The honest take upfront: Claude Opus 4.6 is still the better creative writer. If you’re generating long-form content, prose, or anything where tone and originality matter, 4.6 punches harder per token. That’s not a hot take — it’s just what you notice after working with both.
So why am I writing about 4.7? Because the teammates feature genuinely changes how I think about multi-step AI workflows.
What the Teammates Feature Actually Is
Opus 4.7 ships with first-class support for agent-to-agent collaboration. Think less “one model doing everything” and more “a team of specialized agents handing off work with context intact.” Each agent in a teammates setup can see what the others have done, flag blockers, and delegate subtasks — without you stitching it together manually with prompt engineering hacks.
In practice, this means you can set up a research agent, a drafting agent, and a review agent, and have them work through a task in sequence — or in parallel — with Opus 4.7 orchestrating the hand-offs. The coordination overhead that used to live in your prompt now lives in the model’s architecture.
The Token Cost Is Real
Let me not pretend otherwise. Opus 4.7 is significantly more token-hungry than 4.6. For single-agent, single-turn tasks — writing a blog post, summarizing a document, generating copy — you’re paying a premium for capabilities you’re not using. That’s a bad trade.
But for workflows where the task itself spans multiple agents, tools, or reasoning steps, the math flips. The cost of coordinating agents manually (engineering time, brittle prompts, context loss between hand-offs) often exceeds the raw token delta. 4.7 absorbs that coordination cost into the model itself.
Where I Actually Use Each
I reach for Opus 4.6 when:
- The task is self-contained and creative
- I want the most expressive, nuanced output possible
- I’m running a single-shot generation where agent coordination is irrelevant
I reach for Opus 4.7 when:
- The task spans multiple steps with hand-offs
- I’m building a Claude Code agent team (sprint orchestration, parallel research, review loops)
- Context continuity between agents matters more than per-token creative quality
Why This Distinction Matters for Builders
The failure mode I see most often: people defaulting to the latest model number as though it’s a strict upgrade. It isn’t. 4.7 is a different trade-off, not a universal improvement over 4.6. Anthropic didn’t deprecate 4.6 for a reason — it still has a distinct use case.
If you’re building a product where users need a single, high-quality creative output, 4.6 is likely your better default. If you’re building an agentic pipeline — something with orchestration, delegation, and multi-step reasoning — 4.7’s teammates feature is genuinely worth the token cost.
How to Actually Switch Between Them
Both models are available simultaneously — Anthropic doesn’t force a migration when a new model releases. In Claude Code specifically, you have a few ways to pick:
- CLI flag:
claude --model claude-opus-4-6orclaude --model claude-opus-4-7 - In-session command:
/modelopens a picker mid-conversation - Settings file: set
"model": "claude-opus-4-6"insettings.jsonfor a persistent default /fastmode: exclusively runs Opus 4.6 at 2.5× speed — higher token cost, but if you want 4.6 with less latency, this is your lever
Via the API it’s just the model ID: claude-opus-4-6 or claude-opus-4-7.
The Bigger Shift
What Opus 4.7 signals to me is less about any single feature and more about where Anthropic sees AI going. The focus on agent collaboration, on models that work well with each other rather than just with humans, is a meaningful architectural bet. Whether that bet pays off depends on how quickly the tooling ecosystem catches up.
For now, I’m using both — and I’m a lot more deliberate about which one I reach for.