Claude Fable 5 launched on June 9, 2026, as the first Mythos-class model Anthropic has released publicly. It sits above the Opus tier, carries the label “most capable widely released model,” and costs exactly twice what Opus 4.8 costs. Opus 4.8 stays active, retains its own designation as the most capable Opus-tier model, and is not being deprecated.2
The pricing gap is clean: Fable 5 at $10 per million input tokens and $50 per million output tokens versus Opus 4.8 at $5/$25.1 That 2x multiplier is not a speculative early-access premium. It is the general-availability rate effective from day one, and from June 23, 2026, it applies even for subscription plan users who currently get Fable 5 access through their plan.1
The question is not whether Fable 5 is better. Anthropic says it is, and the benchmark placements support that claim. The question is whether the 2x cost is recoverable in your actual workload.
What Fable 5 Adds Over Opus 4.8
Both models share the same 1M-token context window, 128K maximum output, and the tokenizer introduced with Opus 4.7.2 On API surface, the relevant difference is one constraint: Fable 5 does not accept an explicit thinking: {type: "disabled"} parameter. Where Opus 4.8 allows disabling adaptive thinking, Fable 5 returns a 400 error on that call. Omit the parameter to get default behavior.2
The capability differences Anthropic reports are concentrated in three areas.
Agentic autonomy at long context. Fable 5 is described as working autonomously longer than previous Claude models, operating across millions of tokens in extended task runs.1 Opus 4.8 already offered strong long-horizon performance; Fable 5 extends the practical envelope further for tasks that require sustained multi-step reasoning without human checkpoints.
Frontier coding benchmarks. On FrontierCode, the Cognition benchmark for coding at medium effort, Fable 5 holds the highest score among frontier models.1 CursorBench results are described as state-of-the-art.1 On SWE-Bench Pro, Anthropic reports Fable 5 at 80.3%, versus Opus 4.8’s 69.2% and GPT-5.5’s 58.6% — a roughly 11-point lead over the next frontier model. [Updated June 2026] These SWE-Bench Pro figures are Anthropic’s vendor-reported scores using Anthropic’s own scaffolding; independent third-party replication on a neutral harness is still pending. FrontierCode and CursorBench rankings remain without published numeric scores.
Science and knowledge work. Fable 5 is the first model to break 90% on an unnamed core analytics benchmark.1 On a senior-level reasoning finance benchmark from Hebbia, it scores highest among tested models.1 Anthropic reports drug and protein design acceleration approximately 10x faster than prior workflows, and novel hypothesis generation in molecular biology.1 Stripe reports compressing months of engineering into days using Fable 5.1
On ViBench, Anthropic reports Fable 5 as the highest-performing model tested; the announcement does not detail what the benchmark measures.1
The Benchmark Caveat
[Updated June 2026] Anthropic has published numeric SWE-Bench Pro scores for Fable 5 (80.3%, versus Opus 4.8’s 69.2% and GPT-5.5’s 58.6%), but those figures use Anthropic’s own scaffolding rather than a neutral harness. For FrontierCode and CursorBench, Anthropic released rankings without numeric figures. For ViBench, Hebbia’s finance benchmark, and the core analytics benchmark, neither scores nor methodology have been published.
For buyers evaluating the upgrade, the evidence base for Fable 5’s coding advantage is now partially numeric — but the SWE-Bench Pro lead should be read as a vendor-reported figure until independent replication is available. Third-party evaluators on neutral harnesses have not yet produced a comprehensive Fable 5 leaderboard row. A deeper look at what FrontierCode, CursorBench, and ViBench actually measure is in Claude Fable 5 Benchmarks: What FrontierCode, CursorBench, and ViBench Show.
Use Cases Where 2x Cost Is Justified
Frontier research and scientific computing. The drug design, protein design, and molecular biology claims from Anthropic’s announcement are the clearest ROI case.1 If a researcher’s time costs $100 per hour and Fable 5 reduces a task from four hours to twenty minutes, the token cost difference between Fable 5 and Opus 4.8 is noise. The same arithmetic applies to complex financial modeling, where senior-level reasoning gaps translate directly to analyst time.
