AI model
GPT-5 is OpenAI's large language model family, succeeding GPT-4, powering general-purpose reasoning, coding, and agentic tasks in OpenAI's products and APIs.
Aleph Alpha published Kolibri, a 78B-parameter English and German mixture-of-experts model, under Apache 2.0 with about 3.46B parameters active per token. Teams that cannot send prompts to outside servers can host it themselves if they pay for the GPUs.
Reality
Coinbase replayed 16,140 real transactions and found each newer version of three AI model families caught less fraud than the one it would replace. The result undercuts the idea that a model upgrade is a free improvement for a working fraud screener.
Reality
A study of eight frontier models on SWE-bench Verified puts agentic coding at roughly 1,000 times the token cost of code chat, dominated by input, with the models' own pre-run estimates correlating no better than 0.39.
Reality
A dev.to post makes a statistical case for reading your own code, since agents reproduce whatever pattern already dominates the repo, and its author goes hands-on in greenfield work while easing off once conventions are established.
Reality
Sonnet 4.5 still leads GPT-5 on the coding leaderboards, and GPT-5 lists about 46 percent below it on a 5:1 token mix. Anthropic's current Sonnet undercuts both of Sonnet 4.5's list prices, and that complicates a routing plan built on the older pair.
Reality
Frontier labs are citing AI-designed bioweapons to argue for new controls on synthetic DNA. Researchers who run labs told WIRED the hard part is assembling and verifying a virus, and that no fully autonomous lab exists.
Reality
Harvard and Alberta researchers coded 54,861 subreddit posts around two real AI updates and surveyed 1,452 people. The reaction to Replika's silent change was larger, and it lasted longer than the reaction to GPT-5.
Reality
A dev.to walkthrough moves Codex onto DeepSeek's API using two local config files. Codex accepts the capability figures you write into them, and the cost case in the post compares one metered API against another.
Reality
Headless 360 makes capabilities that used to sit behind the Salesforce console callable as APIs, MCP tools and CLI commands, and more than 60 of the new tools point coding agents such as Claude Code and Cursor at live orgs.
Reality
CISA, the NSA and the FBI want US model providers to catch industrial-scale distillation hiding inside traffic that looks like a busy enterprise customer, then quietly degrade the answers, which makes false positives a product decision.
Reality
The study counts only complexity and dead code, because those are the two pyscn metrics that stay exact when you analyze just the files a patch touched. The human's own commit trips the same rule 24% of the time.
Reality
The published floor is six months for generally available models, three for Codex and chat variants, and as little as two weeks for anything with preview in the name. The dated entries in the log sit on the floor.
Reality
Reaching GPT-6 Astra costs a few days of patience rather than a $200 seat, so the gate worth budgeting against is Astra Pro, which Plus plans do not get, plus the credits sold on top of existing allowances.
Reality
Programming a furnace is the eye-catching part, but the piece worth copying is a verification module that refuses any figure it cannot resolve to a logged result, which cuts reported fabrication to 4 percent without telling you what the number means.
Reality
One generated Datadog monitor per 75 lines of code, agents allowed to write fixes and delete their own alerts, and an admission that no current model can decide what needs attention.
Reality
JetBrains has taken the assembly work out of running a coding agent offline. What it could not take out is the hardware, and that is now the thing deciding who adopts.
Reality
A rollback, a new internal eval category and a co-authored study put "unhealthy emotional dependence" in the vendor's own hand. That is the evidence base a duty-of-care claim starts from.
Reality
Krasyn ran its clinical note checker against Omi Health's open benchmark and published the disagreements. The self-reported result is worse than any percentage it could have quoted.
Reality
A benchmark called CIMemories reports frontier models pushing sensitive attributes into tasks that do not need them, with GPT-5 violations climbing from 0.1% at one task to 9.6% across 40.
Reality
Google Research reports Gemini-3-Pro and GPT-5 encode 95-98% of tested facts yet fail to recall 26-34% of them, moving the fix from pretraining scale toward post-training and inference.
Reality