Anthropic pushed Claude Code 2.1.111 on April 20, 2026, and the headline feature is immediate: Claude Opus 4.7 is now available in Claude Code with a new `/effort` command that unlocks an xhigh effort level, and Auto mode is live for Max subscribers. This is the first time engineering teams can dial up reasoning intensity directly from the CLI without touching a single API parameter. For teams running agentic coding workflows, that's a meaningful control surface. But there's a catch in this release that deserves your full attention before you migrate.

Three things matter in this changelog:

The `/effort` command is the workflow change you'll feel immediately. Previously, effort calibration happened implicitly through prompt engineering or model selection. Now you can invoke it explicitly:

That's a cleaner interface than the workarounds teams were using before. Auto mode takes it further: Opus 4.7 reads task complexity and self-selects the appropriate effort tier, which reduces the cognitive overhead of deciding when to push the model hard.

Released April 16, 2026, Opus 4.7 is Anthropic's production-tier model sitting between Opus 4.6 and the restricted Claude Mythos Preview. The number that engineering leaders should care about: a 10.9 point improvement on SWE-bench Pro over Opus 4.6, which measures real-world GitHub bug-fixing performance. That's not a marginal gain. For comparison, OpenAI's o3 sits at 8.7% on standard SWE-bench. Anthropic is claiming a genuine lead on the task that matters most for agentic coding pipelines: finding and fixing bugs in production code.

Other capability upgrades in 4.7:

The vision upgrade is underrated. Teams running automated UI testing or feeding architecture diagrams into coding agents have been hitting resolution limits. The jump to 3.75MP opens up workflows that were previously impractical.

Here's where this release gets complicated. Opus 4.7 ships with a new tokenizer that inflates token counts by up to 35% for code, JSON, CSV, and non-English text compared to Opus 4.6. Pricing is unchanged at $5 per million input tokens and $25 per million output tokens, but when your codebase suddenly tokenizes 35% heavier, that's an effective price increase that Anthropic didn't advertise. The community noticed immediately. A Reddit post titled "Opus 4.7 is not an upgrade but a serious regression" collected 2,300 upvotes within 48 hours. An X post on the same theme hit 14,000 likes. That's not noise. That's a signal that developers running code-heavy workloads felt the cost change before they saw any capability benefit. The backlash deserves context, though. ArtificialAnalysis data shows approximately 10% net cost savings from fewer output tokens in evals, meaning Opus 4.7 produces more concise responses. For teams running prose-heavy workflows, NLP pipelines, or documentation generation, the economics may actually improve. The pain is concentrated in code, JSON, and structured data inputs, which is exactly what most engineering teams are pushing through Claude Code. This is the honest breakdown:

If your Claude Code usage skews toward large codebase context windows and multi-file operations, model this cost change before committing to Opus 4.7 as your default.

The SWE-bench Pro lead is real, and it positions Claude as the strongest production option for agentic bug-fixing specifically. But the full landscape looks like this in April 2026:

The xhigh effort level is Anthropic's attempt to give teams a practical way to access Opus 4.7's ceiling capabilities without waiting for Mythos Preview to become generally available. It's a smart bridge strategy.

The 2.1.111 release has genuine value, but it requires deliberate evaluation rather than a blanket migration. Here's the priority order: This week:

Next 30 days:

Hold pattern: If your primary workload is code review and refactoring on large codebases with significant JSON or structured data context, the tokenizer inflation may offset the capability gains in the short term. You're not missing a step-change by staying on Opus 4.6 for another few weeks while the cost picture clarifies.

The `/effort xhigh` command is the architectural tell in this release. Anthropic is building explicit effort control directly into the CLI because the teams that will define the next generation of software engineering are running agents, not just autocomplete. They're assigning entire workflow phases to AI systems that need to reason deeply before acting. The Navy SEAL framing applies here: a team of 5 engineers using Claude Code with xhigh effort on complex refactors can outproduce a team of 50 running brute-force manual code review. But those 5 engineers need the tooling to be right. The SWE-bench Pro lead Anthropic is claiming is evidence they understand what that team actually needs. The tokenizer inflation is evidence they still have operational kinks to work out. The companies that win the next 24 months of software development won't be the ones that waited for perfect tooling. They'll be the ones that built internal evaluation discipline, hired engineers who know how to leverage agents effectively, and adapted their workflows fast enough to compound the capability gains before competitors did. Opus 4.7 with xhigh effort is a meaningful step in that direction. Measure the cost impact, run the benchmarks, and move.