Claude Fable 5 vs Mythos 5 vs GPT-5.5: Public Launch
Claude Fable 5 vs Mythos 5 vs GPT-5.5: The Public Mythos Launch
Anthropic finally released a public Mythos-class model on June 9, 2026. Here’s what changed — and how Fable 5, Mythos 5, and GPT-5.5 actually compare for developers in June 2026.
Last verified: June 10, 2026
TL;DR
| Question | Answer |
|---|---|
| Best for agentic coding | Claude Fable 5 (SWE-Bench Pro 80.3%) |
| Best for chat / multimodal | GPT-5.5 (ChatGPT ecosystem, 5M+ Codex users) |
| Best for unrestricted cyber research | Mythos 5 (trusted-access only) |
| Cheapest for most workloads | Claude Opus 4.8 (half the price of Fable 5) |
| Best value cheap-tier | Gemini 3.5 Flash ($1.50/$9 per MTok) |
What changed on June 9, 2026
Anthropic announced two related models:
- Claude Fable 5 (
claude-fable-5) — the first publicly available Mythos-class model, with automated guardrails on cybersecurity and biology - Claude Mythos 5 (
claude-mythos-5) — the full unrestricted model, gated behind a trusted-access program for governments, cybersecurity firms, and ~150 enterprise partners
Both ship with:
- 1M-token context window
- 128k token max output per request
- $10/$50 per million input/output tokens (batch: $5/$25)
- Same base architecture and weights — the difference is the safeguard layer
This is the first time Anthropic has shipped a Mythos-class model to general developers. The previous Mythos Preview was capped at vetted partners and priced at roughly $25/$125 per MTok.
Benchmarks: Fable 5 vs the field
| Benchmark | Fable 5 | Mythos 5 | GPT-5.5 | Opus 4.8 | Gemini 3 Pro |
|---|---|---|---|---|---|
| SWE-Bench Pro | 80.3% | 80.3% | 58.6% | 71.4% | 64.2% |
| Terminal-Bench 2.1 | 84.1% | 84.1% | 67.8% | 76.2% | 76.2% |
| MCP Atlas | 88.7% | 88.7% | 73.1% | 83.6% | 83.6% |
| GPQA Diamond | 87.8% | 87.8% | 82.1% | 81.0% | 86.4% |
| AIME 2025 | 96.2% | 96.2% | 92.0% | 91.3% | 94.1% |
| CharXiv Reasoning | 86.6% | 86.6% | 80.2% | 80.9% | 84.2% |
| AISI Cyber Eval | restricted | frontier | restricted | restricted | restricted |
The signal: Fable 5 is roughly 8–10 points ahead of Opus 4.8 across agentic coding benchmarks, and roughly 20 points ahead of GPT-5.5. Reasoning and multimodal are tighter — Gemini 3 Pro is still competitive on long-context and image tasks.
Pricing comparison
| Model | Input ($/MTok) | Output ($/MTok) | Context |
|---|---|---|---|
| Claude Fable 5 / Mythos 5 | $10 | $50 | 1M |
| Claude Opus 4.8 | $5 | $25 | 500k |
| Claude Haiku 4.5 | $0.80 | $4 | 200k |
| GPT-5.5 | $5 | $25 | 400k |
| GPT-5.5 mini | $1.20 | $6 | 400k |
| Gemini 3 Pro | $5 | $25 | 1M |
| Gemini 3.5 Flash | $1.50 | $9 | 1M |
Fable 5 is the most expensive frontier model on the market — 2x Opus 4.8 and 2x GPT-5.5. For most coding tasks, the orchestrator-subagent pattern (Opus 4.8 orchestrator + Haiku 4.5 fan-out + Gemini 3.5 Flash for cheap calls) is still more economical.
What Fable 5 will refuse to do
The automated guardrail layer blocks:
- Offensive cybersecurity — writing exploit code, payloads, reverse shells, ransomware
- Vulnerability research at scale — chained CVE analysis, attack-path generation
- Biology / bioweapons — synthesis routes, gain-of-function reasoning
- Some dual-use research — cases vary; Anthropic acknowledges false positives
Early testing reports false-positive refusals on legitimate red-team work, CTF challenges, and academic security research. If your workflow includes any of those, you may need to apply for the trusted-access program to use Mythos 5.
Coverage: Should you wait for Claude Mythos public release? — written before the launch, now resolved.
Which model should you use?
| Workload | Recommended |
|---|---|
| Autonomous coding agent (frontier) | Claude Fable 5 |
| Routine coding / refactoring | Claude Opus 4.8 |
| Fan-out subagent worker | Claude Haiku 4.5 or Gemini 3.5 Flash |
| Chat / multimodal consumer | GPT-5.5 |
| Long-context (1M+) reasoning | Fable 5 or Gemini 3 Pro |
| Cheap fast inference | Gemini 3.5 Flash |
| Offensive security / cyber research | Mythos 5 (trusted access) |
| Free on-device (Apple) | Foundation Models |
What this means strategically
Anthropic kept Mythos restricted, but lower-priced
The Mythos Preview was $25/$125. Fable 5 at $10/$50 cuts cost more than half while opening it to all developers. Anthropic is monetizing the safety advantage rather than the cyber capability.
OpenAI’s response will probably come fast
With Fable 5 leading SWE-Bench Pro by 20+ points, OpenAI’s next model (GPT-5.6 or “Daybreak 2”) almost certainly ships before Q3 2026. The confidential S-1 filing creates urgency.
Apple’s Gemini-Siri bet looks better
By licensing Gemini rather than building their own, Apple sidesteps the frontier model arms race. With Fable 5 raising the bar yet again, Apple’s “buy frontier intelligence” strategy looks prudent.
June 22 usage-credits change
Anthropic confirmed that on June 22, 2026, the new usage-credits system kicks in for claude.ai Pro/Team/Enterprise subscriptions. Fable 5 will cost more credits per request than Opus 4.8 — budget accordingly. See: Anthropic Claude Code June 15 billing change.
Related reading
- What is Claude Fable 5?
- Should you upgrade from Opus 4.8 to Fable 5?
- Claude Fable 5 cybersecurity restrictions explained
- Anthropic IPO vs OpenAI IPO race
- Dynamic Workflows 1000-subagents cap
Sources
- Anthropic Newsroom: Claude Fable 5 and Claude Mythos 5 (June 9, 2026)
- TechCrunch: Anthropic’s Claude Fable 5 is a version of Mythos the public can access today (June 9, 2026)
- Reuters: Anthropic rolls out public version of Mythos without cybersecurity capability (June 9, 2026)
- Bloomberg: Anthropic releases Mythos-like model without cyber capabilities (June 9, 2026)
- llm-stats.com: Claude Fable 5 Review, Benchmarks and Pricing
- Vellum AI: Claude Fable 5 & Mythos 5 Full Benchmark Breakdown
- CSO Online: Anthropic releases Mythos-class Fable 5 model with safeguards for cyber risks
- The Guardian: Anthropic releases ‘safe’ version of Claude Mythos AI model to public