Claude Opus 5.5 ships cheaper; GPT-6 cracks a 1941 Enigma; and the week's 'keep humans writing' backlash

Anthropic ships Claude Opus 5.5 — near-frontier capability at a discount, with agent work as the whole pitch. Claude Opus 5.5 is Anthropic's new flagship model, released September 22 as the first of a new 5.5 generation, and the announcement is aimed squarely at agent work: software jobs a model is expected to start and actually finish. Anthropic says it matches Claude Fable 5.1, another top-of-range model in the family, on most kinds of work while costing 40% less to run than the previous Opus 5.

If you're budgeting agent fleets for clients, that combination — frontier-adjacent capability at a materially lower run cost — is the story, and it improves further within weeks, when Claude Sonnet 5.5 and Haiku 5.5 arrive carrying many of the same upgrades at cheaper tiers. Anthropic is also leaning on the safety angle for enterprise buyers: it says Opus 5.5 posted the strongest result of any model so far on its behavioral audit, an in-house suite that plays out thousands of simulated scenarios to see how a model behaves. Pricing is $4 per million tokens in and $20 per million tokens out, 20% below Opus 5, and GitHub's product chief reports that in his testing it finished real tasks with unusually few tokens and steps. Anthropic

“Claude Opus 5.5 used among the fewest tokens and steps we measured.”

'I don't want to read what you didn't write': readers are revolting against AI prose. Colin Breck's essay names a quiet problem inside companies: AI-generated design documents, tickets and meeting summaries that are fluent but carry no real context and no human voice — and readers increasingly resent them. The numbers he cites are blunt about the penalty.

If you hand documents to clients, take note: in a survey he references, 78% of readers said they would stop reading an article they believed was AI-written and 71% would avoid that author, while 98% preferred his own prose to an AI rewrite of it. Bryan Cantrill has gone furthest — all public writing at his company, Oxide, must pass Pangram, an AI-detection tool, as human-authored. Breck is not anti-AI: while producing an academic paper he leaned on it heavily, and it caught a subtle notation slip that four expert human reviewers had all missed — yet it wrote none of the text. That's a workable model for delivery work: machines checking, humans authoring. blog.colinbreck.com

“To use an LLM to write is to void the social contract between writer and reader”

'AI has no wisdom': a practitioner's warning that vibe-coded projects rot. Alexandru Nedelcu's bluntly titled essay makes a maintainability argument: projects assembled mostly by prompting decay into an unmaintainable mess, because models were never taught what makes code keepable — maintainability offers no reward you can measure immediately — and they trained largely on the mediocre code that dominates the wild.

The caution for anyone steering a client's adoption: letting agents loose on scaffolding and well-specified tasks is fine; letting them own anything with a long life is not. He also sees teams now outsourcing both the writing and the reading of code, and notes that even the best current models are poor at simplifying, splitting functions in ways nobody would reuse. He's no holdout — he uses LLMs daily for tedious work and coaches colleagues — but he closes with a prediction worth keeping: companies will eventually advertise NO-AI policies as a competitive advantage, the way food brands advertise the absence of preservatives. alexn.org

“they no longer take responsibility for mistakes in coding and no longer learn from those mistakes”

Markdown in /src: check your specs in next to the code, and let agents compile them. Carson Gross, htmx's creator, proposes a convention for agentic codebases: keep the Markdown documents describing the software — README, TODO, overviews, plus notes on features, data, the API and infrastructure — in a /src/md directory beside the code, and generate both code and tests from those documents rather than from one-off prompts in a chat window.

His claim is that Markdown has quietly become the real source of an agentic project, with the generated code turning into a build artifact — a view that pairs neatly with the essays above: humans author the layer that decides what the software is, agents do the rest. It's also a cheap experiment to run on a client project, since the same specs can be regenerated when a better model ships. In the proposed layout, a README indexes everything and serves as the agent's entry point — and Gross is explicit that agents should not generate much of it; the point is human-authored, human-curated specs. htmx.org

“the actual code that is generated by the agent is sort of becoming a low-level implementation detail”