Astra's Quiet Wins: Cached Reasoning Swaps, Cross-Window Notes & a Twitter Clone in Minecraft
Saturday after the Astra rollout, so the volume was lower than the last three days but the signal-to-noise was better. The launch hype has mostly turned into people reporting what the model actually does differently, and several of the most useful findings are things OpenAI never put on a slide.
GPT-6 Astra, Week One Findings
Change reasoning effort without losing the cache
Eleanor Berger noticed (321 likes, retweeted by Peter Steinberger) that with Astra you can switch reasoning effort mid-conversation without invalidating the context cache. Start on Astra light, which she puts at roughly Sol medium, and turn effort up only when a task turns out to need it. Kristoph Jeffers pointed out that you could already switch effort between prompts; Eleanor's answer is that the old mechanism appended a message at the top of the chat and blew the cache, while the new one appends at the end. One reply reports a Fable 5.1 conversation allows the same thing. Another reports checking request logs after a switch and seeing a cold cache miss, so verify against your own logs before you build a workflow around it. Someone else made the sharper point: an orchestrator agent can do the switching for you, and they had already seen the reasoning level flipping on its own mid-run and assumed it was a bug.
The compaction mode that keeps notes
Gabriel Chua's thread (1,130 likes, 78 reposts) quotes OpenAI's description of Astra's new context handling in Codex: instead of repeatedly compressing history into one summary, the model keeps notes across context windows and can search earlier windows, including messages and tool calls. It is off by default. Gabriel posted the instructions as a screenshot and suggested feeding it to Codex to set up, which drew a fair reply from drnaqi: the whole thing is one TOML line, roughly features.context_management.experimental_mode = true, so why make people burn tokens on an image. Peter Steinberger called it "this one weird trick to make compaction and long running goals way better." Open questions from the replies that nobody answered: whether the notes live on OpenAI's servers rather than locally, which matters for zero-data-retention setups, and whether it drains quota faster. One user reported it broke their Codex install minutes before a fantasy football draft. Dan McAteer's original writeup of the mechanism was in yesterday's issue.
The hallucination drop nobody is talking about
Haider argues (589 likes) that Astra's biggest improvement is the hallucination rate, and that OpenAI buried it on pages 20 and 21 of the system card instead of leading with it. The pushback in the thread is worth reading alongside: John Baehr notes the figures are curated and the sample sizes small, another reply says Astra is still nowhere near the bottom of Artificial Analysis's hallucination table, and DBrodniak makes the standard objection that fewer hallucinations can just mean more "I don't know" rather than more correct answers. The most useful reply is from truevined: grade whether a cited source actually exists as a separate metric, because a non-expert can check that and it does not track the overall score.
Gert Labs: the widest gap since Opus 4.5
Leo Linsky evaluated Astra (957 likes, rankings) in 100 unsaturated multi-agent coding environments where models compete and cooperate on open-ended tasks. His summary: new frontier by a landslide, more dominant than the Fable 5 release was, and 80% cheaper and 30% faster than Fable 5.1 on the same benchmark. He also points out that some of these environments are almost a year old and still separate frontier models from deprecated ones like Sonnet 4, which he reads as evidence that open-ended evals resist saturation. swombat flagged Gemini 3.8 Flash's score as the sleeper result; Leo credits Google's RL pipeline. The "post-AGI world" line at the end of his post got the predictable argument in replies. One practical note from DragonStacker: a 20x subscription's weekly limit can go in a single eight-hour dev shift.
LlamaIndex: record on short documents, 31.7% on long ones
Jerry Liu benchmarked Astra on ExtractBench and ParseBench (extractbench.ai, parsebench.ai). One-shot extraction hits 97.2% on short documents, a new record, and 90.6% on medium ones. The caveats are the story: 11 cents per page, ten times LlamaIndex's cost-effective mode; one-shot on long documents falls to 31.7%; and native OCR is not much better than Fable 5.1 or GPT-5.6 Sol, with strong table understanding but weak charts, layout and semantic formatting. Several replies ask which harness and effort level were used, which Jerry had not answered at fetch time. Timotei Molnar asked for Gemini numbers, saying 3.7 Flash beat Sol in his own document benchmark. Jerry's related thread from Friday reports 10 to 20 point improvements in LlamaParse's cheap mode over the last three months.
