Claude Opus 4.7 launched on April 16 2026. It improves on Opus 4.6 in software engineering with particular strength on the hardest tasks. Teams now hand off complex long running coding work that once needed constant supervision and see reliable completion with built in checks.
The model pays strict attention to instructions. It devises verification methods for its own outputs and maintains consistency over long runs. This reduces the oversight burden that has limited agent adoption in many organizations. Testers across finance legal and engineering report accelerated velocity because the model catches faults in the planning phase instead of later.
Vision sees a substantial step forward. Opus 4.7 processes images at up to 2576 pixels on the long edge. That is over three times the resolution of previous Claude models. The difference shows in higher quality interfaces slides docs and accurate extraction from complex diagrams or dense screenshots. Life sciences teams use it for chemical structures and patent workflows with fewer errors.
This chart shows results from independent evaluations. CursorBench reaches 70 percent compared to 58 percent. Visual acuity for computer use hits 98.5 percent versus 54.5 percent. The 13 percent lift on the 93 task coding benchmark includes four tasks that neither prior Opus nor Sonnet could solve. Rakuten reported three times as many resolved production tasks with gains in code and test quality.
Feedback from two dozen companies paints a consistent picture. Replit observed the same quality at lower cost on log analysis and bug fixing. It pushes back on technical points to improve decisions. Notion saw it pass implicit need tests and continue through tool failures that halted earlier versions. Genspark highlighted loop resistance consistency and error recovery for their Super Agent. Databricks noted stronger document reasoning with 21 percent fewer errors on OfficeQA Pro.
Vercel described it as more correct and complete on one shot tasks with new behavior like doing proofs on systems code. Factory Droids measured 10 to 15 percent task success increase with fewer tool errors and better validation follow through. The pattern holds for Hex Quantium Ramp and Bolt. These reports indicate the model supports managing multiple agents in parallel rather than one at a time with heavy guidance.
API updates address real production needs. The new xhigh effort level sits between high and max. It offers granular control over reasoning intensity and latency for difficult problems. Task budgets in beta let developers guide token spend and set priorities across long executions. Default effort in Claude Code rises to xhigh.
Claude Code adds the ultrareview slash command. It launches a dedicated session to review changes and flag bugs or design issues that a meticulous reviewer would catch. Auto mode now reaches Max users for longer tasks with fewer interruptions. These build directly on the routines and desktop features I covered previously in my post on Claude Code routines for cloud automation.
Anthropic reduced cyber capabilities relative to Mythos Preview during training. They deployed safeguards that detect and block prohibited high risk cybersecurity requests. The Cyber Verification Program opens access for legitimate vulnerability research penetration testing and red teaming. This approach lets them gather data on safeguards before wider Mythos release. See my post on Claude Mythos Preview for more on their staged cyber work.
Pricing matches Opus 4.6 at five dollars per million input tokens and twenty five per million output tokens. Use the identifier claude-opus-4-7 on the API. Availability spans claude.ai the Claude Platform Amazon Bedrock Google Vertex AI and Microsoft Foundry.
Migration from 4.6 involves two main adjustments. The updated tokenizer maps text to 1.0 to 1.35 times more tokens depending on content. Higher effort levels generate more output tokens from additional thinking especially in multi turn agent scenarios. Internal coding evaluations showed favorable net token usage but measure on your traffic. The migration guide provides specific tuning advice.
Prompts need retuning because the model follows instructions literally. Previous versions sometimes skipped elements or applied loose interpretations. Opus 4.7 executes exactly as written. This produces unexpected outputs from old prompts but leads to fewer omissions when specifications are complex. Take time to update your harnesses and prompts for the new behavior. The shift rewards precise prompt writing and reduces the guesswork that crept into earlier interactions.
Safety profiles align closely with Opus 4.6. Modest gains appear in honesty and resistance to prompt injection attacks. The system card describes it as largely well aligned though not perfect. Alignment work continues with Mythos Preview showing the strongest results so far. The deliberate capability reductions on cyber tasks reflect a careful path toward broader releases.
Long context memory sees improvement too. The model makes better use of file system based notes across multi session projects. This reduces the need to reload context and supports sustained work on finance models legal analysis or multi day engineering efforts. One research agent benchmark gave it the top efficiency score with strong performance on general finance at 0.813 versus 0.767 for the prior version. Consistent long context scores across modules stand out as a key advantage for real enterprise document and data work.
The release arrives about seventy days after Opus 4.6. It matches the pattern of targeted jumps that have defined the past months. Leaks had pointed to Claude Code enhancements around parallelism and review tools. This delivery focuses those capabilities into production ready form and aligns with the pre launch speculation I tracked in earlier posts.
For my own use I plan to test Opus 4.7 on the longest agent runs and highest resolution visual tasks first. The combination of self verification reduced tool errors and literal instruction adherence addresses the main friction points I have seen in current agent setups. If the reduced supervision holds across my workloads it will shift how I allocate time between human review and parallel agent management. Start with high or xhigh effort on challenging problems. Track quality cost and output token counts closely. The data from early users suggests that once prompts align with its style the model delivers cleaner code fewer iterations and more shippable outputs on complex projects.
Overall this release advances the state of practical agentic coding and multimodal professional work. It does not require overhauling your stack but it does reward measurement and prompt iteration. Teams already invested in the Claude ecosystem will find the upgrade straightforward. Others should benchmark against their specific coding analysis and visual workflows before committing. The fast cadence means another improvement will follow soon. Stay model agnostic and keep testing on the tasks that matter to you.