Fable 5 for the full trial period. Log every decision the agent makes alongside the decision it would have made on Opus 4.8. Score the differences. By June 22, you'll know which workflows to keep on Fable 5 at full price and which to roll back.
Skipping the trial means making the decision blind after the 23rd. That's the most expensive choice on offer.
Why audit-grade attribution decides whether the premium survives procurement
A 3x cost increase on any production line item is a procurement event. Finance teams will ask, reasonably, which specific Fable 5 calls produced which specific business outcomes. "We upgraded the model and accuracy went up 4%" is not a defensible answer for a $1.6 million annual line item. The defensible answer is: "These 14,000 Fable 5 calls in May processed contracts worth $2.4 billion. They flagged 412 risk clauses that the prior model missed, three of which would have cost more than the entire annual Fable 5 budget if they had shipped."
That answer requires per-step attribution: every model call logged with the workflow step it served, the input it consumed, the decision it produced, the human checkpoint it satisfied, and the business outcome that followed. Most enterprise agent stacks do not log at this granularity today. Production agent platforms that treat audit attribution as a first-class field — not an afterthought — do, which is the difference between defending the premium in next quarter's procurement review and having to roll it back.
The Fable 5 premium is not a model decision. It's a platform decision. The model is good enough to justify the premium on the right workflows. The platform is what proves which workflows those are.
What to do this week
Three concrete moves before June 22.
One. Pick the three highest-stakes workflows in your production agent stack. Reroute them to Fable 5 during the free-trial window. Log everything.
Two. Build (or confirm you have) per-step cost attribution. If you can't show finance which workflow ran which Fable 5 call this week, you can't show them why to keep paying for it next month.
Three. Decide your routing policy by June 22. Workflows that pay back stay on Fable 5. Workflows that don't roll back to Opus 4.8, or down further to Haiku or open models where appropriate. The platform should make this routing trivial to change per workflow without rewriting the agent.
After June 22, the question isn't whether Fable 5 is the best model on the leaderboard. It clearly is. The question is whether your production agent stack can prove, workflow by workflow, where the new bar is worth clearing. The teams that can will have the strongest reasoning model in production with finance signoff. The teams that can't will either underuse the capability or overpay for it.
Anthropic's $965 billion IPO filing last week was a vendor concentration moment for enterprise agent buyers. Fable 5 is the procurement moment that follows. Both decisions get easier when the platform layer underneath does the attribution work that justifies the spend.
