Anthropic says Opus 5 can nearly match its top-performing model for half the price

Antrhopic

Just arsenic it was starting to consciousness for illustration Anthropic had begun to neglect its Opus model, nan institution has released a caller type of its one-time flagship LLM. According to Anthropic, Opus 5 offers capabilities that travel adjacent to Fable 5, Anthropic's astir powerful commercially disposable model, astatine half nan price. More connected pricing successful a bit, but arsenic for performance, Anthropic says Opus 5 is amended astatine knowledge activity than immoderate of its past models. It's besides substantially amended astatine analyzable coding tasks than Opus 4.8. Anthropic claims it's amended for technological investigation too, wherever it excels astatine tasks for illustration predicting really variations successful a macromolecule series mightiness impact nan measurement nan molecules function.

At nan aforesaid time, Anthropic says it has managed to make nan exemplary much resistant to being tricked. Additionally, nan institution notes Opus 5 "exhibits nan lowest rates of deceptive behavior." So hooray for that. One important favoritism betwixt Opus 5 and Mythos, which is presently only disposable to a constricted number of vetted organizations done Anthropic's Project Glasswing initiative, is that nan institution has specifically avoided training nan caller exemplary connected cyber-related tasks. Due to much its powerful capabilities, Opus 5 is broadly amended astatine those tasks than its predecessor, making it much useful for uncovering cybersecurity vulnerabilities, but nan institution says Opus 5 is "substantially behind" its flagship exemplary astatine exploiting those vulnerabilities.

With today's release, Anthropic expects Opus 5 will go nan exemplary astir group will scope for erstwhile they request to get immoderate activity done, and it's easy to spot why nan institution believes that. One of nan frustrations group had pinch Fable were nan safeguards nan institution imposed. Anthropic implemented a strategy wherever Claude would way prompts related to immoderate topics done to Opus 4.8 alternatively of Fable 5. Many felt nan restrictions were heavy-handed, and made Fable consciousness unusable for immoderate tasks. To that point, Anthropic says "Claude Opus 5's safeguards are designed to let beneficial uses of nan exemplary successful some cybersecurity and biology." The institution has strengthened immoderate of nan model's cyber-related guardrails, but notes it did truthful on a "narrow range" of circumstantial tasks. "Based connected our testing, we expect nan classifiers to intervene astir 85 percent little often than they do for Fable 5," Anthropic said. With Opus 5, Anthropic besides isn't including it successful its precocious announced 30-day information retention policy, which nan institution introduced alongside Fable and Mythos 5.

As for pricing. Anthropic says API costs for Opus 5 will stay astatine $5 per 1 cardinal input tokens and $25 per 1 cardinal output tokens. Anthropic has besides added an "effort" paper for Opus that users tin tweak to show nan exemplary whether they want it to beryllium much thorough aliases accelerated and businesslike to conserve tokens. Had nan institution announced nan caller exemplary earlier China's Moonshot came retired pinch its Kimi K3 model, which costs $15 per 1 cardinal output tokens and offers Fable-like capacity connected immoderate tasks, nan guidance to Opus 5 would astir apt beryllium very positive. Instead, I'm judge group will grumble astir pricing. That said, entree to Opus 5 is included pinch each of Anthropic's paid tiers. Right now, you can't motion up for 1 of Kimi's subscriptions plans, moreover if you wanted.