Anthropic rolled out Claude Opus 4.7 on April 16, 2026, positioning it as the company’s strongest publicly available model for software engineering and long-running agent tasks. Developers can now hand off intricate coding jobs with less oversight. The model shines on SWE-bench Pro at 64.3%, edging out OpenAI’s GPT-5.4 at 57.7% and Google’s Gemini 3.1 Pro at 54.2%, according to The Next Web. That’s a jump from Opus 4.6’s 53.4%. On SWE-bench Verified, it hits 87.6%, surpassing its predecessor at 80.8% and Gemini’s 80.6%.
But. Opus 4.7 isn’t the frontier. Anthropic’s Mythos Preview, restricted to select enterprise partners for cybersecurity testing, crushes it across the board—93.9% on SWE-bench Verified, 77.8% on SWE-bench Pro. The company cites safety risks, including Mythos’s prowess at spotting zero-days, as reason for the lockdown, as detailed in Anthropic’s announcement. Opus 4.7 gets cyber safeguards to block high-risk requests. It’s the practical choice today.
Coding pros get tangible gains. CursorBench rises to 70% from 58%. Terminal coding climbs to 69.4% from 65.4%. Agentic feats improve too: 14% better multi-step workflows, one-third fewer tool errors. It infers tools without prompts—first Claude to pass implicit-need tests. Multi-agent coordination handles parallel streams, like code review alongside docs analysis. Resilience kicks in on failures; it adapts without crashing.
Vision upgrades handle images at 2,576 pixels on the long edge—three times prior resolution. Perfect for UI mocks or dense diagrams. Context stays at 1 million tokens. GPQA Diamond nears ceiling at 94.2%, matching GPT-5.4 Pro’s 94.4%. Long-context research ties top at 0.715.
Pricing holds steady: $5 per million input tokens, $25 output. Prompt caching saves up to 90%; batch API cuts 50%. Access hits Claude plans, API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry. AWS confirmed availability in key regions same day, per AWS News Blog.
This builds on Claude 4 family roots. Opus 4 launched May 2025 as top coder at 72.5% SWE-bench Verified, Sonnet 4 at 72.7%, per Anthropic’s Claude 4 post. Sonnet 3.7 Sonnet debuted February 2025 with hybrid reasoning—quick replies or extended thinking—and Claude Code CLI, as in its announcement. Opus 4.7 refines that path, sustaining hours-long focus.
Industry buzz lit up X. “Claude Opus 4.7 is OUT Bigger reasoning leaps, agentic coding on steroids,” posted @HussInvestments. Developers note less babysitting: stricter instruction adherence, self-verification. One engineer built a full Rust text-to-speech engine autonomously, checking against Python refs.
New controls emerge. ‘xhigh’ effort mode balances think time and speed—default in Claude Code. Task budgets cap tokens for predictability. /ultrareview flags bugs like a human. Token use may rise 1.0-1.35x from deeper reasoning.
Comparisons sharpen the picture. Opus 4.7 lags GPT-5.4 on web search (79.3% vs 89.3% BrowseComp), cybersecurity flat at 73.1%. Mythos leads everywhere: 82.0% terminal vs 69.4%, per X threads and VentureBeat. CNBC noted it’s “less broadly capable” than Mythos, following Opus 4.6 in February, via its report.
Enterprise adoption accelerates. Databricks integrated it for document reasoning, 21% fewer errors on OfficeQA Pro than 4.6. GitHub eyed Sonnet 4 for Copilot agents earlier; Opus 4.7 extends that edge.
Anthropic paces releases tightly—Opus 4.6 in February, now 4.7. Safety cards for prior models stressed agentic risks like reward hacking, mitigated via training and monitoring, from Claude 4 system card. Opus 4.7 closes gaps without over-refusals.
Developers switch workflows. Less retries. Full-resolution vision. Predictable costs. It verifies outputs. Handles failures. Builds end-to-end.
The gap to Mythos? Safety. Performance ceiling hit human levels in spots. Agentic coding shifts from hype to hourly grind. Opus 4.7 delivers now. Mythos waits.
WebProNews is an iEntry Publication