Real workflows. Real results.

You were not imagining it. Claude's output quality genuinely degraded between March and April. Responses got shorter, less thorough, and less reliable. Many of us adjusted our workflows, rewrote prompts, and wondered if we were doing something wrong.

Last week, Anthropic published a postmortem explaining exactly what went wrong. Three separate bugs: the default effort level was silently lowered (so Claude was doing less work per request), a caching issue served stale responses, and a routing problem sent some tasks to the wrong processing path. Fortune covered it extensively. These were not your fault. They were Anthropic's.

What matters for operators is what comes next. Anthropic named the bugs, explained how they shipped, and committed to process changes. That is unusually honest for a company at this scale. It does not erase a month of degraded output, but it does set a standard: when the tool breaks, the company tells you why. If you build your operations on a platform, that transparency matters. Our Claude Code guide covers the specific bugs and the workarounds.

The other headline: Amazon committed up to $25 billion in Anthropic. The practical implication is infrastructure. Anthropic has been struggling with demand, and that struggle shows up as rate limits and outages. This deal gives them significantly more capacity. For operators, that should translate to fewer "Claude is busy" messages and more reliable scheduled tasks over the coming months.

Meanwhile, OpenAI launched GPT-5.5. The question operators are asking: should I switch? The short answer is no, but keep both in your toolkit. Claude is still stronger for the work most operators do: writing, analysis, document processing, financial modelling, and anything that requires careful reasoning. GPT-5.5 is better at autonomous browsing and multi-step tasks that involve navigating websites and systems. Different strengths for different jobs. Our switching guide has the full comparison.

Opus 4.7 has been live for two weeks. That is enough time to separate the benchmarks from the reality. Here is what we have found.

What improved. Coding performance jumped meaningfully. SWE-bench Pro improved by 10.9 points, which translates to Claude handling more complex, multi-file code changes correctly. Vision processing now supports 3.75 megapixels, roughly three times the resolution of Opus 4.6. If you work with detailed charts, dense spreadsheets, or high-resolution scans, the difference is noticeable. The new "xhigh" effort level gives Claude more room to reason through complex problems. And Task Budgets let you set token spending limits per request, which is useful for controlling costs on batch workflows.

What regressed. This is where it gets complicated. Long-context recall (measured by MRCR, a benchmark for finding specific information buried in long documents) dropped from 78.3% on Opus 4.6 to 32.2% on Opus 4.7. That is a significant regression. If you work with legal contracts, financial filings, or any workflow that requires Claude to find and accurately reference details across a long document, this matters. BrowseComp (a browsing accuracy benchmark) also dropped 4.4 points. Token costs are 12-18% higher in practice, which users attribute to changes in how the model processes text. And safety refusals are more aggressive, particularly on content that involves risk analysis, competitive intelligence, or anything Claude interprets as potentially harmful.

The community reaction. Some users called it "legendarily bad." That is overstated for most operator use cases. The coding gains are real and substantial. The vision improvements make document processing workflows genuinely better. But the long-context regression is also real, and if your workflows depend on it, you will notice.

Our practical recommendation. Use Opus 4.7 for coding, vision-heavy work, and structured analysis. Pin to Opus 4.6 for long-document workflows (contract review, financial analysis, research synthesis) until Anthropic addresses the MRCR regression. If you use Claude Design, you are on 4.7 by default, and it performs well there.

Full specs and migration guidance: Models

On 30 April, Anthropic is retiring support for large documents on two older Claude models: Sonnet 4.0 and Sonnet 4.5. If any of the tools, automations, or integrations your team uses run on those models, they may stop working when someone tries to process a long document.

Does this affect you? If you use Claude through Chat, Cowork, or Claude Code directly, no. Those interfaces already use newer models. This only matters if your team has custom integrations, third-party tools, or developer-built workflows that connect to the Claude API. If you are not sure, ask whoever set up your Claude integrations whether they use Sonnet 4.0 or 4.5.

What will happen. Any integration still on those older models will fail when it tries to process a document larger than about 150 pages. Shorter requests will keep working, which makes this easy to miss until someone uploads a large contract or quarterly report and the workflow breaks.

What to do. The fix is straightforward for your developer or IT team: switch the model to Sonnet 4.6 or Opus 4.7, which handle large documents natively. If you manage any integrations yourself (Zapier, Make, or similar tools with a Claude connection), check which model is selected in the connection settings and update it.

The bigger picture. Sonnet 4.0 and Opus 4.0 are retiring entirely on 15 June. If your team is updating integrations for the context deadline, it is worth moving to the latest models now rather than doing this twice.

What to tell your team. Forward this to whoever manages your Claude integrations: "Check if any of our Claude API connections use Sonnet 4.0 or 4.5. If they do, switch them to Sonnet 4.6 or Opus 4.7 before 30 April. The older models lose support for large documents on that date."

Our Models page has the full version history. The plan comparison shows which models are available on each plan.

Amazon $25B investment (20-21 April). $5B immediate at $380B valuation, up to $20B milestone-based. $100B+ in AWS spending over a decade. 5 gigawatts of compute capacity on Trainium2/3 chips. The largest single AI investment to date.

Anthropic postmortem (23 April). Three bugs confirmed as the cause of Claude Code's month-long performance decline. Fortune covered it extensively. Unusually transparent for a company at this scale.

Claude Code briefly removed from Pro (21 April). Anthropic pulled Claude Code access from Pro plans for roughly 24 hours before reversing the decision. A signal that Pro-tier Claude Code access may be repriced or restructured in the future. If you rely on it, keep an eye on announcements.

Mythos unauthorised access (21 April). An unauthorised group accessed Anthropic's restricted Mythos model via a contractor endpoint. The access has been revoked. No operator action required, but it is a reminder that even AI companies have security gaps. Our security and privacy guide tracks these developments.

NEC partnership (23 April). 30,000 NEC employees will get access to Claude. Another large enterprise deployment, following the pattern of company-wide rollouts rather than team-by-team pilots. Our team rollout guide covers how to structure these.

Anthropic Sydney office opens (27 April). Fourth APAC office, with Theo Hourmouzis appointed as ANZ General Manager. Good news for Australian and New Zealand operators who want local support and engagement.

Valuation crosses $1T in private markets (27 April). Private secondary trading now prices Anthropic above $1 trillion. IPO still targeting October 2026.

Claude Design token consumption warning. One user reported burning 58% of their Pro weekly usage limits in just two design sessions. Claude Design is powered by Opus 4.7 and consumes tokens aggressively. Budget your usage accordingly, especially on Pro plans. Our Claude Design guide has the details. For managing costs, see the cost management guide.

One thing to do this week: Forward the Operator Workflow section above to whoever manages your Claude integrations. If any of your tools run on older Sonnet or Opus models, they need to be updated before Wednesday. Also: if you are on Opus 4.7 and working with long documents, test your results against 4.6. The long-context regression is real, and for legal, financial, or research workflows, pinning to 4.6 may be the better choice until it is resolved.

P.S. This issue is a day late. We are, it turns out, still human. 

Claude for Operators is an independent publication. Not affiliated with Anthropic.

Just getting started? Visit our Operator's Toolkit for free guides on every workflow.

Got a workflow worth sharing? Reply to this email.

Need custom agentic workflows for your organisation? Get in touch or check out our services here.