Full Breakdown
Claude Opus 5: Anthropic’s Mid-Tier Model Promises Near-Flagship Performance at Half the Cost
7/26/2026, 10:08:23 PM
Core Event – Launch of Claude Opus 5
Anthropic released Claude Opus 5 on July 24. Marketed as an “everyday workhorse,” the model approaches the intelligence of the flagship Claude Fable 5 while charging half the per-token price: $5 per million input tokens and $25 per million output tokens.
Background & Context
Enterprise buyers have been pressing AI providers for cheaper options after token-based bills made frontier-model usage hard to justify. Competitors—including OpenAI, Microsoft, Amazon and Chinese startups—have rolled out lower-priced models, prompting Anthropic to accelerate a cost-focused release. Earlier in 2026 Anthropic briefly withdrew its models to comply with an export-control directive, then reinstated them after two weeks of talks.
Data & Statistics
- Token pricing: Opus 5 – $5 / MTok input, $25 / MTok output; Fable 5 – $10 / MTok input, $50 / MTok output.
- Benchmark scores (Anthropic internal)
- Frontier-Bench v0.1 (coding): Opus 5 43.3 % vs Fable 5 33.7 % vs GPT-5.6 Sol 34.4 %.
- GDPVal-AA (knowledge work): Opus 5 1,861 points, highest among tested models.
- ARC-AGI 3 (reasoning): Opus 5 30.2 % vs Opus 4.8 1.5 % vs GPT-5.6 Sol 7.8 %.
- OSS-Fuzz (vulnerability detection): Opus 5 79.4 % (Mythos 5 80 %); exploitation success 4/14 vs Mythos 5 13/14.
- OSWorld 2.0 (computer use): Opus 5 70.6 % (surpassing Fable 5).
- Safety classifiers: interventions 85 % less frequent than for Fable 5.
Official Statements & Responses
- “We're building and continue to consistently deliver frontier intelligence and bring that as accessibly as possible with every model generation,” — Dianne Penn
- New API features include a Fast mode (?2.5× speed at double price), an Effort setting that trades token usage for depth, and automatic fallbacks that route safety-blocked requests to a less capable model.
Criticism & Opposition
- Reviewers note Opus 5’s verbose default responses require prompt-library rewrites.
- Lenny’s Newsletter called the model “brilliant (but annoying),” echoing complaints about longer outputs and the need to simplify system prompts.
Conflicting Reports & Gaps
- Some outlets claim Opus 5 “more than doubled” Opus 4.8’s Frontier-Bench score and “approaches” Fable 5 on coding benchmarks, while others report Fable 5 still leads on the hardest version of SWE-bench (80 % vs Opus 5 79 %).
- All figures rely on Anthropic’s internal testing; independent verification is limited, leaving day-to-day performance uncertain.
Verbatim Quotes
- “Across your system prompt, skills, and CLAUDE.md files, you may need to simplify just like we did,” — Thariq Shihipar, Claude code engineer
- “We're building and continue to consistently deliver frontier intelligence and bring that as accessibly as possible with every model generation,” — Dianne Penn
What’s Next
Opus 5 is now the default model for Claude Max subscribers and the strongest option on Claude Pro. Anthropic will monitor safety-filter performance and may adjust automated fallback routing based on usage data, with no specific future dates announced.
