Models · facts verified 08.07.2026
New version: Claude Fable 5.1.
Anthropic's smartest publicly available model and the first Mythos class - a tier above Opus.
Use it for the highest quality ceiling: architecture, large-scale migrations, tasks where Opus 4.8 hits its limit. Do not use it for a high-volume daily pipeline - $13.41/task versus $7.23 for GPT-5.5, a premium.
$10/$50
api per 1M tokens · in subscriptions - usage credits from 13.07
1M
4.7+ tokenizer: same text ≈ +30% tokens
01.2026
model knowledge cutoff
through workarounds
From Russia: not directly
Should you get it
The short answer before the tables and comparisons.
Use it if you need
- +quality ceiling: architecture, migrations at scale, tasks where Opus 4.8 hits its limits
- +long agent sessions: adaptive thinking and task retention for days
- +top baseline: check cheaper model outputs against it
- +complex reviews and plans where the cost of error is orders of magnitude above token cost
Skip it if you need
- −daily high-volume pipeline - $13.41/task vs $7.23 for GPT-5.5: overpaying
- −response speed: comparative latency "Slower" - the slowest tier in the lineup
- −organizations with zero data retention: Mythos class requires 30-day retention - model unavailable
- −direct payment from Russia - does not work (see Connection)
Access by platform
One model - different names and channels: Claude Code, Claude API, OpenRouter, RU aggregator. Start with Claude Code: the model is enabled there with one command.
Claude Code
Not default on any plan: enabled explicitly; from 13.07 - usage credits on top of the plan.
claude
/model fableClaude API
Canonical id - claude-fable-5; dateless id = pinned snapshot.
curl https://api.anthropic.com/v1/messages \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-d '{"model": "claude-fable-5", "max_tokens": 1024, "messages": [...]}'OpenRouter
Same price $10/$50; crypto or a foreign card.
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_KEY" \
-d '{"model": "anthropic/claude-fable-5", "messages": [...]}'RU aggregator
Anthropic-compatible format: base_url and key change. Rubles, no VPN, ~1.5-2× markup.
curl https://api.proxyapi.ru/anthropic/v1/messages \
-H "Authorization: Bearer $PROXYAPI_KEY" \
-d '{"model": "claude-fable-5", "messages": [...]}'For API - a platform.claude.com key with balance; from 13.07, the subscription needs usage credits on top of the plan. Neither can be paid for directly from Russia.
Benchmark standing
An agent score is always an “agent + model” pair: this is where Fable 5 lives in the lineup.Model score ≠ tool score: Fable 5 is measured both solo and paired with Claude Code. DeepSWE measures the xhigh configuration; vendor figures were taken on Anthropic's proprietary scaffold.
AA Coding Agent Index 23.09.2026 · Terminal-Bench 4.0 23.09.2026 · DeepSWE v1.1 22.09.2026 · “—” - pair not measured
Read honestly: on DeepSWE Fable has been second since 10.07 - GPT-5.6 Sol delivered 72.7% vs 69.9% at half the task price ($8.39 vs $13.41). The vendor's 80.3% SWE-bench Pro was taken on its proprietary scaffold and is disputed; on Vending-Bench (business agent tasks, Andon Labs), unrestricted Mythos 5 lost even to Opus 4.7. SWE-bench Verified is contaminated - we do not use it. Everyday downside (Ethan Mollick): Fable's ornate style leaks into interface copy even against instructions.
Traits and limits
What Fable 5 feels like: class, reasoning, context capacity, and whether the weights are closed.
Mythos class
Mythos-class - a tier above Opus. Fable 5 and Mythos 5 are the same model: Fable's cybersecurity, bio/chemistry, and distillation classifiers roll the response back to Opus 4.8 (under 5% of sessions on average, per vendor claim). Mythos 5 without some restrictions is available only to Project Glasswing partners; no self-service signup.
Adaptive thinking · always on
Extended thinking is unavailable as a separate toggle - the model always thinks adaptively. The vendor's comparative latency is «Slower»: this is the lineup's slowest tier.
1M context · tokenizer 4.7+
Opus 4.7+ generation tokenizer: the same text produces roughly 30% more tokens than with models before 4.7 - actual window capacity is lower than Opus 4.6 with the same «1M».
Closed weights
Open-weight alternatives for code - Kimi K2.6, GLM-5.x, DeepSeek V4; they trail Mythos-class models in coding quality.
Pricing and access from Russia
$10/$50 per 1M tokens (api per 1M tokens · in subscriptions - usage credits from 13.07).
API - $10/$50 per 1M tokens (Bedrock, Vertex, Foundry - the same). Subscriptions: the window closed 12.07.2026; from 13.07 - usage credits on top of the plan. In Claude Code - /model fable. From Russia - OpenRouter or RU aggregators.
