A retirement date for three models, and a behaviour score for verified bots
GitHub will retire three Copilot models on 16 November with a month of notice and named replacements, the second monthly round run that way. Cloudflare now scores its verified bots on measured behaviour and will publish the trust bands.
How we Build
GitHub put a month of notice and a named successor on three retiring models
On 15 October GitHub announced the next Copilot model retirements. Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra all go on 16 November 2026, across every Copilot surface. Claude Opus 5.5 and GPT-6.1 Sol or GPT-6 Luna are named as the replacements. Under default settings they switch on automatically. Opus 4.8 stays for annual individual subscribers. It is the second consecutive monthly retirement round run on these terms.
We ask that whatever a team publishes for others carries its version and a written account of the ending, so that a consumer can plan.
Every interface the team publishes carries a version and a deprecation promise that a consumer can plan against: a notice period or a withdrawal condition, or an explicit declaration that the interface is supported without end.
The largest developer platform now keeps that promise as routine: a date, a named successor and the carve-outs in writing, twice in two months. A notice is not yet performance. The first round’s retirement lands on 19 October, and whether consumers migrate in time is measurable but not yet measured.
How we Assure
Cloudflare will publish a behaviour score for each bot it has verified
Cloudflare announced the change on 15 October. Its verified bots are now scored on what they do over time. The system watches whether a bot’s traffic respects the preferences a site publishes in robots.txt, and whether the bot slows down after an HTTP 429 response. Both signals feed an explainable trust band rather than a permanent verdict. The scores will be public on Radar, and customers will choose the trust threshold at which traffic is handled. Perplexity’s bot was de-listed in 2025 for evading no-crawl directives.
Our position is that a fleet of agents is held to a floor no headcount can buy past, owned by a named person, with health shown as a spread of results rather than one summed figure.
A quality floor exists that no volume of additional agents can buy past, it has an identified holder, and fleet health is reported as a distribution rather than as a total.
Identity was never the floor. The floor is behaviour, continuously measured, with somebody able to act on the score. The signals here are the ones a network operator can see, path respect and rate adaptation, and the system is in internal use rather than in customers’ hands.