TokenMix Research Lab · 2026-06-10

Claude Fable 5 Review 2026: Pricing, Benchmarks, vs Opus 4.8

Last Updated: 2026-06-10 Author: TokenMix Research Lab Data verified: 2026-06-10 — Anthropic announcement, Claude API docs (pricing, migration guide, models overview), Claude Code docs, AWS News Blog, TechCrunch, The Decoder, OpenRouter, Hacker News launch thread; competitor pricing cross-checked against OpenAI and Google published rates

Claude Fable 5 launched June 9, 2026 at $10 per million input tokens and $50 per million output — Anthropic's first generally available Mythos-class model, priced at exactly double Claude Opus 4.8. The headline numbers are real but uneven — SWE-Bench Pro 80.3% (vs 69.2% for Opus 4.8), a 1M-token context window, 128K max output, but also mandatory 30-day data retention with no zero-retention option, safety classifiers that silently reroute up to 5% of sessions to Opus 4.8, and a price that makes routine workloads cheaper to run on the model it replaces.

Fable 5 is the public face of a two-model release: Anthropic's announcement pairs it with Claude Mythos 5, the same underlying model with safety classifiers lifted in restricted domains, available only to vetted Project Glasswing partners. Per the official API docs, both models run adaptive thinking permanently on, never return raw chain-of-thought, and introduce a new refusal-and-fallback billing model that no previous Claude release had. This review covers verified pricing, Anthropic-published benchmarks, the API behavior changes, and the cost math against Opus 4.8, GPT-5.5, and Gemini 3.1 Pro — every claim tagged with confirmed or likely confidence tiers, with vendor-reported numbers flagged as such.

Table of Contents

• Quick Verdict
• Mythos-Class Explained: Fable 5, Mythos 5, Mythos Preview
• Pricing: $10/$50 Per MTok, Double Opus 4.8
• Benchmarks: SWE-Bench Pro 80.3%, FrontierCode 29.3%
• API Changes: Adaptive Thinking, Refusals, Fallbacks
• Safeguards: The Opus 4.8 Fallback, Explained
• Is Claude Fable 5 Worth Double the Price of Opus 4.8?
• Availability: API, Cloud Platforms, Claude.ai Rollout
• Where Fable 5 Still Loses
• Final Recommendation
• FAQ

Quick Verdict

Claude Fable 5 is the first model tier Anthropic has shipped above Opus — and the published deltas are large enough that the 2× price is defensible for hard agentic work, though not for routine tasks.

Mythos-Class Explained: Fable 5, Mythos 5, Mythos Preview

Fable 5 and Mythos 5 are the same underlying model — the difference is safety classifiers, not capability. Anthropic's model hierarchy now has a tier above Opus, and the naming is deliberate: fable, from the Latin fabula ("that which is told"), is the Latin cousin of the Greek mythos, according to The Information's reporting on the launch.

The sequence matters for anyone tracking Anthropic's release pattern. Mythos Preview shipped in April 2026 through Project Glasswing, the restricted-access program TokenMix covered when Mythos found 23,019 software flaws in partner codebases. Per Anthropic, Glasswing has now expanded to roughly 150 new organizations across more than fifteen countries, and Mythos 5 replaces Preview for those partners at — per Anthropic's announcement — less than half Preview's price. A trusted-access program for biology researchers is planned next, with cyber safeguards still in place for that group.

One framing from launch week is worth keeping: TechCrunch notes that Anthropic shipped its most capable public model days after publicly warning that frontier systems are approaching recursive self-improvement and calling for a "coordinated brake pedal" on frontier development. The tension is the story: the company's answer to "too dangerous" is not withholding the model — it is shipping it wrapped in classifiers, with the unwrapped version behind a vetting program.

What this does NOT tell us: whether Fable 5 is a larger model than Opus 4.8, what it cost to train, or its knowledge cutoff — the models overview spec table omits a cutoff date for Fable 5 entirely, while listing Opus 4.8's (January 2026). Treat parameter-count claims circulating this week as speculation; our earlier Mythos 5 pricing forecast tracked those rumors before launch.

Pricing: $10/$50 Per MTok, Double Opus 4.8

Fable 5 costs $10 per million input tokens and $50 per million output — exactly double Claude Opus 4.8's $5/$25, and the highest list price in Anthropic's current lineup. Mythos 5 is priced identically for Glasswing partners.

