OpenAI confirms an Agents API container billing bug
A Reddit user reported a $1,600 Agents API bill. OpenAI confirmed container overbilling. I separate the resolved incident from the cleanup work apps need.
Releases
When a lab ships a model: what it actually changes for people who build software.
GLM improved its inference. Engineers still set the goal
Z.ai reports 3× serving throughput with help from a GLM coding agent. I examine what improved, what humans controlled, and what developers can reuse.
GPT-6 Sol and Luna: the price cut is the upgrade
GPT-6 Sol and Luna cut API prices while independent benchmarks show smaller gains. I compare the scores, ChatGPT access, and mixed first reports on Reddit.
Opus 5.5 improves coding. Task cost is less clear
Opus 5.5 gains ground in independent coding tests and cuts token prices. I compare the task costs, effort settings, and first Reddit reports on daily use.
Claude leads 26% of AI R&D at Anthropic. Who checks it?
Anthropic says Claude leads a quarter of its AI research work under human supervision. I look at what the figure measures and who still checks the results.
Codex CLI adds voice. The recovery fixes matter more
Codex CLI 0.155 adds experimental voice and safer worktree cleanup. I explain the fixes for lost prompts, restarted sessions and the next-day hotfix.
Fable 5.2 rumors: better answers need better evidence
Reddit users suspect a hidden Fable upgrade. I traced the claims to their sources and checked what would actually establish a new Claude model release.
DeepSeek V4.1 Flash: low API prices, huge open weights
DeepSeek V4.1 Flash activates 8B parameters on input, but its full weights are enormous. I explain the API prices, coding scores and server requirements.
OpenAI Agents API: what the Codex harness takes over
OpenAI now hosts the Codex agent loop for your application. I explain how the Agents API differs from the SDK, what persists and which costs remain.
GPT-6 Astra costs 2.5x Sol per token. It talks like a person
OpenAI's GPT-6 Astra is metered at 2.5 times Sol's rate and feels more human in conversation. I checked the limits, the benchmarks, and the first reactions.
Fable 5.1 beats Fable 5. Opus is harder to replace
Fable 5.1 leads its predecessor on long agent tasks, but Opus 5 stays close at half the base token price. I checked the benchmarks and first reports.
Codex Plus can stop work with 84% of weekly usage left
OpenAI's five-hour gate can interrupt concentrated Codex work while weekly usage remains. It limits when Plus can work, not necessarily total usage.
Claude will watermark its text. The EU is the reason
Claude models will watermark their text through the words they pick, not hidden characters. I trace the reason to the EU AI Act, not the distillation war.
Astra is real. A GPT-6 launch this week is not
OpenAI confirmed Astra as its next major model, then slowed internal work over cyber risk. The GPT-6 launch date came from a withdrawn leak and a joke.
Codex resets were a marketing stunt. The problem was real
OpenAI reset Codex usage five times in ten days while Sol drained limits faster than expected. I think the free refills helped hide the real usage problem.
DeepSeek V4 Flash lost all 9 benchmarks to Opus 4.8
V4 Flash costs $0.14 in and $0.28 out per million tokens. I checked whether its 97 to 99 percent discount makes DeepSeek's benchmark losses worth it.
Opus 5 is a great employee but a frustrating boss
Opus 5 matches Fable 5 on benchmarks at half the token price, yet it can be painful to supervise. Here is where Fable still earns its place.