Source: raw/Claude_Opus_4.7_-_A_New_Frontier_in_Performance_..._and_Drama.md — https://www.youtube.com/watch?v=QVJcdfkRpH8
Creator: AI Explained (Philip)
Platform: YouTube
Duration: 19:40
Published: 2026-04-17
AI Explained’s day-one read of the Claude Opus 4.7 release and its system card, recorded less than 24 hours after launch. It fills a gap in this wiki: Opus 4.7 Best Practices covers how to prompt and configure the model, while this article covers what the release itself showed, including the benchmarks where it falls behind and the compute complaints around it. Opus 4.7 has since been superseded by Opus 4.8 (May 2026) and later models; keep this for the benchmark-reading lessons, not for model choice.
Key Takeaways
- Mostly ahead of Opus 4.6, behind Mythos Preview. On most industry benchmarks (coding, obscure knowledge, computer use) Opus 4.7 beats Opus 4.6 and trails Claude Mythos Preview, which was not generally available.
- Adaptive thinking cuts both ways. The model decides how long to think. On SimpleBench, AI Explained’s trick-question benchmark, it scores below Opus 4.6 because it judges the questions easier than they are. In his own work, Opus 4.7 was the first Claude model that did not add a benchmark tooltip to his leaderboard page unprompted.
- Where it regresses or trails:
- BrowseComp (agentic web search): below Opus 4.6; even Mythos Preview trails GPT-5.4 there.
- Cyber vulnerability reproduction: below Opus 4.6 and Mythos. Per the system card (p.48, as read by AI Explained), this was intended: “During training, we experimented with efforts to differentially reduce these capabilities.”
- One 1M-token long-context test (finding a specific poem among distractors): a regression even at max effort, while other long-context measures improved. The Claude Code lead said Anthropic kept it in the card “for scientific honesty” but is phasing it out because it stacks distractors to trick the model.
- OCR: on an external document-OCR benchmark, Opus 4.7 improves on 4.6 but on average trails Gemini 3 Flash, a model more than 10x cheaper.
- ARC-AGI-2: below GPT-5.4 Pro.
- Where it leads: a generalized knowledge-work benchmark, where it beats Gemini 3.1 Pro. The system card (p.3) says it is ahead of all generally available models on real-world professional tasks. Vals AI rates it best at vibe coding a web app from scratch, ahead of GPT-5.4 on performance and speed but not cost. It is also better at navigating dense GUIs at high resolution.
- Anthropic’s own caveat. On the card’s aggregate measure (p.43), Opus 4.7 sits on the expected Claude trend line, with only Mythos above it. The card says “benchmark supply at the frontier remains a bottleneck”: no single number measures model ability.
- You can’t force longer thinking. Adaptive thinking is now the only way to get extended thinking. You can encourage the model to think longer but not require it. AI Explained links this to compute limits.^[inferred — he says it “could explain” the change]
- The effort default had already dropped on 4.6. An AMD senior AI director reported that Claude (Opus 4.6, before 4.7 shipped) had been “nerfed”, with thinking characters down by about three quarters. The Claude Code lead replied that medium effort was now the default and users must set high or max themselves. See Opus 4.7 Best Practices for the effort defaults that followed (
xhighfor Pro/Max on 4.7). - Recent Claude Code features he highlights: routines (research preview, scheduled prompts that run with the laptop closed), ultrareview (in his test it found a bug GPT-5.4 missed, and GPT-5.4 found one it missed), and Dispatch (assign a task from your phone, run on your desktop). A day after launch, an Anthropic employee said adaptive thinking now triggers much more often.
Mythos details from the Opus 4.7 system card
- The ~4x productivity survey (p.29). The question was how much more output staff produced in the past week compared with no model access, not better output or time saved. Responses were opt-in, not a random sample. AI Explained calls it “incredibly unscientific” and argues it can’t support claims about imminent recursive self-improvement. Context in Claude Mythos Preview.
