Claude Fable 5 is Anthropic's flagship model officially launched June 9, 2026, API ID claude-fable-5. Positioned as the 'strongest broadly-available Claude model' — distinct from the invite-only Claude Mythos 5 also released the same day.
Core specs: 1M token Context Window; 128K max output (2× Sonnet 4.6's 64K); Adaptive Thinking always enabled; Extended Thinking not supported; $10/M input, $50/M output.
Positioning vs Opus 4.8: Opus 4.8 is 'the most powerful Opus-tier model for complex reasoning and agentic coding' with knowledge cutoff January 2026. Fable 5 is the highest-capability broadly-available model in the entire Claude family, designed for 'the most demanding reasoning and long-horizon agentic work' — one capability tier above Opus 4.8.
Available platforms: Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, Microsoft Foundry. Bedrock ID: anthropic.claude-fable-5; Vertex AI: claude-fable-5.
Fable 5 doesn't support Extended Thinking — what are the concrete impacts on existing applications? What migration considerations are needed?
This is the most important Fable 5 limitation currently — affecting all applications using Extended Thinking on Sonnet 4.6:
What breaks on Fable 5: API request thinking: {type: "enabled", budget_tokens: N} — this parameter is ignored on Fable 5. No error thrown, but no Extended Thinking effect either.
Applications needing special attention: 'display reasoning steps to users' features — Fable 5's Adaptive Thinking doesn't output a visible thinking process; if your UI has a 'watch Claude think' design, this disappears after switching. Application logic depending on thinking content blocks — if your code parses thinking content blocks from API responses, Fable 5 responses won't have this block. Extended Thinking output used as subsequent step input — some Agent workflow designs use thinking steps as next-step context; Fable 5 doesn't have this output.
Unaffected usage: if you just want 'better answers' without depending on visible thinking process output, Fable 5's Adaptive Thinking usually provides equivalent or better results — just swap API ID.
Fable 5 costs 3.3× Sonnet 4.6 — when is this cost premium worthwhile?
Direct cost calculation (1,000 daily API calls, 3,000 input / 1,000 output tokens average): Sonnet 4.6 ~$216/month; Opus 4.8 ~$360/month; Fable 5 ~$720/month. Fable 5 vs Sonnet 4.6: ~$504/month additional — a substantial gap.
Specific conditions where Fable 5 is worthwhile: task 'success rate improvement' generates significant business value — if each task failure costs a lot (legal compliance review, high-risk architecture decisions, complex multi-step agents), Fable 5's improved success rate saves more in manual remediation than the cost premium. Opus 4.8 repeatedly falls short — if thorough testing on Opus 4.8 confirms quality instability on certain high-difficulty tasks, Fable 5 is a reasonable upgrade. Small volume but high per-task value — few dozen high-priority tasks daily where premium is small vs total business value.
Recommended cost control: route only the 5-10% of requests that genuinely need Fable 5's capability to it; use Sonnet 4.6 or Opus 4.8 for the rest. Typical tiered routing reduces overall costs 60-70% while giving critical tasks the strongest model.
What are the differences between Claude Mythos 5 and Fable 5? What conditions allow access to Mythos 5?
Both models have nearly identical technical specs (1M token context, 128K max output, Adaptive Thinking always enabled, $10/$50 pricing), but completely different access and positioning:
Claude Fable 5: broadly available — anyone with an Anthropic API account can use it directly. Officially available on Claude API, Bedrock, Vertex AI, Microsoft Foundry.
Claude Mythos 5 (claude-mythos-5): via Project Glasswing, primarily for defensive cybersecurity workflows. No self-service access — requires contacting your Anthropic, AWS, or Google Cloud account team to apply. 'Limited availability' on Bedrock and Vertex AI.
Mythos Preview relationship: claude-mythos-preview is Mythos 5's predecessor, still available as a research preview model in Project Glasswing for defensive cybersecurity workflows.
Practical meaning for regular developers: if you're not on Project Glasswing's approved list, Mythos 5 is unavailable. Your top available model is Fable 5. Mythos 5 access is only realistic for specific defensive cybersecurity work with needs submitted through Anthropic partners.
A legal automation services company evaluating whether to upgrade their contract risk analysis workflow from Opus 4.8 to Fable 5. Their task: automatically analyze cross-border M&A contracts (200-400 pages average), identify all risk clauses, evaluate impacts across different legal jurisdictions, generate rigorously-argued risk matrices.
Problems on Opus 4.8: for the most complex contracts (5+ jurisdictions, extensive cross-reference clauses), Opus 4.8 occasionally misses third-tier clause conflicts — not complete failure, but '80-90% accuracy when legal work requires 98%+.'
Fable 5 A/B test results: testing on 50 most complex contracts, Fable 5 improved accuracy in 'identifying cross-jurisdiction clause conflicts' from Opus 4.8's 84% to 93%. The 128K max output also allows more complete risk matrices in one output without segmentation.
Cost decision: ~200 such complex contract analyses per month. Fable 5 additional cost (~$150/month vs Opus 4.8's ~$75/month) vs 30 minutes reduced manual review per contract (lawyers at $400+/hour) gives very clear ROI. Decision: complex cross-border contracts use Fable 5, standard domestic contracts continue on Opus 4.8.
Claude Fable 5's core trade-off: capability ceiling vs cost and limitations. It represents Anthropic's highest broadly-available reasoning capability currently, but with high cost (3.3× Sonnet 4.6) and one important feature limitation (no Extended Thinking). This trade-off means Fable 5's optimal use cases are the intersection of 'task difficulty genuinely exceeds Opus 4.8's ceiling' AND 'application doesn't depend on Extended Thinking's visible output' AND 'cost premium is justified by business value.' Tasks not meeting all three conditions are usually better served by Sonnet 4.6 or Opus 4.8.