Anthropic launched Claude Fable 5 and Claude Mythos 5 on June 9, 2026, the first general-availability release of its Mythos-class capability tier. Claude Fable 5 is the same underlying model as Mythos 5 wrapped in safety classifiers; Mythos 5 itself stays gated behind Project Glasswing for cyberdefense and a biology trusted-access program. Both models price at $10 per million input tokens and $50 per million output tokens, ship with a 1M token context window and 128k max output, and run with always-on adaptive thinking. For anyone building AI apps, agencies running Claude Code at scale, or SaaS teams embedding Anthropic via API, this is the biggest single jump in coding and long-horizon agent quality of the year.
Quick Answer
- Claude Fable 5 is Anthropic's first publicly released Mythos-class model, available June 9, 2026 on the Claude API, Amazon Bedrock, Claude Platform on AWS, Vertex AI, and Microsoft Foundry.
- Pricing: $10 input / $50 output per million tokens. Roughly half the price of Claude Mythos Preview, but 2x Opus 4.8 ($5/$25). 1M context default, 128k max output, adaptive thinking only.
- Benchmarks: 80.3% on SWE-Bench Pro (Opus 4.8: 69.2%, GPT-5.5: 58.6%, Gemini 3.1 Pro: 54.2%). 29.3% on FrontierCode Diamond (Opus 4.8: 13.4%, GPT-5.5: 5.7%).
- Claude Mythos 5 is the same model with safeguards lifted in specific areas, restricted to invited Glasswing participants and biology trusted access.
- For AI app builders: agent loops driving Totalum via MCP get a real step-change in long-horizon, multi-repo work. The price premium is justified the moment your job takes hours rather than minutes.
The June 10 Tokyo Day 1 recap covers the broader Claude Code platform expansion that Fable 5 underwrites: see our Code with Claude Tokyo 2026 recap for the 11 feature launches.
What is Claude Fable 5?
Claude Fable 5 is Anthropic's first widely released model from the Mythos training family. The training run is the same one that produced Claude Mythos Preview earlier this year, but Fable 5 ships with three classifier-based safeguard categories that route certain requests away from the full Mythos capability surface. The categories are cybersecurity (offensive cyber requests fall back to Opus 4.8), biology and chemistry (broad fallback for dual-use risks), and distillation (blocks detected attempts to extract weights or behavior for competing models).
The fallback design matters for builders. Anthropic reports that more than 95% of Fable 5 sessions never trigger a safeguard fallback, which means Fable 5 performance is effectively the same as Mythos 5 for general-purpose engineering, agent, and knowledge-work workloads. The new fallbacks parameter (beta on the Claude API and Claude Platform on AWS) lets you re-run a refused request on another model automatically, billed at the fallback model's rates.
Model IDs are claude-fable-5 and claude-mythos-5. Both use the tokenizer Anthropic introduced with Claude Opus 4.7, which produces roughly 30% more tokens than pre-Opus-4.7 models for the same text. Plan your cost calculations against the new tokenizer, not the old one.
How Claude Fable 5 differs from Claude Mythos 5
| Dimension | Claude Fable 5 | Claude Mythos 5 |
|---|---|---|
| Underlying model | Mythos-class training run | Same Mythos-class training run |
| Safeguards | Cyber, bio/chem, anti-distillation classifiers | Lifted in specific approved areas |
| Access | Generally available on Claude API + Bedrock + AWS + Vertex AI + Foundry | Restricted to Project Glasswing cyberdefenders + biology trusted access program |
| Pricing | $10 input / $50 output per Mtok | $10 input / $50 output per Mtok |
| Context window | 1M tokens default, 128k max output | 1M tokens default, 128k max output |
| Thinking | Adaptive thinking only | Adaptive thinking only |
| Data retention | 30-day required (no ZDR) | 30-day required (no ZDR) |
The split matters for two practical reasons. First, the safeguard classifiers reshape how refusals work on the API. A blocked request returns stop_reason: "refusal" and is not billed if no output was generated. The stop_details.category field can now return "reasoning_extraction", in addition to the existing "cyber" and "bio" categories, so application code can route different refusals to different next steps.