Robot arms: 19 of 20 on the bowl, still stuck on the puzzle
Robocurve published GPT-6 Astra on robotic manipulation (HN, 157 points), a follow-up to their Fable 5 versus Fable 5.1 comparison on the same YAM arms and Inspect Robots policy. Block into bowl: Astra 19 of 20, Fable 5.1 8 of 20, Fable 5 1 of 20, at 2.5 minutes and $0.94 per run against Fable 5.1's 6.8 minutes and $2.12. Astra used about 2,100 output tokens per run to Fable 5.1's 12,900. The puzzle piece into a groove is unchanged: 2 of 20 for both Astra and Fable 5.1, stalling at the same final insertion step. On HN a commenter who spent two days trying to get Astra to control an Adeept tank arm via OpenClaw reports the opposite experience: servo directions wrong, gripper extent miscalculated, and a lot of unit tests instead of working code.
The Minecraft Twitter clone
Adi's post (19,700 likes, 1.9 million views, quoted by Theo as "Minecraft is solved") is Astra's run on his recurring test, "make a Twitter clone in Minecraft," which he first ran on Fable 5 versus GPT-5.5 in June. The follow-up post says it even built a back panel. This is a computer-use task, not a code generation one, which is why the replies read it as confirmation of OpenAI's "best computer use" claim. Theo's own contribution to the genre is a note that his Blender-built browser game cost under $30 in tokens, not the thousands people assumed.
Simon's Blender TIL and the developer video
Simon Willison wrote up Using Blender with coding agents on macOS: install the full Blender app, then prompt Codex with "Use the already installed /Applications/Blender to render a scene of a pelican riding a bicycle," followed by "add a background and a lot of flair" and "make it a whole lot better." The models drive Blender's Python API, can hand back editable .blend files, and can render image sequences into video with ffmpeg. He also linked OpenAI's "Introducing GPT-6 Astra for developers" video, which has a pelican on a bicycle at 1m59s, and quoted VB Srivastav's Ferrari recreation as an example of the model researching specs, regulations and photos before modeling. Ethan Mollick's 3D Zork, which kept the 1977 plot and puzzles and built characters and environments in Three.js, is the one Simon called "a whole lot better than any of my experimental games."
Agentic Coding & Agent Harnesses
Theo's shipping chart
Theo posted (628 likes) a chart of merged PRs on T3 Code with the caption that Fable 5.1 and Astra "have both had a profound impact on how much we are able to ship," and pointed doubters at the open source repo. A reply reads the numbers off the chart: the Fable 5.1 step took the count to 93 and Astra to 179. Peter Steinberger replied that he is now limited by CI and GitHub API calls, and Theo said the same. The skeptical reply worth noting: Theo and his team have had Astra for about a month, so the last few bars are not a clean before-and-after. Theo also asked how people's usage limits are holding up and said he would pay for a $10,000 plan if one existed.
Steinberger's harness: OpenClaw, crabbox, and a slopmeter
Peter Steinberger says OpenClaw is "one step closer" to the harness he wants to use. The missing piece is cloud sessions that start in seconds, which needs snapshotting instead of fresh repo clones; he counts clone plus environment prep as one thing and says snapshots will be able to fork into parallel sessions via crabbox, which supports around 40 providers. Asked what he used before, he said OpenClaw was his agent on the go and Codex did the heavy work, but Codex being local was the limit. OpenClaw v2026.9.2 shipped the same day: Astra and Muse Spark 1.3 support, interrupted tasks resume after restart, mid-response corrections on OpenAI API connections, a task workspace that centers a dashboard, browser, terminal or file view, and guided local-model setup that checks the model can actually call a tool before making it the default. The release is 1,245 PRs from 232 contributors. Read the upgrade note if you run a shared Gateway: omitted settings now let agents with session tools read other agents' conversations, including other users' transcripts.
He also built a slopmeter into his team server, a chart of test lines versus application lines where you can see the point they started pushing for more tests. His target is a 2:1 test-to-app ratio. When someone said the slopmeter itself looks like slop, he agreed: it is a one-shotted prompt, and the Discord agent publicly shames anyone who submits too much slop or deletes tests to game the ratio. Both OpenClaw and the prompt are open source.
