Anthropic released Claude Opus 5 on July 24, 2026, positioning it plainly: it approaches the intelligence of flagship model Fable 5 on many tasks, at half the price. If you remember the safety-routing system we covered in our piece on Claude Fable 5 — high-risk requests get automatically redirected to Opus 4.8 — this piece covers a new wrinkle: that routing setup just got updated again with Opus 5's launch.
Opus 5 vs. Opus 4.8: Same Price, a Big Jump in Capability
Here's the most direct point first: Opus 5 is priced identically to its predecessor, Opus 4.8 — $5 per million input tokens and $25 per million output tokens, with a Fast mode available at twice the base price. Anthropic's own framing is that Opus 5 “provides greatly improved performance for the same cost as its predecessor.” In other words, this isn't a price hike in exchange for more capability — it's the same budget buying a meaningfully better model.
Opus 5 is now the default model on Claude Max and the strongest model available on Claude Pro, while also being available via the API (model ID claude-opus-5) and integrated into Claude.ai, Claude Code, and Claude Cowork.
The Benchmarks Anthropic Named: Coding, Knowledge Work, Visual Output
The evaluations Anthropic cited on release are fairly wide-ranging: Frontier-Bench, CursorBench 3.2, ARC-AGI 3, Zapier AutomationBench, OSWorld 2.0, GDPval-AA, and OSS-Fuzz — the last of which specifically tests vulnerability discovery and exploitation. Anthropic emphasizes that Opus 5 is stronger at verification and iterative problem-solving, particularly excels at software engineering and knowledge-work tasks, and produces stronger visual output as well.
What Does “Still Behind Mythos 5” Mean? First, Who — or What — Is Mythos
One line in the official release post is worth zooming in on: Opus 5 is now the new state-of-the-art on most evaluations, but it “remains behind Mythos 5” specifically on cybersecurity and biology-research tasks. You might be wondering who or what Mythos 5 even is. We checked Anthropic's own model page, and the answer turns out to be simpler than it sounds: Mythos 5 and Fable 5 are the same underlying model. Fable 5 is the public version with a full safety-safeguard layer applied on top; Mythos 5 is the version with that layer stripped away and capability left uncapped, currently offered only to a small set of testing partners for cybersecurity work (with biology research access said to be coming soon) — it is not a product the general public or enterprises can sign up for.
This design logic is a continuation of the same safety-routing approach we covered when Fable 5 launched: the same vulnerability-finding capability has to be filtered, in the public-facing version, so the part of it that could be used for harm doesn't ship to everyone. Anthropic's own language is direct here too — Mythos 5 “could be used both for good and for harm” given how capable it is on cybersecurity and biology research, so for now it's offered only through a trusted-access program to a small, but growing, set of customers.
So how big is the actual gap between Opus 5 and Mythos 5? Based on the OSS-Fuzz results Anthropic disclosed, Opus 5's ability to find vulnerabilities is already close to Mythos 5's — but the step of actually turning a vulnerability into a working exploit is where Opus 5 falls substantially behind. Anthropic explains this as a deliberate training choice: as with Opus 4.8, they intentionally avoided training Opus 5 on cyber-offense tasks, so this capability gain is purely a byproduct of the model getting generally smarter, not something deliberately cultivated. Biology research follows a similar pattern: Anthropic says Opus 5 still shows meaningful limitations on long-running, autonomous research tasks — which is exactly where they expect AI models to pose the most substantial biology-related risk — and states plainly that “Mythos 5 remains the stronger model” for that kind of biological work.
A Detail That Extends Our Earlier Piece: The Routing Rules Got Updated Again — and There Are Two of Them, Moving in Opposite Directions
If you read our earlier piece on Fable 5, you'll remember the routing mechanism it described: high-risk requests get redirected from Fable 5 to Opus 4.8. With this Opus 5 release, Anthropic actually disclosed two separate routing rules, and they point in opposite directions — worth separating out clearly here rather than lumping them together.
The first is Opus 5's own built-in cyber classifiers. Anthropic explains that Opus 5's cyber classifiers are proportionally less restrictive than the equivalent mechanism on Fable 5 — they let Opus 5 find vulnerabilities in source code, but still block “binary-based” vulnerability scanning (a method Anthropic says is more often associated with malicious actors), penetration testing, and exploit generation. Based on Anthropic's own testing, they expect these classifiers to intervene roughly 85% less often than the equivalent ones on Fable 5. But here's the key part: on Claude.ai, Claude Code, and Claude Cowork, any request that does get flagged falls back to Opus 4.8 by default — and that same fallback can be enabled on the API as well. If your team is already a vetted cybersecurity research outfit, Anthropic offers a separate channel called the Cyber Verification Program (CVP), through which enrolled enterprises and researchers get direct access to a version of Opus 5 with fewer restrictions.
The second rule is the one specific to biology-related requests: requests that get blocked on Fable 5 for biology-related reasons now route to Opus 5, rather than to Opus 4.8 as before. The reason is that Opus 5's biology safeguards are similar to Opus 4.8's, which makes it, in Anthropic's telling, now the most capable generally available model for scientific research — so Anthropic upgraded this particular fallback path from Opus 4.8 up to Opus 5.
