Anthropic's internal merge rate jumps twelvefold as AI coding-trust gap widens

Anthropic's internal merge throughput jumps from 2,500 to 30,000 per day

Anthropic revealed on 2026-10-16 that internal code-change merges rose from roughly 2,500 per day in May 2026 to over 30,000 in September, a more than twelvefold increase inside one organization. The shift is tied to Scion, a version control system Anthropic built internally to absorb the throughput. For practitioners, the practical signal is that AI-assisted coding at this scale is forcing bespoke infrastructure: a generic host cannot absorb a ten-times-plus jump in daily merges without custom plumbing around branching, code review queues, and rollback paths.

Teams planning similar rollouts should expect to invest in their own pipeline layer rather than rely on off-the-shelf repository hosts.

Adoption is near-universal, but developers stay cautious when picking tools

A 2026-10-16 industry read found only 17.2% of developers say they do not use AI tools at all, with the rest exercising considerable caution in how they select and configure those tools. Reporting also described developers actively switching between AI assistants as their workflows evolve. The practical read is that "use AI" is no longer the dividing line — tool fit, configurability, and exit cost now matter more than adoption itself. Developers who treat an AI tool as a fixed, long-term commitment are likely to repeat the switching work their peers are doing.

A controlled trial found experienced developers 19% slower with AI while believing they were 20% faster

A separate 2026-10-16 piece reported a controlled study in which experienced developers using AI tools were measured 19% slower than a control group, while the same developers self-reported being 20% faster. The gap between perceived and measured productivity is the story, not the artifact: it implies existing benchmarks for AI coding assistants are weak signals of real throughput, and that team-level velocity metrics need to be instrumented before rollout rather than after. Practitioners evaluating new assistants should plan a measurement window of their own rather than trust vendor or self-reported numbers.

Meta debuts Muse Code while one developer documents switching to Pi

On 2026-10-16 reporting also noted Meta has debuted Muse Code, billed as a first-party AI coding tool, and that developer Miloš published a note on why he settled on the assistant Pi, citing its adaptability as his needs changed. For practitioners, two adjacent signals reinforce the same point: the major-vendor assistant market is still moving, and individual developers are publicly documenting their migrations rather than staying put. Tool-selection decisions made today should keep the switching cost visible from day one.

AMD opens FastFlow SDK for model experimentation

AMD introduced the FastFlow SDK on 2026-10-16, giving access to the tooling used to build and customize its flow-based models and letting developers experiment with model configurations and integrate new capabilities. The release widens the on-device and accelerator-side surface available to developers beyond the major AI-vendor assistants, which matters for teams that want to route specific workloads outside a hosted assistant. Practitioners building AI-assisted pipelines now have another knob to turn when a step is too latency-sensitive, too private, or too costly for a hosted model.

What to watch next

The next concrete checkpoint is whether Anthropic publishes further throughput figures tied to Scion, since the May-to-September arc is the only public anchor a practitioner has for an AI-native repository pipeline. The controlled productivity trial also implies a near-term need for in-house benchmarks before any further assistant swap, and AMD's FastFlow SDK landing on 2026-10-16 opens a short window in which teams can evaluate a non-hosted option against their current default. No pending release date in the evidence is firm enough to cite verbatim, so treat each of these as a "watch and measure" item rather than a deadline.