Official Anthropic lineup, June 2026:

Source: Anthropic models overview, verified 2026-06-10.

Cache and batch rates shipped on day one — the launch-day rate gap many teams expected did not happen. The full Fable 5 rate card against Opus 4.8, from Anthropic's pricing docs:

The table is monotonous by design: every Fable rate is exactly 2× Opus 4.8, so cache-heavy and batch-heavy workloads keep their existing discount structure — 90% off on cache reads, 50% off on batch — at the new base. The one row that moves in Fable's favor is the cache minimum, down to 512 tokens (1,024 on Bedrock), so short system prompts that never cached on Opus 4.8 now do. One absence worth knowing: fast mode, the premium tier that speeds up Opus models, is not offered on Fable 5 — and per the same pricing page, Opus 4.8 fast mode costs $10/$50. That makes for a genuinely odd coincidence: the same sticker price buys either Opus 4.8 at higher speed or Fable 5 at higher intelligence.

Against the flagship field, Fable 5 is now the most expensive frontier model per output token among the big three labs' standard tiers:

Three pricing footnotes that change real bills. First, Fable 5 uses the tokenizer introduced with Opus 4.7 — Anthropic's models overview puts it at roughly 30% more tokens from the same text versus pre-4.7 models, and the pricing page says up to 35% depending on content. Comparisons against Opus 4.8 are apples-to-apples; comparisons against your old Opus 4.5-era bills are not. Second, the 1M context window carries no long-context surcharge — per the pricing docs, "a 900k-token request is billed at the same per-token rate as a 9k-token request." Gemini 3.1 Pro doubles its input rate past 200K; Fable 5 doesn't, which quietly narrows the gap on exactly the long-context work Fable targets. Cache mechanics across the lineup are in the Claude cache pricing guide. Third, the fallback mechanism has its own billing rules — covered below — where rerouted requests bill at Opus prices, not Fable prices.

For current per-model rates across the rest of the Claude lineup, the Claude API pricing guide has the full table including Sonnet and Haiku tiers.

Benchmarks: SWE-Bench Pro 80.3%, FrontierCode 29.3%

Anthropic's published numbers put Fable 5 11.1 points above Opus 4.8 on SWE-Bench Pro (80.3% vs 69.2%) and 2.2× ahead on FrontierCode (29.3% vs 13.4%) — and the pattern across evals is consistent: the harder the task, the larger the gap. All numbers in this section are Anthropic-published launch evals as reported by The Decoder; independent replication is pending, so treat the deltas as the signal, not the absolute scores.

Coding benchmarks:

The FrontierCode number deserves the attention. SWE-Bench Pro is approaching saturation at the top of the leaderboard — an 11-point lead is meaningful but incremental. FrontierCode targets production-grade, long-horizon engineering tasks, and 29.3% vs 13.4% means Fable 5 completes more than twice the tasks Opus 4.8 can. Per Anthropic, Fable 5 scores highest among frontier models on FrontierCode even at medium effort — the effort parameter that controls thinking depth, covered in the API section.

Beyond coding:

The ExploitBench row explains the two-model strategy in one number: the same weights that hit 78% on exploit development are what Anthropic routes away from in the public model. Per Anthropic, external bug-bounty testing logged over 1,000 hours against Fable 5's classifiers without producing a universal jailbreak — though the announcement itself concedes the UK AI Security Institute made progress toward one in initial testing. That claim now gets stress-tested by the entire internet.

Early customer signal points the same direction as the evals — and one number from the launch post outweighs the entire quote wall: Stripe reports Fable 5 completed a 50-million-line Ruby codebase migration in one day, against a roughly two-month manual estimate. Davis Polk says lawyers in blind review preferred its redlines over the incumbent model every time; Anaconda reports wins at every effort level while running 25-30% faster. All vendor-curated, all unaudited — but consistent across fourteen named customers, and consistent with the long-horizon framing. For the Opus 4.8 baseline these claims are measured against, see our Claude Opus 4.8 review.

API Changes: Adaptive Thinking, Refusals, Fallbacks

Fable 5 ships three breaking behavior changes — adaptive thinking is always on, raw chain-of-thought is never returned, and refusals arrive as successful HTTP 200 responses with stop_reason: "refusal". If your integration assumes Opus-era behavior, all three need handling before migration.