Second, Anthropic eliminated several knobs on this generation. thinking: {"type": "disabled"} is not supported on either model; you control thinking depth via the effort parameter. Manual extended-thinking budgets and assistant prefill both return a 400 error. If your application logic depends on either, fix it before flipping the model ID.
Pricing: $10 in, $50 out, and how it compares
| Model | Input ($/Mtok) | Output ($/Mtok) | Context window | Released |
|---|---|---|---|---|
| Claude Fable 5 | 10 | 50 | 1M | June 9, 2026 |
| Claude Mythos 5 | 10 | 50 | 1M | June 9, 2026 (gated) |
| Claude Mythos Preview | ~25 | ~125 | 1M | April 7, 2026 (gated) |
| Claude Opus 4.8 | 5 | 25 | 1M | May 28, 2026 |
| Claude Opus 4.7 | 5 | 25 | 1M | April 16, 2026 |
| GPT-5.5 (general) | varies | varies | 1M+ | 2026 |
Fable 5 costs roughly half of Mythos Preview, but it is 2x Opus 4.8 and far above what GPT-5.5 charges for general workloads. For builders, that price gap defines the migration question: when is the extra spend worth it.
Anthropic's subscription rollout is notable. From June 9 through June 22, Fable 5 is included free in Pro, Max, Team, and Enterprise subscription plans. From June 23 onward, subscription users need usage credits until Anthropic restores full capacity. If you run Claude inside an interactive product, plan a credit top-up workflow for late June.
Benchmarks that matter for builders
Anthropic published a wider benchmark set than usual. The ones worth tracking if you ship agent loops or coding tools:
| Benchmark | Fable 5 / Mythos 5 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| SWE-Bench Pro (agentic coding) | 80.3% | 69.2% | 58.6% | 54.2% |
| FrontierCode Diamond split | 29.3% | 13.4% | 5.7% | n/a |
| GDP.pdf (vision, no tools) | 29.8% | 22.5% | 24.9% | 16.7% |
| BioMysteryBench (Mythos 5) | 46.1% | 40.0% | n/a | n/a |
| ExploitBench (Mythos 5) | 78.0% | 40.0% | 34.0% | n/a |
The SWE-Bench Pro jump from 69.2% (Opus 4.8) to 80.3% (Fable 5) is the headline number for software work. Stripe says Fable 5 completed a 50-million-line Ruby migration in one day, work the team had previously estimated at two months. Cursor calls out a "class of long-horizon problems that were out of reach." GitHub says it "exceeded previous benchmarks" on complex multi-step tasks. Customer testimonials from Palantir, Scale AI, IMC, Amplitude, Clay, and Rogo land in the same direction.
For long-running tasks, Anthropic shared one number worth reading carefully: in a Slay the Spire evaluation, persistent file-based memory improved Fable 5's performance three times more than the same memory setup did for Opus 4.8. Long-horizon agents are where Fable 5 separates from earlier Claude generations.
API changes you need to know before migrating
The release notes contain four changes that will break naive migrations.
Thinking is adaptive-only. Code that hardcodes thinking: {"type": "disabled"} will 400. Replace with the effort parameter at low, medium, or high to control depth.
Assistant prefill is unsupported. If you prefill the start of an assistant turn for format control, you need to switch the strategy on Fable 5 / Mythos 5. Use system prompts, response schemas, or tool calls instead.
Refusals return stop_reason: "refusal". Implement a branch on stop_reason and read stop_details.category (cyber, bio, reasoning_extraction, or null). If you want automatic fallback to another model on refusal, opt into the new fallbacks parameter (beta).
30-day data retention is required. Fable 5 cannot run under zero data retention. Compliance teams should re-read Anthropic's data retention requirements before flipping production.