Anthropic blocks Russia by IP and cards. Working routes: OpenRouter (anthropic/claude-fable-5 - same price, crypto or a foreign card), RU aggregators in rubles without VPN (~1.5-2× markup). DeepSeek is the only strong option that works directly from Russia and is free.
All vendors reworked billing into token credits in 2026; figures change every few weeks. Every row above has a verification date: do not trust figures without a date, including ours.
Risks and policy
What to check before bringing a Mythos-class model into production: data storage, export controls, closed weights.
- DataAnthropic news · 08.07.2026
- 30-day retention for Mythos-class; ZDR unavailableFable 5 has a 30-day data retention policy (class-level safety monitoring) - organizations with zero data retention do not get the model: this is stricter than regular Claude models. Data is processed on foreign servers - important for 152-FZ.
- Export controlAnthropic: Redeploying Fable 5 · 08.07.2026
- 06/12-06/30: suspension at the request of the US governmentOn June 12, the US applied export controls to Fable 5 and Mythos 5 (after Amazon researchers reported bypassing safeguards); without reliable real-time citizenship verification, Anthropic suspended the models globally for everyone. Controls were lifted on June 30; on July 1, the model returned with a new classifier blocking the discovered jailbreak in >99% of cases.
- WeightsClaude Platform docs · 08.07.2026
- Closed weights - cannot be deployed locallyOpen-weight alternatives for code - Kimi K2.6, GLM-5.x, DeepSeek V4; they trail Mythos-class models in coding quality.
Fable 5 or GPT-5.5 - historical snapshot 10.07
Historical snapshot 10.07: both models were the pinnacle of the closed class. The numbers preserve the round's trade-off - the highest ceiling on the hardest tasks versus nearly the same quality at half the price.
- →Summer 2026 working pattern (Mitchell Hashimoto): Fable plans the task and judges the result, GPT-5.6 writes the code (GPT-5.5 filled this role until 09.07) - full-cycle quality for a fraction of the price. Use Fable end-to-end only where the cost of an error is orders of magnitude higher than the cost of tokens.
Quick answers
How much does Claude Fable 5 cost - including in rubles?+
Via API, $10 input / $50 output per 1M tokens - twice as expensive as Opus 4.8. The subscription window closed on July 12, 2026; from July 13 - $10/$50 usage credits on top of the plan. Through RU aggregators - the same tokens in rubles with a ~1.5-2× markup. A typical task in our DeepSWE snapshot - $13.41.
What is Mythos-class, and how does Fable differ from Mythos?+
Mythos-class - Anthropic's tier above Opus. Fable 5 and Mythos 5 are the same model: Fable's safety classifiers (cybersecurity, bio/chemistry, distillation) roll responses back to Opus 4.8 - under 5% of sessions, per vendor claim. Mythos 5 without some restrictions is available only to invited Project Glasswing partners.
Was Fable 5 removed from subscriptions? What was the window through July 12?+
The model was planned to move from subscription limits to usage credits from July 8, but on the evening of July 7 Anthropic extended the window through July 12. It closed, and from July 13 Fable 5 is available through usage credits ($10/$50 per MTok) on top of the plan; Opus 4.8, Sonnet, and Haiku remain within subscription limits. Policy changes weekly - check the date on this card.
Fable 5 or GPT-5.5 - which is better for code?+
On independent DeepSWE, Fable 5 outperforms GPT-5.5: pass@1 of 69.9% vs. 67.0% (both in the highest mode). But a Fable task cost ~1.9 times more ($13.41 vs. $7.23). Before the 07/09 transition, GPT-5.5 was the cost-effective brain of the daily pipeline; GPT-5.6 now fills that role. Fable remains the ceiling for the most difficult tasks and long agent sessions.
Fable 5 or Opus 4.8 - when is Opus enough?+
Opus 4.8 is twice as cheap ($5/$25), faster, stays within subscription limits, and remains the Claude Code default. Use Fable for tasks where Opus hits a wall: architecture decisions, over a day of agent work, large-scale migrations. For daily PRs and edits, Opus 4.8 is the workhorse, Fable is the power reserve.
How do I get access from Russia?+
Anthropic blocks Russia by IP and cards. Two routes: OpenRouter (slug anthropic/claude-fable-5, same $10/$50 price, payment by crypto or foreign card) and RU aggregators in rubles without VPN with a ~1.5-2× markup. A Claude subscription cannot be paid for directly from Russia.
Can it be deployed locally?+
No, weights are closed, and Mythos-class also has 30-day retention with safety monitoring - the model is unavailable to organizations with zero data retention in principle. For self-hosting, see the open-weight class: Kimi K2.6, GLM-5.x, DeepSeek V4.
Why does the model sometimes respond worse than usual?+
Fable 5's classifiers on restricted topics (cybersecurity, bio/chemistry) roll the response back to Opus 4.8 - under 5% of sessions, per vendor claim. If the topic is far from restricted areas but quality has dropped - check whether the usage credits limit was triggered and the session switched to another model.