Notice that these two rules point in exactly opposite directions — a flagged cybersecurity request falls back “down” to Opus 4.8, while a flagged biology request gets routed “up” to Opus 5. For developers and enterprises who already have Claude API integrations, this is worth adding to your monitoring checklist rather than assuming every flagged request lands on the same model. We made this point in our Fable 5 piece and it's worth repeating here: don't just look at the total bill — where possible, log which model version actually handled each request, so you don't misread your budget estimates or performance monitoring. This is our own practical suggestion based on the routing rules Anthropic disclosed, not an operating procedure Anthropic itself requires.
Alignment: Anthropic's Own Claim to Be Its “Least Deceptive” Model Yet
Beyond performance and safety-routing, Anthropic also disclosed an alignment data point: in its automated behavioral audit, Opus 5 scores 2.3 on overall misaligned behavior — the lowest of Anthropic's recent models — and the company describes it as their most aligned model to date, with the lowest rates of deceptive behavior. This figure is a result Anthropic disclosed and evaluated itself; we have no third-party data to cross-verify it, so we're presenting the official claim as-is here without additional endorsement or extrapolation.
What This Means in Practice for Enterprise Procurement and Token Budgeting
Pulling the pieces above together, there are at least three practical implications worth noting for any enterprise currently doing model selection or procurement evaluation.
First, if you'd previously split work by budget — reserving important tasks for Fable 5 and routing lower-priority work to Opus 4.8 — it's worth re-examining that split now that Opus 5 offers meaningfully more capability at the same price as Opus 4.8. Some tasks that used to require Fable 5 may now be handled adequately by Opus 5, trimming a real chunk of cost.
Second, safety-routing rules aren't a one-time setting — they keep getting adjusted as new models ship. This time it was the biology-related fallback moving from Opus 4.8 to Opus 5; there's no reason to assume it won't be updated again in the future. If your product sits anywhere near this kind of sensitive-content boundary, it's worth adding “periodically check whether Anthropic's official routing rules have changed” to your routine maintenance list, rather than treating a one-time setup as permanent.
Third, Mythos 5 — this access-restricted top-tier model — isn't something most enterprises can get today, but its existence itself is a signal: there's a deliberate safety buffer built between the ceiling of what a model can technically do and what ordinary users are actually given access to. When evaluating any AI vendor, rather than only asking “what's your most capable model,” it's worth also asking “how far apart is the version I can actually get from the version you're technically capable of building” — that framing tends to be a more useful way to assess risk boundaries.
FAQ
How should I choose between Claude Opus 5 and Claude Fable 5?
If your budget is limited and your tasks don't require Fable 5's top-tier capability, Opus 5 approaching Fable 5's performance at half the price is likely the better value. If your tasks demand very high capability and budget allows, Fable 5 remains the highest-capability option among currently public models. In practice, it's worth benchmarking both against your own actual tasks rather than relying solely on Anthropic's published benchmark numbers.
Can the general public sign up to use Mythos 5?
Not currently. Per Anthropic's own description, Mythos 5 is offered only through a trusted-access program to a small set of testing partners doing cybersecurity work (with biology-research access said to be coming soon) — it isn't a public product that ordinary users or enterprises can directly apply for.
Is Opus 5 more expensive than Opus 4.8?
No. Anthropic states the two are priced identically — $5 per million input tokens and $25 per million output tokens, with Fast mode at twice the base price. There's no price increase here.
What should I watch for if my application touches biology- or cybersecurity-related content?
Anthropic updated two routing rules that point in opposite directions this time: biology-related requests that get blocked on Fable 5 now route to Opus 5 (rather than the previous Opus 4.8), while requests flagged by Opus 5's own cyber classifiers fall back by default to Opus 4.8 on Claude.ai, Claude Code, and Claude Cowork. If your product sits near this kind of boundary, it's worth logging which model version actually handles each request and periodically checking Anthropic's latest routing rules — don't assume every flagged request lands on the same model.
Can Opus 5's “alignment score” be taken at face value?
This is a result from Anthropic's own automated behavioral audit — a first-party disclosure with no third-party data available to cross-verify it. It's worth treating as the vendor's own claim rather than an independently verified conclusion.
Source
This piece's topic was triggered by a YouTube video, “Claude Opus 5 in 8 Minutes” (channel: Developers Digest, published within a few days of Anthropic's official announcement). The writer did not watch this video or obtain its transcript, so this piece contains no retelling of anything from the video's footage or narration. This piece also did not use an automated view-count scraping pipeline for this topic selection, so no view count is reported here rather than substituting an estimate. Every specific fact in this piece — pricing, availability, benchmark names, the Mythos/Fable relationship, the routing-rule update, and the alignment score — was verified directly against Anthropic's official release post (anthropic.com/news/claude-opus-5) and official model page (anthropic.com/claude/mythos), by directly retrieving and checking the full raw HTML of both pages, not by relying on the video or any secondhand account.
