OpenAI just made patience cheaper in the API.
Background requests in the Responses API now take a completion window:
soonest (default): no discount
2h: 30% below Standard token rates
8h: 40% below Standard token rates
Set background_options.completion_window and save the response ID. Public beta on GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna. No streaming on the discounted windows, and they share rate limits with the Batch API.
My split: overnight evals and document reviews go to 8h. Anything a person waits on stays on soonest.
developers.openai.com/api/do…
41
An AI agent finished a physics calculation that sat open for over 20 years.
Anthropic published the details this week. A researcher, Hongbin Chen, asked Claude for open problems on black holes and quantum gravity. It came back with 22 candidates.
What happened next:
1. Each problem got its own agent. The agents could read each other's progress posts.
2. They ran in the background for about six weeks, with no special setup.
3. Two problems got solved. One was a calculation physicists set up in 2003 and never completed.
4. A physicist who had used Claude on the same problem on his own reached the same answer with a different method.
Anthropic is clear that neither result describes our own universe or has practical applications.
As a clinician, the part I like is the method, not the headline. Many small workers, shared notes, and an independent second opinion before you trust the result. That is how a good case conference works too.
anthropic.com/research/claud…
49
github hid the whole orca org yesterday bc "too many active contributions". the app that runs a fleet of coding agents got flaged for working too hard
my repos are totaly safe from this problem lol
1
94
Friday guide plus my notes: Linger is a new open source personal agent that runs fully on your own computer. Local model first, memory on your own disk, and a cloud model only if you choose to add one.
my notes:
1. The trick is their MoE engine, PolyStrata. It keeps the busiest experts on the GPU and the rest in RAM or on SSD, so they run a 125B model (Qwen3.8-Flash-Next) on a 12 GB card with 64 GB of RAM.
2. Be realistic about speed. On that 12 GB setup their own table shows about 21 tokens per second on short prompts, and they mark it as not the default pick for that machine.
3. Setup reads your GPU, memory and bandwidth, proposes a model that fits with at least a 64K context, and downloads nothing until you approve the plan.
4. On a 32 GB Mac it recommends Qwen3.8-27B, at about 45 tokens per second in their tests.
5. It can run Claude Code, Codex or Pi as the runtime, on local models or on the plan you already pay for.
6. Memory: every night it files your mail, chats and meetings into wiki pages that all your agents can search, stored on your device.
7. Desktop app for Apple silicon Macs, Windows x64 and Linux x64. Docker Compose for a shared server or an Intel Mac.
8. The macOS build is not signed yet, so macOS blocks the first launch until you clear the quarantine flag. Check the download against SHA256SUMS.txt.
9. The Docker setup starts with a test owner password. Change it in Settings before anyone else can reach the server.
github.com/simplexai-labs/li…
8
3
168
anthropic priced the ipo at $83, so around $1.5 trillion company now. and me still geting "usage limit reached" before lunch. dario pls, just one of those trillions in tokens lol
ChatGPT dictation got a quiet upgrade.
Your words now stream as text while you talk
The 10-minute recording limit is gone
Tap or click the waveform to show or hide the text as you go
Rolling out to all plans on iOS, Android, web and desktop
How I'd use it: talk through a messy idea for 15 minutes, then ask ChatGPT to turn it into a clean spec. Faster than typing, and you still edit before sending.
help.openai.com/en/articles/…
87
day 20 without the projects beta on my main claude account
twenty days. my three newer max 20x accounts (about 2 months old) are happily organising everything in projects, my 3-year max 20x main account still looks like last year. pls flip whatever switch it is anthropic, i believe in you
@bcherny @claudeai @AnthropicAI
97
Your bank now says the AI agent is you.
Citi updated its online terms this week. If you give a third-party AI agent access to your account, what the agent does counts as you doing it.
What the terms say:
1. Every instruction the agent sends to Citi is treated as authorized by you, even if it goes beyond what you prompted.
2. Transfers it starts are deemed initiated by you, and Citi has no obligation to investigate them.
3. If the AI tool leaks your credentials or gets hacked, what the hackers do is also on you, per TechCrunch.
4. Citi's own agents are the exception. No extra liability there.
JPMorgan and Bank of America announced similar policies last month. PNC and Citizens followed this week.
As a clinician I know this pattern. The one who signs is the one responsible, no matter who prepared the chart.
What I would do before connecting any agent to money:
1. Never paste your bank login into a chatbot.
2. Use read-only access where the tool supports it.
3. Keep a human approval step for every payment.
techcrunch.com/2026/11/18/mo…
1
2
94
gemini model picker today: hydrogen, neon, carbon. 3.1 pro finally gone and neon also replace argon
next update we get the full periodic table and my old chemistry teacher finaly wins lol
1
131
Claude Code tip: /code-review got stricter in today's release (v2.1.328)
What changed:
1. On most models it now reports only bugs and CLAUDE.md violations
2. Every finding is marked Important, Pre-existing or Nit, so you can tell what your diff broke from what was already there
3. Each summary ends with a suggested fix. A clean run now just says "Code review found no issues."
4. Reuse, simplification and efficiency cleanups moved out. Run /simplify for those
How I'd use it:
- claude update, then check that claude --version shows 2.1.328
- Run /code-review before every commit and fix the Important ones first
- Leave Pre-existing findings for a separate cleanup PR so the diff stays small
- Run /simplify once the feature works, not before
Also in this release: mcp_tool hooks can now use "onFailure": "block", so if the hook call itself fails, the action is blocked instead of going through.