Long-context autonomous agents. An agent running across millions of tokens without human steering is paying substantial token costs regardless of which model it uses. The per-token price difference is real for these long runs, but if Fable 5 completes a task in one sustained run where Opus 4.8 requires two or three with human re-orientation, the extra per-token cost may be net lower. The efficiency gain on task completion, not raw token count, determines the cost comparison.
High-stakes coding at medium effort. FrontierCode and CursorBench placements indicate Fable 5’s coding ceiling is higher than Opus 4.8’s.1 For a team shipping production software where a missed bug costs significant engineering time to debug and patch, the quality delta may be worth the premium. This is not the same as general coding assistance.
ViBench tasks. Workloads in the categories ViBench covers, where Anthropic reports Fable 5 as the highest-performing model tested.1
Use Cases Where Opus 4.8 Is the Right Choice
Standard coding assistance and review. Opus 4.8 is still the “most capable Opus-tier model”.2 For everyday code generation, refactoring, documentation, and review, the capability difference is unlikely to close in 2x less time on enough tasks to make Fable 5 economical. Cursor, GitHub, and Replit are cited as Fable 5 customers,1 but their platform defaults will be calibrated against volume economics that individual developers do not share.
Chatbots, classification, and extraction. These workloads rarely bump into the ceiling that differentiates Fable 5. Paying double for tasks that both models complete equally well is straightforwardly wasteful.
Batch processing pipelines. Fable 5 supports the standard Batch API at $5/$25 per million tokens (half the synchronous rate). However, the 300K-output extended beta — documented for Opus 4.8, 4.7, 4.6, and Sonnet 4.6 — does not list Fable 5 as a supported model.2 For pipelines that depend on that extended output limit, Opus 4.8 remains the documented path.
Cost-constrained products. A 2x model-cost increase that flows through to per-request costs has direct unit economics implications. If your product’s margin depends on inference cost, Opus 4.8 is the more defensible default until you have user-level evidence that the quality difference retains customers or enables higher pricing.
The Mythos 5 Distinction
Fable 5’s API model ID is claude-fable-5. A companion model, Claude Mythos 5 (claude-mythos-5), uses the same underlying model but has certain safeguards lifted in specific areas.12 Mythos 5 carries the same $10/$50 pricing and shares the 1M context window, but its access is restricted to Project Glasswing partners (cyber defenders, infrastructure providers) and select biology researchers.2 There is no self-serve path to Mythos 5.
On Fable 5 itself, cybersecurity classifiers block offensive cyber tasks with zero compliance across 30 tested jailbreak techniques.1 Biology and chemistry classifiers apply broad protections; flagged prompts fall back to Opus 4.8.1 Fable 5 also ships with distillation protection against model capability extraction, though Anthropic has not published how that mechanism works.1 A first-principles analysis of what the feature can and cannot prevent is in Fable 5 Distillation Protection: How Anthropic Blocks Model Copying.
Pricing Context: The Full Current Lineup
For reference, the complete model pricing from Anthropic as of June 2026:2
- Fable 5 (
claude-fable-5): $10/$50 per million input/output tokens, 1M context, 128K output - Opus 4.8 (
claude-opus-4-8): $5/$25 per million input/output tokens, 1M context, 128K output - Sonnet 4.6 (
claude-sonnet-4-6): $3/$15 per million input/output tokens, 1M context, 64K output - Haiku 4.5 (
claude-haiku-4-5): $1/$5 per million input/output tokens, 200K context, 64K output
Fable 5 costs the same per token as Opus 4.8 fast mode.1 Users on Opus 4.8 fast mode are already paying $10/$50; migrating those workflows to Fable 5 standard mode costs the same per token and gains a more capable model.
