Anthropic released Claude Opus 5 on July 24, 2026, a new flagship model in the Opus tier. The company positions it as approaching the frontier intelligence of Claude Fable 5 at half the price, while keeping rates level with Opus 4.8 — 5 USD per million input tokens and 25 USD per million output tokens.
Key takeaways
- Released July 24, 2026. Pricing unchanged from Opus 4.8: 5 USD/M input, 25 USD/M output.
- Frontier-Bench v0.1: doubles Opus 4.8's score at a lower cost.
- CursorBench 3.2: within 0.5% of Fable 5 at half the price.
- Default model on Claude Max, strongest on Claude Pro — available via API, Claude.ai, Claude Code and Claude Cowork.
- Not subject to the 30-day data retention that applies to Fable and Mythos models.
A new peak for the Opus tier
Anthropic describes Opus 5 as a "thoughtful and proactive" model that comes close to the frontier intelligence of Claude Fable 5, but at half the cost. It is the first major update to the Opus tier since Opus 4.8, which launched on May 28, 2026.
In June 2026 the company shipped Mythos 5, Fable 5 and Sonnet 5. Haiku is still awaiting its version 5 upgrade.
The positioning is clear: Opus 5 is the default model on the Claude Max plan and the strongest model available on Claude Pro. Anthropic offers it through the API, the Claude.ai app, Claude Code and Claude Cowork. The company also claims stronger visual output generation and a greater tendency for the model to verify its own work.
Results: near Fable 5 at a fraction of the cost
The strongest message from the benchmarks is the quality-to-price ratio: Opus 5 gets close to Fable 5 at a much lower cost. The jump is also clear in scientific tasks and on the ARC-AGI reasoning test.
| Benchmark / area | Opus 5 result |
|---|---|
| Frontier-Bench v0.1 | 2x Opus 4.8's score, lower cost |
| CursorBench 3.2 | within 0.5% of Fable 5, half the price |
| OSWorld 2.0 | beats Fable 5 at one-third the cost |
| ARC-AGI 3 | 3x the next-best model |
| Organic chemistry | +10.2 pts vs Opus 4.8 |
| Protein tasks | +7.7 pts vs Opus 4.8 |
Customers who tested the model before launch stress the same point — Fable-level performance at Opus pricing.
Claude Opus 5 approaches Fable-level performance at half the cost.
— a quote from a customer testing the model, the CEO of Cognition, maker of the Devin coding assistant. Cursor's co-founders offered similar assessments ("near Fable 5 intelligence at Opus speed and cost"), as did Lovable ("the biggest leap in the Opus family since 4.5").
Safety: best alignment, hard limits on cyber
Anthropic claims Opus 5 is the best-aligned model in its history — scoring 2.3 on an automated behavioral audit?Behavioral audit: An automated test measuring how often a model acts against its creators' intent. A lower score means less misalignment. and adhering most closely to Claude's "Constitution".
At the same time, the company keeps cyber restrictions: the model cannot scan binaries for vulnerabilities, though it may analyze source code. Safety classifiers are expected to fire about 85% less often than with Fable 5, cutting false blocks on legitimate tasks.
In offensive cybersecurity and biology research, Opus 5 still trails Mythos 5 — a model Anthropic deliberately keeps stronger in these sensitive areas.
Pricing and API additions
For API users, Anthropic is adding a beta feature called "Automatic Fallbacks" — automatically switching to another model when the primary one is unavailable. Unlike Fable and Mythos, Opus 5 is not subject to 30-day data retention?Data retention: How long a provider stores API inputs and outputs. No retention eases compliance for enterprises., which matters for companies with strict information-retention rules.
Why it matters
Opus 5 shows how quickly the gap is shrinking between the most expensive "frontier reasoning" and mid-tier models. If Anthropic's claims hold, the Opus tier catches up to Fable 5's capabilities at half the cost.
That reshapes the economics for companies building long-running agents that make thousands of model calls. For those use cases, the price per token matters more than a single benchmark point.
The focus on self-verification and iteration signals that Anthropic is tuning the model for real agentic workflows. It answers pressure from cheaper rivals — Chinese open-weight models and GPT and Gemini — that have competed on price-to-quality for months.
What's next?
- Haiku is still awaiting its version 5 — the missing piece of Anthropic's 5 model family.
- No date has been given for the full rollout of "Automatic Fallbacks" (currently in API beta).
- Mythos 5 remains stronger in offensive cyber and biology, marking the boundary of Opus 5's use.
Sources
- Anthropic — Introducing Claude Opus 5
- TechCrunch — Anthropic launches Opus 5