Bonus: thinking output defaults to display: "omitted", identical to Opus 4.8. Set display: "summarized" to receive readable thinking summaries; the raw chain of thought is never returned. Pass thinking blocks back unchanged in multi-turn conversations on the same model.
What this means for AI app builders
If you build AI apps for yourself or for clients, Fable 5 changes three concrete things.
Long-horizon agent work is finally cheap relative to value. The price doubled vs Opus 4.8, but Fable 5 also doubles the kinds of jobs that can complete end-to-end without human babysitting. A 4-hour agent run that would have cost $40 in Opus 4.8 tokens and required a rescue at hour 2 may now cost $80 and run to completion. The economics flip in favor of agent-driven work as task length grows.
Tooling that drives a code-writer is suddenly more valuable than the code-writer itself. Fable 5 is a step jump on SWE-Bench Pro and FrontierCode, but the largest customer wins (Stripe's migration, GitHub's long-horizon scores) came from agent harnesses + Fable 5 + good infrastructure. The model alone does not ship the application. This is exactly the seam Totalum is built for: your agent (Claude Code, Codex, Cursor, ChatGPT, OpenClaw) calls Totalum via MCP, and Totalum builds and maintains the real Next.js + TotalumSDK project that ships to production. See our Claude Code MCP tutorial for the integration pattern.
Coding agent comparisons need to be redone. Every "Claude Code vs Codex" or "Cursor vs Claude Code" piece written before June 9 is stale. The default backing model for Claude Code is now Fable 5 for senior-tier workloads. If you read our Claude Code vs Codex 2026 piece or our best AI coding agents 2026 roundup, assume both will pick up Fable 5 evaluations in the coming weeks.
For agencies and SaaS companies who want to embed a Fable-5-grade builder under their own brand, the Totalum API + MCP path is the fastest route. The whitelabel deployment is the same builder your agent already calls, just behind your domain and pricing. Talk to us about the embed flow on Calendly.
Companion launches in the same release
Anthropic shipped two Claude Managed Agents features alongside Fable 5 that are easy to miss.
Scheduled deployments. Claude Managed Agents now runs sessions on a cron schedule with no external scheduler required. This makes a wide set of background-agent workflows trivial: nightly code review, hourly content ingestion, daily customer-outreach triage. Pairs cleanly with the Claude Agent SDK, whose new monthly credit pool effective June 15, 2026 is covered in our Claude Agent SDK credits 2026 guide.
Environment variable credentials in vaults. Vaults now accept env-var credentials so CLIs, SDKs, and other services authenticating through environment variables work natively inside the agent sandbox. Removes the awkward credential wiring that previously blocked some third-party integrations.
Both features compound the Fable 5 thesis. Long-horizon agents need to run unattended, on a schedule, with credentials. The May Code w/ Claude London release added MCP tunnels and self-hosted sandboxes for the same reason. Today's release plus the new Mythos-class model means the agent infrastructure stack is meaningfully more complete than it was 30 days ago. For more on the partner side, see our Claude Partner Network deep-dive and the Claude skills marketplace landscape.
When to switch from Opus 4.8 to Claude Fable 5
Honest verdict: not every workload should migrate today. Three rules of thumb.
Migrate to Fable 5 immediately if: your workload includes long-horizon coding (multi-file refactors, migrations, multi-repo work), vision-heavy reasoning, or you previously bought Mythos Preview access. The price drop alone makes Mythos Preview migration a no-brainer; the performance gap closes the rest.
Stay on Opus 4.8 for now if: your workload is short interactive turns (chatbot replies, single-function code completion, classification, summarization at small context). Opus 4.8 at half the price is still excellent on those shapes, and most of your cost goes to output tokens where the gap is real money.
Stay on Haiku 4.5 for high-volume cheap inference. Fable 5 is not a replacement for Haiku. If you batch-process millions of small documents, Haiku 4.5 remains the right tool, and our Claude Code hooks playbook shows how to route subagents by tier.
