v0 taps Claude Opus 4.6 to level up code generation

If you have used frontier AI models for complex projects over the past year, you know the frustration of context rot. After a dozen turns, the model forgets its core constraints, confuses variable names, and starts hallucinating.

Anthropic is tackling that friction head-on. Just three months after releasing Opus 4.5 in November, the company launched Claude Opus 4.6 on Thursday. The update pushes Opus into the 1M-token context tier while expanding far beyond software development into professional finance and multi-agent coordination.

Multi-agent teams and 1M context

The marquee architectural upgrade in Opus 4.6 is what Anthropic calls agent teams. Instead of a single agent stumbling through a massive list of linear instructions, Claude Code and API users can now deploy teams of agents that divide large jobs, coordinate directly, and run subtasks in parallel. Anthropic product head Scott White compared the setup to directing a full team of human specialists.

Under the hood, Opus 4.6 also introduces a 1M-token context window in beta. In needle-in-a-haystack stress tests like the 8-needle 1M MRCR v2 benchmark, Opus 4.6 scored 76%, demolishing Sonnet 4.5's 18.5%. The API also adds native prompt compaction and adaptive thinking controls to prevent runaway latency and cost.

Benchmark scores followed suit. Opus 4.6 captured top marks on Terminal-Bench 2.0 for agentic coding and led frontier models on Humanity’s Last Exam. In GDPval-AA, which gauges economically valuable knowledge work, it outperformed OpenAI’s GPT-5.2 by 144 Elo points and its predecessor by 190 points.

Targeting Wall Street workflows

Anthropic is actively expanding Claude from a developer tool into an indispensable financial analyst. White pointed out that non-engineers, including product managers and financial professionals, have flooded into Claude Code simply because it gets hard tasks done.

To meet them where they work, Anthropic upgraded Claude in Excel to support pivot tables, custom charts, conditional formatting, and financial-grade layouts. It also rolled out Claude in PowerPoint as a beta preview side panel, letting analysts generate and tweak slides natively without copying back and forth.

For broader autonomous work, the desktop Cowork feature now grants Claude permission to read, edit, and create documents directly inside selected local folders. With plugins, it can automate variance analysis, reconciliations, and journal entries. In internal testing across roughly 50 investment banking and corporate finance tasks, Opus 4.6 beat Sonnet 4.5 by more than 23 percentage points. On external benchmarks from Vals AI, it posted a chart-topping 60.7% on SEC filing research and 76.0% on TaxEval.

Opus 4.6 is live on claude.ai, major cloud hosts, and the API at an unchanged $5 per million input tokens and $25 per million output tokens.

Why it matters

For knowledge workers and enterprise teams, Opus 4.6 shifts AI from a passive copy-paste assistant to an active coworker. The ability to deploy parallel agent teams across a 1M-token context means you can feed in an entire repository or years of financial filings without the model forgetting where it started. Having native Excel and PowerPoint tooling reduces friction for non-technical analysts who need client-ready output on the first pass.

For the AI landscape, the battle lines have moved from raw chat intelligence to agentic reliability in enterprise workflows. Anthropic's multi-agent approach lands directly against OpenAI’s new Frontier platform and GPT-5.3-Codex. Winning this phase is about making models reliable enough to execute multi-hour, multi-step business logic without wandering off course.

We think Anthropic's hyper-focus on high-value finance and corporate work is brilliant. It directly monetizes white-collar operational spend where budgets are deep and patience for basic hallucinations is zero. However, handing autonomous file-editing permissions to multi-agent groups inside local directories carries real organizational risk. When autonomous agents coordinate among themselves, spotting a subtle financial error buried in a complex spreadsheet model will require serious human oversight.