A new wave of complaints is building around Claude Opus 4.6. Technical users say the model no longer holds up as well on long, difficult engineering tasks. This is not just social media noise. TechRadar cited a case from AMD’s AI team, backed by large-scale usage data, arguing that real-world performance has degraded over time. Anthropic’s own documentation also shows that Opus 4.6 uses adaptive thinking, which means the model now decides how much to think instead of always following a fixed reasoning budget.

SiFive has reached a $3.65 billion valuation in a new round backed by Nvidia, as investors continue to bet on RISC-V and open chip design for the AI era. This is not a standard GPU story. The bigger signal is that more value is starting to accrue below the GPU layer, where companies are building the tools and architectures needed for custom silicon, inference optimization, and tighter control over the compute stack. If Nvidia came to define one winning layer of AI infrastructure, SiFive points to the layer underneath it that is now being repriced.

Perplexity can now use bank, card, loan, and investment data through Plaid to answer more concrete financial questions. The bigger signal is that AI becomes far more useful when it connects to permissioned data rails, not just web search.

This is a smaller story, but a clean one. On-device AI is starting to move beyond hobbyist use and toward a personal AI layer that can span Mac, iPhone, and iPad. Cloud AI may win on capability. Local AI may try to win on privacy, ownership, and persistence.

Recent incidents around Sam Altman’s home suggest AI is no longer just an online argument. It is becoming a social and political flashpoint attached to visible individuals. This is not today’s lead story, but it is a signal worth watching.

If a frontier lab can insert itself directly into Word workflows, value does not stay fully inside Office anymore. It shifts toward the AI layer that shapes work inside the interface people already open every day.

Problem: most teams only notice model quality changes after a real workflow starts failing. By then, the damage is already operational. A bug fix takes longer than usual. A refactor introduces new mistakes. A long memo suddenly needs far more editing. Because the decline is gradual, it is hard to prove and easy to dismiss as “just a bad run.”

Here is a simple way to catch that drift early.

First, pick 5 real tasks your team repeats every week. Do not invent benchmark-style prompts. Use work that actually matters. Good examples: fix a difficult bug from a real codebase, refactor a large file without changing behavior, extract structured data from a PDF, summarize a 20-page internal doc, or turn messy meeting notes into a decision log.

Second, save the exact input for each task. Save the prompt, the file or text you used, and the expected output format. If a task needs context, keep that context identical every time. The goal is not to test creativity. The goal is to test consistency.

Third, create a simple scoring table with 4 columns:

Score each output from 1 to 5. If useful, add two operational columns:

Fourth, run the same 5 tasks once a week on the same model. Do not tweak the prompts mid-test. If the model fails, let it fail. That is the signal.

You can use this scoring prompt to make evaluation faster:

At the end of each week, compare the average score to the prior week. If one category drops for two weeks in a row, investigate before your production workflow absorbs the cost.

⏱ Time to implement: ~25 minutes the first time, ~10 minutes per weekly run
🛠 Tools needed: your current model, Google Sheets or Notion
💰 Cost: existing inference cost only

Techzip note
This is less about ranking models and more about protecting your workflow. If a model changes, you want evidence before the team starts arguing from memory.

Our documentation assistant had a problem. RAG pipelines only sent the model page fragments, making responses feel half-baked. The fix was obvious: give the agent full doc access. Quality jumped, but startup hit 46 seconds and costs ballooned to $70K+/year.

The insight: the agent doesn't need a real filesystem. It just needs to think it's in one.

We built ChromaFs, a virtual filesystem that translates standard UNIX commands into queries against our Chroma database. Every doc page becomes a file, every section a directory. The agent explores documentation the way a developer explores a codebase.

The results: startup dropped from 46s to ~100ms, marginal compute cost hit $0 per conversation, and output quality matched the full sandbox, with access control built in.

ChromaFs now powers 30,000+ daily conversations for hundreds of thousands of users. No containers, no cold starts, no invoice surprises.

Mintlify powers docs for 20,000+ companies, reaching 100M+ people a year. We just raised a $45M Series B led by a16z and Salesforce Ventures to keep building the knowledge layer for the agent era.

Try Mintlify today!

This is not an essay or a long-form analysis piece, but it is still worth reading if you want to understand the technical backdrop behind the Opus 4.6 debate. The document explains how Claude 4.6 decides how much to think based on task difficulty instead of always using a fixed reasoning budget.

Why it made the cut: if model behavior changes because of default reasoning mechanics, then many “the model got nerfed” debates may really be about transparency and controllability, not just raw capability.