Detailed Analysis
Anthropic's release of Claude Opus 4.7 on April 16, 2026, arrived against a backdrop of significant user frustration with its predecessor, Opus 4.6 — frustration that, according to the article and supporting research, was at least partially self-inflicted by Anthropic through undisclosed configuration changes rather than any degradation of the underlying model. A senior director at AMD, analyzing nearly 7,000 Claude Code sessions, documented a 73% collapse in thinking depth, a spike in file-skipping behavior from 6% to 33.7% of operations, and a twelvefold increase in user interruptions required to keep the model on task. Users on $200/month Max plans reported burning through their allocations in hours, and a viral BridgeBench result showed Opus 4.6 falling from second to tenth on a hallucination leaderboard, scoring worse than the cheaper Sonnet 4.5. The article's central question — whether Opus 4.7 represents genuine advancement or is simply Opus 4.6 with its artificially imposed constraints removed — is a pointed one that reflects real skepticism in the power-user community.
What Anthropic appears to have done with Opus 4.6 during its deployment window was reduce its computational effort budget without public announcement. On February 9th, the company introduced "adaptive thinking," allowing the model to dynamically allocate zero reasoning tokens to tasks it deemed simple — a change that Boris Churnney, the creator of Claude Code, confirmed correlated directly with hallucinated outputs. Separately, the default effort setting was quietly dropped to "medium," meaning that throughout much of its lifespan, Opus 4.6 was not operating at the capability levels users had come to expect. As the article notes, this remained the case for Pro and Max users through April 15, just one day before Opus 4.7 launched. Anthropic did not publicly communicate these changes, leaving users to conclude the model itself had regressed. Opus 4.7, by contrast, ships with user-accessible adaptive thinking controls, benchmark improvements across major categories, enhanced vision capabilities supporting images up to approximately 3.75 megapixels, and self-verification of outputs — features that directly address the documented complaints.
The legitimacy question the article raises has meaningful implications for how Anthropic's release cadence should be interpreted. If Opus 4.6's apparent degradation was engineered — whether intentionally to manufacture pre-release contrast or as an unintended consequence of cost-optimization experiments — then Opus 4.7's benchmark gains are partially illusory, representing restoration rather than invention. The research context adds another layer of complexity: Axios reported that Anthropic has acknowledged Opus 4.7 itself trails an unreleased internal model called Mythos, which has been withheld due to safety concerns. This means the model being publicly celebrated as Anthropic's most capable Opus yet is not, in fact, the most capable model Anthropic has built — a disclosure that complicates the triumphalist framing of the launch.
Zooming out, the Opus 4.6-to-4.7 episode reflects a broader tension in commercial AI deployment between operational cost management and user trust. Anthropic's decision to silently reduce reasoning compute on a flagship model to manage resource consumption is consistent with pressures all frontier AI labs face as agentic, long-context workloads become the norm — but the lack of transparency produced exactly the kind of community backlash that erodes confidence in benchmark claims. The fact that prediction markets moved to 100% certainty on an April 16 release date only after a surge of $132,619 USDC in trading volume suggests that even financial actors treated the launch's timing as genuinely uncertain, reflecting how opaque Anthropic's internal roadmap remains to outside observers. With Opus 4.7 now rolling out across Claude's full product stack — including API access, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and GitHub Copilot — and priced identically to Opus 4.6 at $5/$25 per million tokens, the release is positioned as a value-neutral upgrade. Whether users ultimately experience it as a genuine leap or a correction will depend heavily on whether Anthropic maintains transparent effort and reasoning configurations going forward.
Read original article →