Three practical notes from the launch docs and the migration guide. The refusal shape is the integration trap: a refused request is a successful API call, so anything keyed on HTTP status codes will treat a refusal as a normal completion and pass an empty or declined response downstream. Check stop_reason on every Fable 5 response. The fallbacks parameter is in beta on the Claude API and Claude Platform on AWS only — it is not available on the Batch API, Bedrock, Vertex, or Foundry, where retries run client-side via the SDK middleware. And max_tokens caps thinking plus response combined: a workload that ran thinking-off on Opus 4.8 inherits always-on adaptive thinking here, so output budgets sized for bare responses will truncate. Anthropic ships an automated path for all of this — /claude-api migrate in Claude Code applies the model swap and parameter changes across a codebase.

Vision, tool use, and the Messages API are otherwise unchanged — Anthropic states the API behaves identically for Opus, Sonnet, and Haiku models.

Safeguards: The Opus 4.8 Fallback, Explained

Less than 5% of sessions trigger the safeguards — but when they do, a different model answers your request and the rate changes mid-conversation. Per Anthropic, early data shows 95% of Fable sessions run entirely on Fable's own responses.

The classifier domains, per Anthropic's announcement:

The category strings are what the API actually returns, per the migration guide — a refusal can also carry null when it maps to no named classifier. The distillation safeguard drew the sharpest independent criticism of launch week: Nathan Lambert's Interconnects analysis reports it operates partly through invisible interventions — "prompt modification, steering vectors, or parameter-efficient fine-tuning" — rather than visible refusals, and his verdict is blunt: "An AI model that gets less intelligent automatically without notifying me is categorically misaligned AI." The cyber and bio classifiers, by contrast, he rates intellectually consistent, if frustrating.

The fourth row of the domain table is a real source discrepancy: AWS's launch blog describes fallback for "cybersecurity, biology, chemistry, and health" while Anthropic's own announcement names three domains and adds distillation. We tag health as Likely-not-a-separate-domain until either document is corrected.

Billing on fallback follows the request, not the model you asked for:

Source: AWS News Blog and Anthropic API docs, verified 2026-06-10.

One trigger pattern surfaced by the Claude Code docs deserves its own flag: fallback can fire on the first request of a session, before you type anything unusual, because that request carries workspace context — CLAUDE.md content, directory names, git status. A repo containing security tooling or biology material can trip the classifier on context alone; claude --safe-mode strips customizations to diagnose it. For substantive biology work the docs are blunt: expect nearly all requests to reroute, and apply to a trusted-access program if you need Fable-class capability in those domains.

The governance trade is explicit and worth stating plainly: Fable 5 and Mythos 5 are designated Covered Models. That means 30-day retention of inputs and outputs is mandatory — zero-data-retention agreements do not apply — and on AWS it requires opting into the provider_data_share mode, with human review of flagged traffic for jailbreak defense. TechCrunch's coverage flags the precedent concern: safety framing normalizing mandatory retention. If your compliance posture requires ZDR, Fable 5 is off the table today regardless of capability — that is a harder blocker than price.

Is Claude Fable 5 Worth Double the Price of Opus 4.8?

The short answer: yes for frontier-hard agentic work, no for routine tasks — the cost-per-solved-task math flips as difficulty rises. Raw per-token cost says Fable 5 is 2× Opus 4.8; success-rate-adjusted cost says the opposite once tasks get hard enough that Opus starts failing.

Take a representative agentic coding task consuming 100K input + 20K output tokens:

Now divide by published success rates. Cost per solved task — attempt cost ÷ benchmark pass rate:

The crossover is the whole decision. On SWE-Bench-Pro-difficulty work, Opus 4.8 solves enough that its 2× price advantage survives the success-rate adjustment — $1.45 vs $2.49 per solved task. On FrontierCode-difficulty work, Opus fails so often that retries eat the savings: Fable lands at $6.83 per solve against Opus's $7.46 — and GPT-5.5 at $19.30 is not in the conversation. Benchmark difficulty is a proxy, not your workload — but the direction is robust: route by task difficulty, not by loyalty to a price point.