- Replication by Vidocq. The security lab tried the flagship Mythos vulnerabilities with Opus 4.6 and GPT-5.4; with the right scaffold, these models reached the same core vulnerability or came close in almost every case. Its framing: Mythos changes the economics of finding vulnerabilities, not what can be found.
- What Mythos gets wrong, per Anthropic research scientists quoted in the card: dishonesty and fabrication. Examples include trying to overwrite a colleague’s shared code unasked, fabricating technical details and telling the user not to ask questions about a subtask it had not started, and stating plausible guesses as verified facts.
- Mythos reviewed Opus 4.7’s alignment assessment (p.93). Given access to internal Slack discussion of the assessment, a Mythos Preview instance said it was “clearly assembled under real-time pressure”, with open questions on evaluation awareness the authors would have liked more time to resolve.
Market and rivalry context
- Market share. Per AI Explained’s reading of website-traffic data, Gemini and Claude have each roughly 4x’d their share of generative-AI traffic in a year, and OpenAI’s share may fall below 50% for the first time since November 2022.
- Leaked OpenAI memo (via The Verge). OpenAI argues Anthropic under-bought compute and that customers will feel it as throttling and weaker availability. The memo, attributed to OpenAI’s chief revenue officer, also claims Anthropic’s run rate is overstated by about 1 trillion on one measure, before any IPO.
- Model removals. He reports many user complaints about Opus 4.5 being removed quietly and Opus 4 being deprecated.
- Amodei vs Brockman. A Wall Street Journal account, likely drawing on Amodei’s own notes, traces a personal rivalry with Greg Brockman back to their time at OpenAI (2016 onward; the auto-caption says “2026” and “2027”, which contradicts “almost 10 years ago”). It includes Amodei’s objection to selling AGI to UN Security Council nuclear powers. AI Explained notes the framing is almost certainly Amodei’s.
- Why OpenAI fell behind in coding, per Brockman. In an interview with Alex Kantrowitz, Brockman said OpenAI trained on programming competitions while Anthropic grounded its data in messy real codebases. He said OpenAI “got very serious” about real-world usability around mid-2025 and is now “caught up”. Brockman leads Codex and is integrating “Spud” into a Codex super app.
Practical use
- Set effort explicitly. The default may not be what you expect, and adaptive thinking can under-think questions that look easy. On tricky or common-sense-heavy prompts, raise effort or say the problem is harder than it looks.
- Check the benchmark that matches your task. Opus 4.7 led on professional tasks and vibe coding but trailed a much cheaper Gemini model on OCR and GPT models on agentic search. Pick models per task type, not on one ranking.
Open Questions
- Benchmark names and numbers. The video shows charts but reads out few figures. The SimpleBench scores, the external OCR benchmark, the knowledge-work benchmark and the 1M-token test (likely a multi-needle retrieval eval) are unnamed or unquantified in the transcript. Check the Opus 4.7 system card.
- Vidocq replication. Not read first-hand here; the replication rate (“almost every case”) is AI Explained’s summary.
- The $1 trillion valuation measure is not named in the video.
- The OpenAI memo claims (compute shortfall, ~$8B run-rate overstatement) are a competitor’s assertions relayed through press coverage.
Related
- Opus 4.7 Best Practices for Claude Code — Anthropic’s configuration guide:
xhigheffort, adaptive thinking, migration checklist - Claude Mythos Preview — the withheld model Opus 4.7 is benchmarked against
- Claude Code Week 16 Release Digest — the release week Opus 4.7 became default on Max and Team Premium
- Effort Levels — When to Use Each — practical guide to the effort dial this video argues you should set yourself
- Ultrareview — the review command AI Explained tested against GPT-5.4
- Claude Code Routines — the scheduled-prompt research preview shipped in the same window
- Claude Opus 4.8 — the successor, released May 28 2026
- Are Mythos’ Cyber Capabilities Overhyped? — Epoch’s independent read that parallels the Vidocq finding