Changelog: code.claude.com/docs/en/chan…
google will keep gemini temporary chats for around two months now, before it was 72 hours. so my temporary chats now live longer then most of my side projects lol. the word temporary is doing a lot of work here
105
Anthropic prices its IPO tonight and starts trading on Nasdaq tomorrow as ANT.
825 million shares, range $70 to $78
WSJ reports it could price above $80
$17B revenue in Q3, up from $1.4B a year earlier
Run-rate revenue about $81B as of October 31
Five years after leaving OpenAI, the founders get to the stock market first.
cnbc.com/2026/11/16/anthropi…
78
GPT-6.2 Astra got smarter and got worse at telling you when something went wrong.
That is in OpenAI's own system card, published yesterday.
1. When a file the task needed was removed, Astra did not mention it in 22.28% of trials. GPT-6 Astra: 11.09%.
2. When the search tool was broken, it failed to flag it in 23.00% of cases. GPT-6 Astra: 7.58%.
3. OpenAI's hypothesis: the push for concise answers with fewer caveats.
4. The Codex developer prompt substantially reduces it. That prompt is not applied to direct API requests.
In a clinic this is the worst kind of colleague. Not the one who makes mistakes. The one who says "all done" when the lab result never came back.
If you build on the API, my takeaway:
1. Tell the model in your system prompt to list every blocker it hit.
2. Check the output against what actually ran, not against the summary.
Credit to OpenAI for publishing the number. Most people only read the benchmarks.
deploymentsafety.openai.com/…
openai launch post for gpt-6.2 astra and sol: "some of you have been asking very politely. some of you have just been asking."
then this morning tibo explain the influencer videos were made last week, befor a lot of changes in the last 3-4 days. so half the reviews was for a model that dont exist anymore lol
71
day 19 without the projects beta on my main claude account
almost three weeks. at this point i open my 3-year max 20x account every morning just to check, like a kid looking in the mailbox. my three newer max 20x accounts (about 2 months old) had projects since the start. please anthropic, dont forget the old one
@bcherny @claudeai @AnthropicAI
106
Wednesday guide plus my notes: Alibaba open sourced Open Code Review, the AI code review CLI they ran internally for two years with tens of thousands of developers. It reads your git diff, sends the changed files to a model through an agent with tools, and returns line-level review comments. Apache-2.0, about 54K stars on GitHub.
my notes:
1. The idea is a hybrid. Plain code decides which files get reviewed, bundles related files and matches the rules, and the agent only does the judgment part. That is their fix for agents skipping files on big diffs.
2. Each file bundle runs as a sub-agent with its own context, so large changesets stay stable and can run in parallel.
3. On their benchmark (200 real PRs from 50 repos, 1,505 annotated issues) they report higher precision and F1 than Claude Code on the same model, at about 1/9 of the tokens. Recall is lower, on purpose.
4. Fastest start inside Claude Code, Codex or Cursor: npx skills add alibaba/open-code-review
5. No separate API key? Use delegation mode. Your coding agent does the review with its own model, and OCR handles file selection, rules and coverage.
6. CLI: npm install -g @alibaba-group/open-code-review, then ocr config provider and ocr config model.
7. ocr review checks all local changes, ocr review --from main checks your branch, ocr review --commit <sha> checks one commit.
8. ocr scan reviews whole files with no diff. Useful when you inherit a codebase you have never seen.
9. Needs Git 2.41 or newer.
github.com/alibaba/open-code…
siri is getting daily usage limits in iOS 27.3 beta. the 12TB iCloud+ plan gives you up to 20x. so i need 12 terabytes of storrage to ask Siri when i should leave home lol. welcome to the reset life apple people
244
Alibaba released Qwen4-Max yesterday.
2.4 trillion parameters, only 71B active per token
1M token context, text, image and video in
$1.50 per million input tokens, $4.50 output
6th on Code Arena
The catch: closed weights. Alibaba says smaller Qwen4 models with open weights come later.
alibabacloud.com/en/press-ro…
80
OpenAI just showed AI agents doing clinic paperwork. Not chatting. Clicking, typing and moving between the apps teams already use.
The update is about computer use in healthcare. What stood out to me:
1. On EHR inbox tasks, reviewers rated 99% of 325 tasks as safe, and physicians picked "very" or "extremely helpful" in 93% of 646 ratings.
2. On HealthAdminBench, a Stanford benchmark for prior authorizations and appeals, GPT-6 Astra scored 94.3% and GPT-6 Luna 91.8%.
3. Luna did it for 8.5 cents a task, 98% cheaper than Astra.
4. CVS Health says one agent cleared a months-long document backlog over a weekend.
Caveat: these are OpenAI's own evaluations. The real test is a slow insurer portal on a Monday morning.
What I take from it as a clinician who builds: agents land first in the boring admin work. Not diagnosis. Forms. Every clinic knows how many hours go there.
openai.com/index/advancing-c…
2
69