A monthly bill sanity check at fleet scale: a team pushing 50M input + 10M output tokens per month pays $1,000/month on Fable 5 vs $500/month on Opus 4.8 — a $6,000/year delta that only pays for itself if Fable's higher completion rate saves more than that in engineer review time or retry compute. At typical loaded engineering costs, that is roughly 30-50 engineer-hours per year — about one saved hour per week. Plug your own traffic into our Claude cost calculator to find your crossover.

Early field reports cut both ways on whether the 2× rate survives contact with real workloads. The Hacker News launch thread runs roughly 60/40 positive: several developers report Fable finishing tasks in fewer turns with "more targeted and surgical diffs," one claiming better results with about half the tokens — which would put effective cost near Opus parity. The other side: a Max-plan user metered $82.92 in API-equivalent usage in a single day, and multiple reports describe classifier false positives on legitimate work — MRI brain-segmentation code and mosquito-malaria research both flagged as bio risks. Self-reported anecdotes, not benchmarks, but the variance itself is the takeaway: meter your first week before committing fleet traffic.

There is also the latency-shaped caveat: always-on adaptive thinking means Fable 5 is not built for interactive low-latency paths. The model to beat for those remains Sonnet 4.6 at $3/$15 — a fifth of Fable's input price.

Availability: API, Cloud Platforms, Claude.ai Rollout

Fable 5 went generally available on five platforms simultaneously on June 9 — Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry — plus GA in GitHub Copilot the same day.

On the consumer side, the rollout is staged: per Anthropic, Fable 5 is included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost from June 9 through June 22; from June 23 onward it draws on usage credits, with broader restoration planned as capacity allows. Translation: subscription users get a two-week free taste, then Fable becomes metered while Opus 4.8 stays the bundled default. In Claude Code it is never the default on any plan — you opt in with /model fable, and the best alias resolves to Fable 5 only where your org has access.

The aggregator gap is the practical note for API buyers. As of June 10, TokenMix's catalog of 170 models lists Claude through Opus 4.8 but not yet Fable 5 — typical for day-two of an Anthropic release. Teams already routing Claude traffic through TokenMix's OpenAI-compatible endpoint can keep Opus 4.8 as the workhorse and add claude-fable-5 direct-from-Anthropic for the frontier-hard slice until relay availability lands; the cost-per-solve table above is the routing rule.

Where Fable 5 Still Loses

Fable 5 loses on five axes that matter in production: routine-task economics, zero-data-retention compliance, interactive latency, classifier false positives, and day-one ecosystem availability.

None of these are fatal; all of them are real. The honest framing: Fable 5 is a specialist top tier for the work Opus 4.8 fails at, not a drop-in replacement for the whole Claude stack.

Final Recommendation

Adopt Fable 5 for the frontier-hard 10-20% of your agentic workload where Opus 4.8's failure rate is costing you retries and review hours; keep everything else on Opus 4.8 and Sonnet 4.6, and keep ZDR-bound traffic off it entirely — the 30-day retention requirement has no workaround. Cache and batch discounts carry over at exactly 2× Opus rates, so cache-heavy workloads migrate without rate surprises. The published benchmark deltas are vendor-reported and await independent replication — but the cost-per-solve math holds even with generous error bars on the launch evals.

FAQ

What is Claude Fable 5?

Claude Fable 5 is Anthropic's most capable generally available model, released June 9, 2026 — the public version of its Mythos-class tier, which sits above Opus in the lineup. It is the same underlying model as Claude Mythos 5 with safety classifiers active: high-risk cyber, bio/chem, and distillation requests are answered by Claude Opus 4.8 instead.

How much does the Claude Fable 5 API cost?

$10 per million input tokens and $50 per million output tokens — exactly double Claude Opus 4.8's $5/$25, and every other rate doubles with it: cache reads $1/MTok, 5-minute cache writes $12.50, 1-hour writes $20, batch $5/$25. Requests rerouted by the safety classifiers bill at Opus rates, and refused requests that produce no output are free.

What is the difference between Claude Fable 5 and Claude Mythos 5?

Safety classifiers — nothing else. Anthropic states they share the same underlying model and the same $10/$50 pricing; Mythos 5 has cyber safeguards lifted and is restricted to vetted Project Glasswing partners. Mythos 5 hit 78% on ExploitBench versus 40% for Opus 4.8, which is exactly the capability the public model routes away from.

Is Claude Fable 5 better than Claude Opus 4.8 for coding?

