Anthropic releases Claude Opus 5 at half Fable 5's price

On July 24, Anthropic announced Claude Opus 5, a new LLM positioned as a cost-reduced alternative to Claude Fable 5. The model is available across all platforms—Claude.ai, Claude API (model ID: claude-opus-5), Claude Code, and Claude Cowork—with API pricing at $3 per 1M input tokens and $15 per 1M output tokens, compared to Fable 5's $10 and $50 respectively. The price cut is scheduled through May 2026.

On standard benchmarks, Opus 5 demonstrates competitive performance. In Frontier-Bench v0.1, Opus 5 scores twice as high as Opus 4.8. On CursorBench 3.2, at maximum effort level, the model reaches within 0.5 points of Fable 5's peak score. On ARC-AGI 3, which tests general problem-solving, Opus 5 scores three times higher than Opus 4.8. On OSWorld 2.0, a task measuring real-world interaction with computers, it achieves roughly one-third of Fable 5's best score. Opus 5 also surpasses Opus 4.8 across life-science benchmarks.

The key architectural difference from Fable 5 lies not in raw capability but in how optimization trade-offs are balanced. Fable 5 was trained on Anthropic's internal Mythos model, which embeds strong cybersecurity and safety constraints; input pricing for Fable 5 starts at $10 per 1M tokens, with output at $50. Opus 5 approaches cybersecurity tasks with a different stance: rather than learning safety explicitly, it achieves comparable performance to Fable 5's Mythos baseline through practical expression of capabilities, and avoids outright refusals on safety-sensitive queries. Anthropic notes that where blocking occurs, Opus 5 routes such requests using an automatic fallback.

Anthropic revised its Constitution—Claude's underlying ethical guidelines—across Opus 4.8, Sonnet 5, and Fable 5 to reduce behavioral brittleness; the new version allows for more nuanced responses while maintaining safety. Overall refusal scores dropped to 2.3, making the models less prone to blanket rejection. On safeguards, Opus 5's design is fundamentally similar to Opus 4.8, but Anthropic relaxed the cybersecurity guardrails and added protections against prompt injection, penetration testing, exploit creation, and other attack vectors. For safeguards flagged at Fable 5's level of ~85 points, Opus 5 now routes these requests rather than blocking them outright. On Claude.ai, Claude Code, and Claude Cowork, flagged requests automatically fall back to Opus 4.8; on the API, Opus 5 and Fable 5 automatically redirect flagged requests to the best available model, a feature currently in beta. The fallback mechanism is intended to offer a safety net without sacrificing the model's ability to serve legitimate use cases. For allowed use, Anthropic rolled out a "Cyber Verification Program" to support legitimate cybersecurity research and red-teaming.

Anthropic also introduced a "Fast mode" offering 2.5× faster inference at standard pricing for search previews and benchmarking. On Claude Platform, standard speed incurs 2× cost; Claude Code charges usage credits based on the tool's consumption. When a request is flagged across Claude.ai, Claude Code, and Claude Cowork, it automatically falls back to Opus 4.8. On the API, some flagged requests can route to a fallback model; both Opus 5 and Fable 5 automatically redirect flagged requests to a safer model—a feature in beta on API. For some prompts, users can disable caching to enable fallback on tools switching between compatible versions. Anthropic also published a System Card (PDF) documenting responsibility and safety scaling policies (RSP); it reports that Opus 5 demonstrates "CB-1" capability—meaning it does not independently develop or use security exploits—on existing knowledge, though new exploits ("CB-2") are evaluated conservatively. Alignment to SL-3 is maintained, matching Opus 4.8. In edge cases where the model cannot confirm safe deployment, Anthropic routes requests to safer fallback, or clarifies uncertainty to the user; in such instances, harmful outputs do not result in automatically applied blocking, but Anthropic can update system prompts on claude.ai to manage risk. The model includes guidelines encouraging users to flag safety concerns and, where Opus 5 cannot be relied upon, to escalate to explicit testing or other verification methods.

Anthropic's release of Claude Opus 5 represents a strategic repositioning in the reasoning model market. The model achieves near-Fable 5 performance on benchmarks such as CursorBench 3.2—where Opus 5 at maximum effort scoring reaches within 0.5 points of Fable 5's peak—while cutting API costs in half. This pricing move directly challenges OpenAI's GPT-5.6 Sol, which costs $5 per 1M input tokens and $30 per 1M output tokens.

The safety posture shift signals a notable change in Anthropic's design philosophy. While Opus 4.8 maintained strict guardrails, Opus 5 relaxes constraints from its predecessor, though Anthropic introduced an automatic fallback mechanism to mitigate risk. The fallback routes flagged requests (including those related to cybersecurity tasks) to a safer model, allowing users to deploy the more capable model in sensitive contexts while retaining a safety net. For cybersecurity teams and enterprises, this may unlock use cases previously blocked by stricter safety rules—though the trade-off is that some requests will be automatically diverted rather than answered.

Anthropically notes that on the CursorBench benchmark, where safety guardrails present obstacles to maximum performance, Fable 5 scores approximately 85 points; Opus 5, despite new safeguards, avoids the blocking behavior of Opus 4.8 and reaches comparable ability on the underlying task. The Cyber Verification Program referenced in the announcement suggests Anthropic is investing in structured validation for high-stakes domains.

No comments yet. Be the first to share your thoughts!