Armin Ronacher: are the models nudging us all toward the same projects?
Armin's weekend post Latent Powers (thread) starts with a hardware project: he wanted a cheap Carlinkit CarPlay dongle to run his own code while passing CarPlay through, and a conversation with Fable and Sol via Pi surfaced CatPlay, a Rust reimplementation of the protocol. The dongle he bought turned out to be a different SoC than the one CatPlay targets, which is where he says he would normally have given up, but working with Kimi K3 and Sol he flashed it and got CatPlay compiling. The interesting part is what happened next: an acquaintance had independently started the same project at roughly the same time, and also learned about the CarPlay hacking community from the models. Armin's question is how much of what the AI builder community builds comes from eliciting the same latent capabilities from the same models, and whether the running joke that everyone is working on the same thing has a mechanical explanation. His example is HTML reports instead of Markdown from Pi, which felt novel when Lucas Meijer proposed it and is now a default. Separately, he broke the pi-coding-agent package for library usage in a release and blamed a gap in the dual-bundling release process.
Short takes
- Nate Berkopec, still true a year later: agents think 99% of problems are solved by adding more code, and your job is to fight that.
- Thibault Sottiaux said Astra was OpenAI's biggest competitive advantage while it was internal-only, and productivity gains pulled some DevDay ships six months forward from mid next year.
- am.will ran Muse Spark 1.3 Max against Astra on his RocketLeagueBench and says they are "not even in the same solar system," offered as a case for why benchmarks alone are not enough. He also thinks SVG generation is now effectively solved.
Other Interesting Stuff
The revolt of the reader
Bryan Cantrill's The revolt of the reader (HN, 305 points) is about people he otherwise respects publishing obviously LLM-authored pieces. His two questions for them: do you think readers can't tell, or don't care? He cites Cynthia Dunlop's survey of 668 developers where 78% stop reading immediately on detecting an LLM and 71% avoid the author afterward, and 98% prefer an imperfect human draft to a polished LLM one. His analogy is email spam, which stopped being an existential threat once it could be identified at scale and being labeled spam became ruinous. He has found Pangram 4 accurate enough, with a very low false positive rate in his months of testing, that Oxide's RFD 576 now explicitly covers public writing. HN's best reply is Lerc's: if a reader does not notice LLM text, they by definition cannot report it, so "readers can always tell" is unfalsifiable from the inside. Several commenters want a Pangram overlay for the HN front page.
LLMs as a cognitive virus
Large-Language Models as a Cognitive Virus (HN, 255 points) by Ricard Solé, David Krakauer, Michael Levin and others models LLM adoption as an epidemic with uncoupled, coupled and persistently dependent user states. The claim is that social transmission plus collective reinforcement produces tipping points and lock-in, so past a threshold small adoption increases trigger rapid population-level shifts toward dependence, with "abrupt losses in cognitive competence." The same model identifies conditions for "cognitive immunization." The HN thread is mostly the Socrates-on-writing quote and the observation that the framing works equally well for cars and cooking. The paper is 12 pages in physics.soc-ph, so treat it as a model, not a measurement.
Benedict Evans on why "just make the tool" misses the point
AI, tools and transformation argues against the idea that AI making tools trivially cheap to build will sweep away enterprise software. Most people are not tool builders and do not think about how their job could be done differently; the hard part was never writing the code but knowing a tool is needed and what it should do; and anything touching 50 or 500 people across departments and regulatory regimes has to be a purchase and a decision, not a spreadsheet. His useful frame is a spectrum from institutionalized software like SAP to improvised substrates like Excel and email, with tasks migrating from improvised to institutionalized as they gain revenue and risk. Relevant background for the "forward-deployed engineer" trend.
Coverage note: 13 of 14 tracked accounts were fetched via the community Nitter instances (tw.eir-nya.gay hit its rate limit partway through, so two feeds and one thread were refetched on x.n0g.xyz and nitter.jaydenha.uk). @potetotes still returns nothing. @karpathy, @swyx, @leerob, @trq212 and @bcherny posted nothing new in the window, and @mattpocockuk's only post was non-technical. Thread pages for am.will's RocketLeagueBench comparison and Victor Nunez's computer-use video timed out, so those are summarized from RSS text only.