The Open-Weight Alternative: GLM-5.2 [Updated June 2026]
The per-token framing above assumes you are buying inference from a cloud provider. GLM-5.2, released by Zhipu on June 13, 2026, offers a structurally different arrangement that changes how the Fable 5 vs. Opus 4.8 price calculation sits in context.6 The model’s weights are published under an MIT license at huggingface.co/zai-org/GLM-5.2 (BF16) and huggingface.co/zai-org/GLM-5.2-FP8, covering a 753B-parameter MoE architecture with approximately 40B active parameters (implied by the “744B-A40B” designation in Zhipu’s README). At 1M input tokens of context (the same ceiling as Fable 5 and Opus 4.8), the model uses IndexShare sparse attention that Zhipu reports reduces per-token FLOPs by 2.9x at that context length.6 On coding specifically, GLM-5.2 posts 62.1% on SWE-bench Pro; for comparison, Opus 4.8 is at 69.2% and Fable 5 at 80.3%.6 The capability gap is real, but the pricing structure is not comparable: self-hosting on your own hardware carries no per-token fee under the MIT license, and Zhipu’s Z.ai cloud subscription starts at $18/month (approximately 400 prompts per week) or roughly $12.60/month on annual billing. For teams whose primary concern is inference cost predictability rather than ceiling capability, GLM-5.2’s flat-rate model sits at a different point on the tradeoff space than either Fable 5 or Opus 4.8 occupies. The model also exposes an Anthropic Messages API-compatible endpoint, so tooling built for Claude (including Claude Code) can route to it with a base URL swap. See how Zhipu is positioning this release against Anthropic’s tightened access and the broader pattern of per-token pricing pressure on AI agent workloads.
The Export Control Suspension [Updated June 2026]
The directive cited a reported jailbreak of Fable 5’s guardrails as a national security concern, specifically the model’s cybersecurity capabilities. Anthropic disputes the characterization, describing the demonstrated technique as narrow and non-universal and noting that comparable results are achievable on other deployed models including GPT-5.5. Anthropic is working to restore access and met with US officials on June 15–16, but as of this writing no restoration timeline has been announced.3
This suspension has a direct effect on the pricing comparison this article covers: if Fable 5 is not available, the 2x pricing question is moot for any workload that cannot fall back to a third-party provider. Teams that began migrating workflows to Fable 5 between June 9 and June 12 now route to Opus 4.8 by default. The Stripe engineering case study (50-million-line Ruby migration) and other early customer work was completed during that three-day window; it cannot currently be reproduced.
The export control episode also provides involuntary data on the robustness of Anthropic’s cybersecurity classifiers. If a jailbreak technique can plausibly surface Mythos-grade behavior from the public Fable 5 model, it undermines the clean separation the classifier architecture was designed to provide. Anthropic’s response — contesting the scope of the concern while complying fully — suggests the company’s internal assessment of the risk differs substantially from the government’s. That gap will need to close before access is restored. See the full export order analysis for regulatory context.
For teams currently evaluating the Fable 5 vs. Opus 4.8 decision, Opus 4.8 is the only viable Claude option until the suspension lifts.
The Upgrade Decision
The 2x multiplier is not unusual for a new frontier tier. The question is whether your specific task mix generates enough quality-differential value to cover it.
Frontier research, long-running autonomous agents, complex scientific computing, and senior-level knowledge work sit in the clear-yes category. These are tasks where Anthropic’s evidence is strongest and where the cost-per-result framing consistently favors capability over per-token price.
Standard coding, chatbots, content generation, and classification sit in the clear-no category. Opus 4.8 is a strong model, not a fallback. For these tasks, the 2x premium is cost with no measurable return.
The middle ground, which includes most real codebases and complex research tasks, requires evaluation against your actual workload. Anthropic’s subscription window through June 22, 2026 provides access to Fable 5 at no additional cost for plan users — assuming it is reinstated before that date.1 The suspension has effectively consumed most of that evaluation window.