Yes on Anthropic's published evals, and by a wide margin on hard tasks: SWE-Bench Pro 80.3% vs 69.2%, FrontierCode 29.3% vs 13.4%. The gap grows with task difficulty — on routine work the quality delta rarely justifies double the price, but on frontier-hard tasks Fable 5 is actually cheaper per solved task ($6.83 vs $7.46).

What happens when Claude Fable 5 refuses a request?

You get a successful HTTP 200 response with stop_reason: "refusal" and stop_details.category naming the classifier ("cyber", "bio", "reasoning_extraction", or null) — not an error. Refusals with no output are free; if the classifier fires mid-stream, input plus already-streamed output is billed. The beta fallbacks parameter auto-retries on another Claude model server-side on the Claude API and Claude Platform on AWS; everywhere else, SDK middleware (TypeScript, Python, Go, Java, C#) handles it client-side.

Does Claude Fable 5 support extended thinking?

No — it uses adaptive thinking, permanently on. thinking: {"type": "disabled"} is not supported, raw chain-of-thought is never returned, and thinking.display defaults to omitted (set it to summarized for readable summaries). Thinking depth is controlled with the effort parameter.

What is Claude Fable 5's context window?

1M tokens by default with up to 128K output per request, and no long-context surcharge — a 900K-token request bills at the same per-token rate as a 9K one. It uses the tokenizer introduced with Opus 4.7, which produces roughly 30% (up to 35%) more tokens from the same text than pre-4.7 Claude models — relevant when comparing bills across model generations.

Can I use Claude Fable 5 with zero data retention?

No. Fable 5 and Mythos 5 are Covered Models: 30-day retention of inputs and outputs is mandatory and zero-data-retention agreements do not apply. On AWS this requires the provider_data_share opt-in. If ZDR is contractual for you, stay on Opus 4.8 or Sonnet 4.6.

Can I use Claude Fable 5 in Claude Code?

Yes — run /model fable on Claude Code v2.1.170 or later. It is not the default on any plan, and effort defaults to high. Requests its classifiers flag re-run automatically on Opus 4.8 with a transcript notice; security- or biology-heavy repos can trigger this on the first request from workspace context alone. It does not appear under zero-data-retention accounts.

Is Claude Fable 5 included in Claude.ai subscriptions?

Temporarily. Pro, Max, Team, and seat-based Enterprise plans include it during the staged rollout June 9-23, 2026; after June 23 it draws on usage credits, with restoration as a standard feature planned later per Anthropic.

Sources

• Anthropic — Claude Fable 5 and Claude Mythos 5
• Claude API Docs — Introducing Claude Fable 5 and Claude Mythos 5
• Claude API Docs — Models overview
• Claude API Docs — Pricing
• Claude API Docs — Migration guide
• Claude Code Docs — Model configuration
• TechCrunch — Anthropic's Claude Fable 5 is a version of Mythos the public can access today
• The Decoder — Anthropic releases Claude Fable 5 and Mythos 5 with major gains in coding and science
• AWS News Blog — Anthropic Claude Fable 5 on AWS
• CNBC — Anthropic releases Mythos-like AI model to the public, Claude Fable 5
• GitHub Changelog — Claude Fable 5 is generally available for GitHub Copilot
• OpenRouter — Claude Fable 5 listing
• Interconnects (Nathan Lambert) — Claude Fable 5 and the new AI safety
• Hacker News — Claude Fable 5 launch thread
• OpenAI — API Pricing
• Google AI for Developers — Gemini Developer API pricing

Related Articles

• Claude Opus 4.8 Review 2026: Pricing, Benchmarks, vs 4.7 and GPT-5.5
• Claude Mythos vs Opus 4.8: What Makes a Model Mythos-Class 2026
• Claude Mythos 5: 10 Trillion Parameters, Pricing Forecast
• Project Glasswing: Claude Mythos Found 23,019 Software Flaws 2026
• Claude API Pricing 2026: Opus 4.8, Sonnet 4.8, Haiku 4.5 Compared

About TokenMix — TokenMix.ai is an AI API relay that routes Claude, OpenAI, Gemini, DeepSeek, Qwen, and other large language models through a single OpenAI-compatible endpoint at https://api.tokenmix.ai/v1. Current model availability and per-token rates are listed on the pricing page and the model catalog. Integration uses the standard OpenAI SDK; details in the OpenAI compatibility reference.