Anthropic launches Claude Opus 4.5 for coding and agents

If you apply for a performance engineering job at Anthropic, you get handed a brutal two-hour take-home exam. When the company fed that exact test to its newest model, Claude Opus 4.5 outscored every single human candidate who had ever taken it.

Anthropic officially rolled out Claude Opus 4.5 on Monday, completing its 4.5 lineup alongside Sonnet and Haiku. The flagship model is now live across the Claude web and mobile apps, the API, and major cloud providers including AWS Bedrock.

Breaking the 80% barrier

Opus 4.5 is officially the first model to cross the 80% threshold on SWE-bench Verified, landing at an impressive 80.9%. It also scored 80.7% on the multimodal benchmark MMMU, setting a fresh benchmark across coding, visual understanding, and complex tool use.

Beyond raw scores, the model shows a strange knack for lateral problem solving. In one customer service benchmark scenario, models were asked to help a traveler modify a non-refundable basic economy ticket. The benchmark expected a straight refusal based on airline policy.

Opus 4.5 went around the rule instead: it upgraded the customer's cabin class first, then processed the flight change. The benchmark technically marked it as a failure because the behavior was unexpected, but Anthropic highlighted it as proof that the model can navigate messy real-world tradeoffs without hand-holding.

The company noted that this kind of boundary-skirting can veer into reward hacking if left unchecked, pointing to upgraded alignment and industry-leading prompt injection defenses built into the release.

Smarter agents, cheaper tokens

Flagship capability usually commands flagship prices, but Anthropic cut the cost of Opus-class intelligence dramatically. At $5 per million input tokens and $25 per million output tokens, AWS confirmed this puts Opus 4.5 at one-third the price of Opus 4.1.

Developers also get a new effort parameter in the API to trade speed for depth. Set to medium effort, Opus 4.5 matches Sonnet 4.5's best coding score while burning 76% fewer output tokens. At maximum effort, it beats Sonnet by 4.3 percentage points while still using 48% fewer tokens.

The tooling around the model got serious attention too. Claude Code now features a Plan Mode that asks clarifying questions upfront and generates an editable plan.md before running code. The desktop app now supports running multiple local and remote agent sessions in parallel.

Consumer apps also solved a long-standing annoyance: endless chats. Paid Claude users will no longer hit context limit walls, as the engine now quietly compresses older conversational context in the background. Claude for Chrome expanded to all Max users, and Claude for Excel opened up across Max, Team, and Enterprise tiers.

Why it matters

We think Anthropic just made long-horizon, autonomous agents practical for actual production setups. For months, developers were forced to compromise, keeping expensive reasoning models on a leash. Cutting Opus pricing by two-thirds and raising usage limits means software teams can finally use a top-tier model as an orchestrator commanding cheaper sub-agents like Haiku 4.5.

Arriving right on the heels of OpenAI's GPT 5.1 and Google's Gemini 3, Opus 4.5 shifts the competitive battleground from raw conversational trivia to end-to-end office and software engineering workflows. Features like dynamic tool search and explicit planning files show Anthropic understands enterprise pain points better than most.

Still, the airline ticket workaround should give enterprise security teams pause. That brand of unscripted ingenuity looks like magic when writing React code, but it looks an awful lot like dangerous overreach when applied to financial ledgers or internal access controls. As Anthropic gives Claude more rein over desktops and browsers, guardrails will need to evolve just as fast as the model's reasoning.