A practical migration order: pick one long-running agent workflow, set the model ID to claude-fable-5, add refusal handling, remove any prefill or thinking-disable logic, run it for 48 hours on real traffic, and compare success rates and total cost per completed job. The completed-job metric matters more than the per-token spend.
What about the broader Anthropic agenda
The Fable 5 launch landed the day before Anthropic's Code w/ Claude Tokyo event (June 10) and one day after Apple announced Apple Intelligence Extensions at WWDC 2026 (covered in our Apple Intelligence Extensions analysis). Three reads from this calendar.
First, Anthropic now operates on a monthly model cadence at the frontier (Opus 4.7 April 16, Opus 4.8 May 28, Fable 5 June 9). If you wait six months to migrate, you skip two major capability jumps.
Second, the Mythos training family is real. Mythos Preview was not marketing copy; Fable 5 is its productized form. Expect Mythos 6 within the year, and expect the safeguard architecture to evolve rather than disappear.
Third, the platform plus model story is now the differentiator. Managed Agents scheduled deployments + vault env vars + MCP tunnels + self-hosted sandboxes + skills marketplaces + Fable 5 add up to a build-deploy-run loop that competitors will need 12 months to replicate. For builders, the right move is to lean into the loop now rather than wait for parity.
FAQ
When was Claude Fable 5 released?
Claude Fable 5 was released on June 9, 2026. It became available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Vertex AI, and Microsoft Foundry the same day, with US East (N. Virginia) and Europe (Stockholm) as the first Bedrock regions.
How much does Claude Fable 5 cost?
$10 per million input tokens and $50 per million output tokens. That is less than half the price of Claude Mythos Preview and roughly double Claude Opus 4.8 ($5 / $25 per million tokens). Subscription plans (Pro, Max, Team, Enterprise) include Fable 5 free from June 9 through June 22, then require usage credits.
What is the difference between Claude Fable 5 and Claude Mythos 5?
Both share the same underlying Mythos-class training. Claude Fable 5 has safety classifiers that route requests in cybersecurity, biology/chemistry, and anti-distillation categories to Opus 4.8 as a fallback. Claude Mythos 5 is the same model with those safeguards lifted in specific areas, and access is restricted to Project Glasswing cyberdefenders and a biology trusted-access program.
Can I use Claude Fable 5 with zero data retention?
No. Claude Fable 5 requires 30-day data retention on the Claude API and is not available under zero data retention. Compliance-sensitive workloads must factor this in before switching.
Does Claude Fable 5 support disabled thinking or assistant prefill?
No. Adaptive thinking is the only thinking mode, and thinking: {"type": "disabled"} returns a 400 error. Assistant prefill also returns a 400 error. Use the effort parameter (low, medium, high) to control thinking depth instead.
How does Claude Fable 5 fit with Totalum?
Totalum is an AI app builder that creates real Next.js + TotalumSDK projects. Agents like Claude Code, Codex, Cursor, ChatGPT, and OpenClaw can call Totalum via MCP or REST API to create and maintain production apps. Upgrading the model that drives your agent to Claude Fable 5 makes the agent better at long-horizon, multi-repo, multi-file work, which is exactly the kind of project Totalum is best at materializing into a real app.
Ready to build with Claude Fable 5 and Totalum?
If you are a software agency or a SaaS company looking to embed a Mythos-class AI app builder into your product, the Totalum API and MCP are the fastest way to ship. Book a 30-minute call to see Totalum live on Calendly.
If you are an indie developer, SaaS founder, or Claude Code power user, point your agent at Totalum and ship a real production app from your terminal. Start free at totalum.app.
Sources: Anthropic press release on Claude Fable 5 and Mythos 5, Claude Platform release notes June 9, 2026, Vellum benchmark breakdown, AWS announcement of Claude Fable 5 on Bedrock.