Developers flag Claude 4.6, Opus 4.7 limits

Developers spent May 19 comparing the practical limits of Anthropic’s Claude 4.6 and Claude Opus 4.7, with discussion centering less on headline benchmark gains than on the constraints that shape real production use. Posts on X and developer forums focused on computer-use image handling, how much reasoning developers can explicitly buy, and when prompt caching lowers costs at the expense of flexibility, according to the social briefing and Anthropic’s model documentation. (platform.claude.com) Anthropic released Claude Opus 4.7 on April 16 and described it as “a notable improvement on Opus 4.6 in advanced software engineering,” while keeping pricing unchanged at $5 per million input tokens and $25 per million output tokens. (platform.claude.com) The company said the model is available across Claude products, its API, Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry. ### What were developers actually flagging on May 19? May 19 discussion singled out three operational limits: image resolution in computer-use workflows, tighter controls around “thinking,” and the economics of caching, according to the source briefings. (platform.claude.com) Those are the kinds of details that matter when teams are running browser agents, screenshot-heavy tasks or long coding sessions rather than one-off prompts. Anthropic’s documentation supports the first two points directly. The company says Opus 4.7 is its first model with high-resolution image support and lifts the maximum image size to 2,576 pixels on the long edge, from 1,568 on prior models. Anthropic also says higher-resolution images consume more tokens, creating a direct cost trade-off for teams sending screenshots or documents into the model. (anthropic.com) ### Why did screen resolution come up in a coding conversation? Anthropic says the vision change is “particularly important for computer use and screenshot/artifact/document understanding workflows.” The company also says Opus 4.7 now maps coordinates 1:1 with actual pixels, removing scale-factor adjustments that earlier workflows needed. That matters because computer-use agents depend on reading interfaces accurately. Anthropic says Opus 4.7 improves low-level perception, including pointing, measuring and counting, and improves image localization. Developers discussing screen caps were effectively talking about how much visual detail the model can reliably use before cost rises. (platform.claude.com) ### What changed in “thinking” depth between 4.6 and 4.7? Anthropic’s migration guide says “extended thinking” with an explicit token budget is no longer supported on Claude Opus 4.7 or later and will return a 400 error. (platform.claude.com) Developers migrating from Opus 4.6 must switch to “adaptive thinking” and control reasoning depth through an effort setting instead. The same guide says adaptive thinking is off by default on Opus 4.7 unless developers enable it. Anthropic’s “What’s new” page adds a new “xhigh” effort level and says effort lets developers trade capability against speed and token spend. (platform.claude.com) In practice, that means developers have less direct budget-token control than they had with older extended-thinking settings, even though Anthropic says Opus 4.7 is stronger on hard software-engineering tasks. ### Where does the benchmark claim fit in? Anthropic said at launch that Opus 4.7 shows better results than Opus 4.6 across a range of benchmarks. (platform.claude.com) A third-party benchmark tracker updated May 19 lists Claude Opus 4.7 (Adaptive) at 87.6% on SWE-bench Verified, behind only Anthropic’s Claude Mythos Preview. The 80.8% figure in developer chatter appears to refer to a different Claude variant or benchmark snapshot rather than Anthropic’s published Opus 4.7 launch materials. Anthropic’s official release and documentation reviewed here do not show 80.8% for Opus 4.7. (platform.claude.com) ### What should developers watch next? Anthropic’s migration guide says teams moving from `claude-opus-4-6` to `claude-opus-4-7` need to update model names, replace extended-thinking parameters, and recalibrate effort settings. The company also introduced beta “task budgets,” which give the model a token target for a full agentic loop, including thinking, tool calls and final output. (anthropic.com) Anthropic’s next practical checkpoint for developers is in those migration and “What’s new” documents, where the company lays out supported parameters, effort levels and budget controls for Opus 4.7 deployments. (anthropic.com) (platform.claude.com)