Claude Opus 5 Is Here: Near-Fable Intelligence at Half the Price, Now on MuiRouter

On July 24, 2026, Anthropic released Claude Opus 5. The official positioning: it may sit slightly below Fable/Mythos on raw capability, but the price is more approachable. Over time it may act as the day-to-day executor under Fable. Anthropic calls it a more proactive, self-checking model that approaches Claude Fable 5 on coding and knowledge work, at half Fable 5’s price — the same $5 / $25 per million tokens as Opus 4.8.

For everyday developers, every Anthropic update is worth watching. For the past few months, Fable was strong but expensive, scarce, and often dinged by safety classifiers; after Fable shipped, talk of Opus 4.8 “getting dumber” got loud. Opus 5 finally offers some hope: same price, big capability step.

MuiRouter now supports claude-opus-5. Same API key, same /v1/chat/completions and /v1/messages — change the model name and try it.

Anthropic update summary

From Anthropic’s Claude Opus 5 announcement and system card:

• Model ID: claude-opus-5, available on the API, Claude Code, and the Claude Platform.
• New default on Claude Max, and the strongest model currently available on Claude Pro.
• Pricing matches Opus 4.8: $5 / 1M input tokens, $25 / 1M output tokens.
• Optional Fast mode: ~2.5× default speed at 2× base price.
• Capability focus: long-running agents, software engineering, knowledge work, and proactive verify-and-iterate. Unlike Mythos, it is not cyber-offense–centric, so it does not need a “partners only” release.

Anthropic has been explicit: on cybersecurity exploitation tasks, Opus 5 is intentionally kept behind Mythos 5.

Benchmarks and cost: lower cost per finished task

Anthropic is not selling a cheaper sticker price per token. It is selling lower cost to finish the same class of work. That has been the market race across major labs recently.

In short: list price unchanged, official “effective output per unit cost” curve clearly up. We hope that holds in practice.

Early community takes: cost-performance is the main theme

One day after launch, public discussion and builder circles are split: some say it rivals Fable; others disagree. In our experience, that usually tracks the agent stack and its built-in harness, not only the base model.

OpenResty’s creator 春哥 (Zhang Yichun) rates Opus 5 highly — in his own evaluation framework, it can even beat Fable. Independent developer Baye takes the opposite view: Opus 5’s thinking is not deep or comprehensive enough, and you still need more effort to manage it.

Across social posts and early trials, a few themes keep recurring:

• “Half price vs Fable.” Fable 5 is about $10 / $50 on the API; Opus 5 is $5 / $25. CursorBench at max effort is claimed within ~0.5% of Fable’s peak. For team budgets, that is a re-pricing signal, not a cosmetic rename.
• Stronger self-check loops. Early stories include opening desktop and mobile layouts, catching products below the fold and off-screen checkout buttons, then fixing them; or finding the root cause a community patch missed in an open-source package manager. People care about fewer pings and fewer rework rounds.
• Pushback without blind stubbornness. Some refactor reports describe the model arguing a bad design, holding until a single disagreement is clear, then offering a compromise. For agent workflows, judgment beats blind compliance.
• Cooler heads. Still re-run on your own harness — leaderboard #1 is not automatic proof on your contract parsing, trading path, or multi-repo refactor. Fast mode trades 2× list price for speed; price that by task type.

Compress the consensus: Opus 5 is currently the “everyday default Opus” most worth evaluating first.

Pricing snapshot

Standard Anthropic prompt caching still applies: cache hits around 10% of input, 5-minute cache writes around 1.25× input. Same advice as with Sonnet 5: decide on finished-task bills, not sticker rates alone — Anthropic’s own charts are effort-based cost–performance curves for a reason.

Fast mode: ~2.5× speed at 2× base price, fine for interactive coding; long batch jobs can stay on the default tier first.

Safety and safeguards

Anthropic calls Opus 5 among its most aligned models so far: lower rates of misaligned behavior on automated audits, harder to trick into misuse, less likely to take hard-to-reverse reckless actions.

On cyber: vulnerability finding is much stronger and closer to Mythos 5; full exploit development remains far weaker. Classifiers are looser than Fable 5’s — source-level vuln discovery is allowed; binary-oriented scanning, pen-testing, and exploit generation are blocked. Flagged requests can fall back to Opus 4.8; enterprises still have the Cyber Verification Program.

For most product teams, that means “can help with defensive review” without becoming an attack tool for everyone. In practice, it should false-positive user intent less often than Fable 5, and sit closer to everyday engineering.

What this means for developers

• Migrating from Opus 4.8: list price is flat — compare pass rate and tokens on your agent/coding harness. In general this should be fine.
• Vs Fable 5: use Fable when you need extreme long-horizon work and can accept budget and guardrails; for most coding and knowledge work, Opus 5 is better value.
• Vs Sonnet 5: high-volume, patterned tasks can stay on Sonnet; judgment-heavy debugging, architecture, and multi-file refactors fit Opus 5 better.
• Effort and Fast mode: start mid effort to map the curve. 春哥 advises against mindlessly maxing effort — overthinking can hurt quality. As for Fast mode: skip it for now.
• Old skills / over-constrained system prompts: some early notes say the model is more opinionated — clearing stale skills or redundant rails can clean results up.

What this means for MuiRouter users

MuiRouter is built around one idea: one API key, one integration style, model switches as configuration instead of rewrites.

We’ve added claude-opus-5 to the catalog at Anthropic’s published $5 / $25 (including cache prices), with the same markup as other Claude models. Point model at claude-opus-5 (or pick it in the Playground); native Messages and OpenAI-compatible Chat Completions both go through the same gateway.

As always, treat production cutover as done only after an end-to-end check on your account — real routing still depends on Anthropic / gateway availability. The point of a unified gateway is: new model, new default, new cost curve, minimal app-code churn.

Bottom line

Claude Opus 5 is a solid everyday release plus a same-price generational step: official numbers and early community feedback both point to a higher coding/knowledge ceiling and lower cost per finished task. 春哥’s take is another reason the capability jump is worth your attention.

If you are still on Opus 4.8, we recommend upgrading sooner rather than later. If Fable 5’s price or guardrails made you hesitate, Opus 5 is likely the better default tool.

MuiRouter is ready — start now.