Claude Opus 5: What It Changes for B2B Outbound Teams

Anthropic shipped Claude Opus 5 on July 24, 2026. It runs a 1M-token context window by default, thinks by default, and costs exactly what Opus 4.8 cost: $5 per million input tokens, $25 per million output.

If you run Clay enrichment, AI call agents, or any research step that reads messy input and has to make a judgment call, this release is worth an afternoon. Here is what changed, and how we would put it to work.

What Anthropic shipped

Claude Opus 5 (claude-opus-5) sits at the top of the Opus tier, above Opus 4.8, which stays available. From the official release notes:

• A 1M-token context window, as both the default and the maximum. There is no smaller context variant.
• 128k max output tokens.
• Thinking on by default. On Opus 4.8 requests ran without thinking unless you asked for it. On Opus 5 the model decides when and how much to think on each turn.
• Effort is the primary control: low, medium, high, xhigh, max, with high as the default.
• Pricing unchanged from Opus 4.8 at $5 / $25 per million input / output tokens.
• The minimum cacheable prompt drops to 512 tokens, down from 1,024.
• Fast mode is in research preview, on the Claude API only, at $10 / $50 per million tokens.
• Available on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.

Anthropic calls it "a step change improvement for the Opus tier powering long-running agents" (Anthropic). The gains it highlights are in deep reasoning, agentic and long-horizon tasks, long-context work, code review, vision, and multi-agent coordination.

Two behavior changes matter before you swap a model ID. Thinking now counts against max_tokens, which covers thinking plus response text, so revisit that number on any workload that previously ran without thinking. And disabling thinking is now allowed only at effort high or below: pair thinking: {"type": "disabled"} with xhigh or max and the API returns a 400. That is a breaking change from Opus 4.8.

Why this matters for outbound teams

Most model releases do not change outbound work. Three things in this one do.

The context window changes what a research step can be

A serious account brief is a pile of unstructured input: the site, recent job posts, a funding note, review pages, past email threads, whatever the CRM already knows. Today most teams chunk that across three or four chained enrichment steps because no single step can hold it all.

At 1M tokens as the default, the chain becomes optional. One step reads everything and returns a structured answer. Fewer steps means fewer places for the workflow to silently return an empty cell.

Effort turns quality into a budget line

The effort ladder is the part we expect teams to underuse. Anthropic's own guidance is to start at high and move in either direction based on evals, stepping down where quality holds and up for the hardest work.

For outbound that maps cleanly onto list tiers. A "does this company sell to contractors, yes or no" column does not need the same reasoning depth as a personalized angle for a 200-account priority list.

The cache minimum is quietly useful

Enrichment prompts are short and run thousands of times. A 512-token minimum means classification and qualification prompts that were too short to cache on Opus 4.8 can now create cache entries with no code change.

Our read: cost per row did not move, but cost per useful row should, because you can now spend deliberately instead of uniformly. It's the same cost lever Clay just pulled from the other direction with open-weight models in Claygent.

How we'd use it

Four plays we are building into client systems this week.

1. Collapse a research chain into one column. Take an enrichment waterfall that currently runs scrape, summarize, classify, then write, and test it as a single long-context step with all raw inputs passed in at once. Keep the old chain running in parallel for a week and compare fill rate and accuracy before you cut over. Long context removes the excuse for chaining, not the need to verify.

2. Tier effort across the list. Run low or medium effort on the mechanical columns (ICP fit, dedupe, industry normalization) where quality holds, and reserve xhigh or max for the accounts that justify a custom angle. Same budget, better distribution.

3. Write real pre-call briefs for AI callers. An AI call agent is only as good as the context it opens with. A long-context research pass that reads the full account history and returns three specific talking points beats a merge field every time. Bland's July 21 release also added PUT, PATCH, and DELETE to Custom API tools, so an agent can now write back to the system it read from during the same call.

4. Triage replies against the whole thread. Reply handling breaks when the model sees only the last message. Feed the full thread, the CRM record, and the sequence context in one pass, and classify intent, extract objections, and draft the follow-up in a single step. Route the result into HubSpot or GoHighLevel rather than a human inbox queue.

FAQ

How much does Claude Opus 5 cost?

$5 per million input tokens and $25 per million output tokens, unchanged from Claude Opus 4.8. Fast mode, currently a research preview on the Claude API only, is priced at $10 / $50 per million tokens.

What is the difference between Claude Opus 5 and Opus 4.8?

Opus 5 has a 1M-token context window as its default and maximum, 128k max output tokens, and thinking on by default. Migration is a model ID change plus two checks: revisit max_tokens now that thinking counts toward it, and stop disabling thinking at xhigh or max effort, which now errors.

Is Claude Opus 5 worth it for lead research and enrichment?

For deep account research and reply handling, yes, because those are exactly the long-context, multi-step judgment tasks it improved on. For high-volume mechanical columns, test a lower effort level first. The right answer is usually a mix, not one model everywhere.