Zach Anderson
Feb 05, 2026 18:43
Anthropic releases Claude Opus 4.6 with 1M token context window and agent groups characteristic, days after $350B tender supply valuation shook tech shares.
Anthropic dropped Claude Opus 4.6 on Thursday, a mannequin improve that quintuples its context window to 1 million tokens and introduces autonomous multi-agent collaboration—arriving simply at some point after the corporate’s $350 billion tender supply valuation triggered a selloff throughout tech shares.
The timing issues. Traders spooked by AI competitors dumped shares on February 4, and now Anthropic is displaying precisely why it instructions that valuation: Opus 4.6 outperforms OpenAI’s GPT-5.2 by 144 Elo factors on GDPval-AA, a benchmark measuring economically useful information work in finance, authorized, and technical domains.
What Really Modified
Three upgrades stand out for enterprise customers.
The 1M token context window (in beta) represents a 5x improve from Opus 4.5’s 200,000 tokens. On MRCR v2—a needle-in-a-haystack retrieval check—Opus 4.6 scores 76% in comparison with Sonnet 4.5’s 18.5%. That is not incremental enchancment; that is a unique functionality class for document-heavy workflows.
Agent Groups in Claude Code lets builders spin up a number of AI brokers working in parallel. Early entry accomplice Invariant Labs reported Opus 4.6 “autonomously closed 13 points and assigned 12 points to the appropriate staff members in a single day, managing a ~50-person group throughout 6 repositories.” The mannequin dealt with each product and organizational choices whereas realizing when to escalate to people.
For enterprise integration, Anthropic added PowerPoint help (analysis preview) alongside upgraded Excel capabilities. The mannequin can now ingest unstructured knowledge, infer construction with out steerage, and execute multi-step modifications in a single move.
Benchmark Efficiency
Opus 4.6 leads on Terminal-Bench 2.0 for agentic coding and tops Humanity’s Final Examination, a multidisciplinary reasoning check. It additionally beat each different mannequin on OpenAI’s BrowseComp, which measures skill to find hard-to-find data on-line.
The 190 Elo level enchancment over its personal predecessor on GDPval-AA suggests significant progress on duties that truly generate income—monetary evaluation, authorized assessment, technical documentation.
Pricing holds regular at $5/$25 per million tokens for enter/output. Premium pricing ($10/$37.50) applies for prompts exceeding 200k tokens.
Security Claims and Aggressive Positioning
Anthropic says Opus 4.6 exhibits “the bottom fee of over-refusals of any latest Claude mannequin” whereas sustaining alignment similar to Opus 4.5. The corporate added six new cybersecurity probes to detect potential misuse, acknowledging the mannequin’s enhanced capabilities reduce each methods.
The discharge positions Anthropic immediately towards OpenAI within the enterprise AI race. With Google’s Gemini and xAI’s Grok additionally competing for enterprise contracts, Opus 4.6’s multi-agent options and prolonged context window characterize Anthropic’s wager that autonomous AI workflows—not simply chatbots—will outline the subsequent section of enterprise adoption.
Opus 4.6 is out there instantly on claude.ai, the API, and main cloud platforms. Builders can entry it by way of the claude-opus-4-6 mannequin identifier.
Picture supply: Shutterstock
