Anthropic’s New Claude Opus 4.5 Reclaims The Coding Crown From Gemini 3

Anthropic coming launched nan latest type of its flagship Opus model: Opus 4.5.

The institution calls it its astir intelligent exemplary yet and notes that it is particularly beardown successful solving coding tasks, taking nan crown from OpenAI’s GPT-5.1-Codex-Max and Google’s week-old Gemini 3 exemplary pinch an SWE-Bench Verified accuracy people of 80.9%.

The institution is besides making Opus 4.5 importantly much affordable to use, pinch API pricing of $5 per cardinal input tokens and $25 per cardinal output tokens, down from $15/$75 per cardinal input/output tokens.

Users connected Anthropic’s subscription plans will besides now spot a spot much headroom to usage Opus 4.5.

Benchmarks

With nan launches of OpenAI’s GPT-5.1 and 5.1-Codex-Max, Google’s Gemini 3 (and its deed Nano Banana Pro image model), it’s been a very progressive November for nan ample exemplary builders. Gemini 3, especially, received a very affirmative reception.

Unlike Google, Anthropic has ne'er focused connected image manipulation aliases video creation, but has stuck squarely to its spot successful coding and productivity usage cases. This latest Opus is nary different and Anthropic stresses that nan exemplary tin now nutrient documents, spreadsheets and presentations “with consistency, master polish, and domain awareness.”

But arsenic usual, it’s coding wherever nan Claude models shine. That’s reflected connected nan benchmarks, wherever Opus 4.5 bests nan title crossed nan board, but benchmarks don’t ever bespeak real-world usage cases, of course.

For this release, Anthropic besides gave Opus 4.5 nan aforesaid trial it gives prospective capacity engineering candidates who want to activity astatine nan company. This test, which solely focuses connected method ability, has a two-hour clip limit and Opus 4.5 scored higher than immoderate of Anthropic’s occupation candidates ever did.

As Alex Albert, nan caput of developer relations astatine Anthropic, besides told me, he’s gotten nan consciousness that nan “model conscionable gets it.” He noted that past models person often been very bully astatine gathering information crossed different channels (like Slack and email), but person had a difficult clip synthesizing each of this accusation effectively.

“I’m uncovering that pinch this model, that is nary longer nan case,” he told me. “I tin really spot it to spell consecutive from those Slack messages to a bully output, and past I’m like, wow, it could person really conscionable sent this. I still americium reviewing it and things, but I really could person conscionable been hands-off here.”

Low, Medium, High Effort

One caller characteristic of Opus 4.5 is that it features an “effort” parameter (low, medium, high), akin to immoderate of its competitors’ models, which allows developers to power really overmuch clip (and really galore tokens) nan exemplary will usage to lick a fixed problem. Set to medium, nan exemplary is connected par pinch Sonnet 4.5 connected nan SWE-bench Verified benchmark but uses 76% less tokens, and moreover astatine nan precocious setting, wherever it thumps Sonnet 4.5, it uses only astir half nan tokens of nan Sonnet model.

That’s a inclination we’ve been seeing and this ratio is thing OpenAI besides stressed erstwhile it launched its latest Codex-Max exemplary past week.

Overall, nan exemplary besides improved upon nan remainder of nan Opus family (and Opus 4.1) successful different areas, including ocular reasoning and math.

Opus 4.5 for Computer Use

Opus 4.5 is besides Anthropic’s champion exemplary for machine usage cases yet, nan institution says. To put this to nan test, Anthropic is now opening up its Chrome hold to each Claude Max subscribers (who salary $100/month+).

Computer and browser usage still feels for illustration it’s successful its infancy and often feels alternatively slow and error-prone, but Anthropic is pushing nan authorities of nan creation up a level here, pinch scores good supra those of its erstwhile models.

Anthropic had been successful an absorbing position lately, wherever nan latest type of its mid-tier Sonnet exemplary often outperformed nan older Opus 4.1 model, giving users very fewer reasons to usage nan much costly exemplary successful their day-to-day work. The idea, however, was ever to person a three-tier exemplary and Opus 4.5 restores nan equilibrium here.

“What’s absorbing astir this release, to maine astatine least, is that it is not needfully like: ‘Oh, everybody needs to now displacement to Opus,’ but it does alteration this caller tier of possibilities,” Albert said. “Now we’re entering this scenery wherever you really do person 3 models from america that each fresh different needs on this curve: you person nan Haiku exemplary that we conscionable released a period ago. You person Sonnet 4.5, which was a period and a half ago. And now this completes nan fold.”

Claude Developer Platform Updates

In summation to nan caller model, Anthropic is besides announcing 2 updates to nan Claude Developer Platform that spell hand-in-hand pinch nan Opus 4.5 release: An upgraded scheme mode for Claude Code and Claude Code support successful nan desktop app.

The caller scheme mode now creates much precise plans for really to lick a problem aliases adhd a caller characteristic and sticks to them much directly, Anthropic says.

And if you usage nan Claude desktop app, you tin now commencement coding tasks successful Claude Code connected your desktop aliases successful nan unreality environment. This now lets you tally aggregate section and distant Claude Code sessions successful parallel.

Group Created pinch Sketch.